Can my computer run this LLM? Memory, CPU offload and speed for GPUs, Macs and mini-PCs

NVIDIA GeForce RTX 4060

NVIDIA RTX 4060 launch spec-comparison image (on https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4060-4060ti/): Memory Subsystem '24 MB L2 / 272 GB/s (453 GB/s effective)', TGP '115 W'. 272 GB/s is the physical DRAM bandwidth; 453 GB/s is NVIDIA's cache-adjusted 'effective' figure. Per-pin data rate not stated by NVIDIA. Power: Total Graphics Power. Compare page (40 Series): '8 GB GDDR6', 128-bit, Total Graphics Power (W) '115'. Product page: https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4060-4060ti/. GeForce news dated May 18, 2023 says RTX 4060 arrives in July (2023). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 4060 Ti (8 GB)

NVIDIA RTX 4060 Ti launch spec-comparison image (on https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4060-4060ti/): VRAM '8GB / 16GB', Memory Subsystem '32 MB L2 / 288 GB/s (554 GB/s effective)', TGP '160 W'. VRAM explainer (https://www.nvidia.com/en-us/geforce/news/rtx-40-series-vram-video-memory-explained/) also: 'an Ada GPU with 288 GB/sec of peak memory bandwidth' in the RTX 4060 Ti test. Per-pin data rate not stated by NVIDIA. Power: Total Graphics Power. Compare page: '16 GB GDDR6 or 8 GB GDDR6', 128-bit, Total Graphics Power '165 or 160' (8 GB = 160 W by position). Available May 24, 2023 per GeForce news dated May 18, 2023. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 5050

Compare page (50 Series) 'Memory Bandwidth' row: '320 GB/sec'. Per-pin data rate not stated by NVIDIA. Power: Total Graphics Power. Later 50-series desktop card (not in the original request list). Compare page and product page: '8 GB GDDR6', 128-bit, Total Graphics Power (W) '130'. GeForce news (https://www.nvidia.com/en-us/geforce/news/rtx-5050-desktop-gpu-and-laptops/, seen via search result) says desktop cards arrive in the second half of July (2025). Second source: https://www.nvidia.com/en-us/geforce/graphics-cards/50-se....

NVIDIA GeForce RTX 3060 Ti

NVIDIA RTX 4060 Ti launch spec-comparison image (on https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4060-4060ti/) lists RTX 3060 Ti: '8GB' VRAM, Memory Subsystem '4MB L2 / 448 GB/s', TGP '200W'. Applies to the original GDDR6 card; no official bandwidth found for the later GDDR6X variant. Per-pin data rate not stated by NVIDIA. Power: Graphics Card Power. Compare page: Standard Memory Config '8 GB GDDR6 / 8 GB GDDR6X', 256-bit, Graphics Card Power (W) '200'. Release year from GeForce news 'GeForce RTX 3060 Ti out December 2' dated December 01, 2020 (https://www.nvidia.com/en-us/geforce/news/geforce-rtx-3060-ti-out-december-2/). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 3070

Ampere GA102 whitepaper Table 10 (p.47, 'RTX 3070 Founders Edition'): Memory Interface 256-bit, 'Memory Clock (Data Rate) 14 Gbps', 'Memory Bandwidth 448 GB/sec', TGP 220 Watts. Same values in RTX Blackwell whitepaper Table 6 (p.54-55). Power: Graphics Card Power. Compare page: 8 GB GDDR6, 256-bit, Graphics Card Power (W) '220'. Product page: https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3070-3070ti/. Release year: RTX 30 Series introduced in GeForce news dated September 01, 2020. Second source: https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-g....

NVIDIA GeForce RTX 5060

Compare page (50 Series) 'Memory Bandwidth' row: '448 GB/sec'. Per-pin data rate not stated on any NVIDIA page found. Power: Total Graphics Power. Compare page / RTX 5060 family page: '8 GB GDDR7', 128-bit, Total Graphics Power (W) '145'. NVIDIA newsroom (April 15, 2025): 'GeForce RTX 5060 graphics cards will be available starting in May at $299'. Second source: https://www.nvidia.com/en-us/geforce/graphics-cards/50-se....

NVIDIA GeForce RTX 5060 Ti (8 GB)

Compare page (50 Series): RTX 5060 Ti column 'Standard Memory Config 16 GB / 8 GB GDDR7', 'Memory Bandwidth 448 GB/sec' (one value for both memory sizes). Per-pin data rate not stated. Power: Total Graphics Power. Total Graphics Power '180' (single value listed for both memory sizes). NVIDIA newsroom (April 15, 2025, https://nvidianews.nvidia.com/news/nvidia-blackwell-geforce-rtx-arrives-for-every-gamer-starting-at-299): 'GeForce RTX 5060 Ti graphics cards, equipped with 16GB or 8GB graphics memory, will be available starting April 16'. Second source: https://www.nvidia.com/en-us/geforce/graphics-cards/50-se....

NVIDIA GeForce RTX 3080 (10 GB)

Ampere GA102 whitepaper Table 2 (p.14-15, 'GeForce RTX 3080 10 GB Founders Edition'): 320-bit, 'Memory Clock (Data Rate) 19 Gbps', 'Memory Bandwidth 760 GB/sec', 'TGP (Total Graphics Power) 320W'. Also RTX Blackwell whitepaper Table 4 (p.49-50). Power: Graphics Card Power. Compare page: '12 GB GDDR6X / 10 GB GDDR6X', '384-bit / 320-bit', Graphics Card Power '350 / 320' (10 GB = 320 W). Product page: https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3080-3080ti/. Release year: RTX 30 Series GeForce news dated September 01, 2020 ('on sale starting September 17th'). Second source: https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-g....

NVIDIA GeForce RTX 3080 (12 GB)

No NVIDIA page found that states either the memory bandwidth or the per-pin data rate for the 12 GB RTX 3080, so bandwidth cannot be stated or derived. Left null. Power: Graphics Card Power. Compare page and RTX 3080 family page: '12 GB GDDR6X / 10 GB GDDR6X', '384-bit / 320-bit', Graphics Card Power '350 / 320' (12 GB = 350 W). release_year 2022 not verified against a dated NVIDIA page. Second source: https://www.nvidia.com/en-us/geforce/graphics-cards/30-se....

NVIDIA GeForce RTX 3060 (12 GB)

NVIDIA RTX 4060 launch spec-comparison image (on https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4060-4060ti/) lists RTX 3060: '12GB' VRAM, Memory Subsystem '3MB L2 / 360 GB/s', TGP '170W'. NVIDIA does not state the per-pin data rate on any page found. Power: Graphics Card Power. Compare page (30 Series): Standard Memory Config '12 GB GDDR6 / 8 GB GDDR6', Memory Interface Width '192-bit / 128-bit' (12 GB = 192-bit), Graphics Card Power (W) '170'. Release year from GeForce news 'GeForce RTX 3060' dated January 12, 2021 (https://www.nvidia.com/en-us/geforce/news/geforce-rtx-3060/). The 8 GB/128-bit variant is not included. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 4070

RTX Blackwell whitepaper Table 6 (p.54-55, 'RTX 4070'): '12 GB GDDR6X', 192-bit, 'Memory Clock (Data Rate) 21 Gbps', 'Memory Bandwidth 504 GB/sec', TGP 200 W. These apply to the original GDDR6X card; NVIDIA states no bandwidth or data rate for the later GDDR6 variant. Power: Total Graphics Power. Compare page / RTX 4070 family page: '12 GB GDDR6 / 12 GB GDDR6X', 192-bit, Total Graphics Power '200'. Release year from GeForce news 'GeForce RTX 4070' dated April 12, 2023 (https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4070/). Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/black....

NVIDIA GeForce RTX 4070 SUPER

NVIDIA CES 2024 spec image for RTX 4070 SUPER (on https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4080-4070-ti-4070-super-gpu/): Frame Buffer '12GB G6X', Memory Subsystem '48MB L2 / 504 GB/sec', TGP '220 W'. Per-pin data rate not stated. Power: Total Graphics Power. Compare page: 12 GB GDDR6X, 192-bit, Total Graphics Power '220'. Release year from GeForce news dated January 08, 2024 (40 SUPER series launching January 2024). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 4070 Ti

RTX Blackwell whitepaper Table 5 (p.51-52, 'RTX 4070 Ti'): '12 GB GDDR6X', 192-bit, 'Memory Clock (Data Rate) 21 Gbps', 'Memory Bandwidth 504 GB/sec', TGP 285 W. Power: Total Graphics Power. Compare page / RTX 4070 family page (https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4070-family/): 12 GB GDDR6X, 192-bit, Total Graphics Power '285'. Available January 5, 2023 per GeForce CES 2023 news (seen via search result only). Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/black....

NVIDIA GeForce RTX 5070

Compare page 'Memory Bandwidth' row: '672 GB/sec'. RTX Blackwell whitepaper Table 6 (p.54-55): 'Memory Clock (Data Rate) 28 Gbps', 'Memory Bandwidth 672 GB/sec', TGP 250 W. Power: Total Graphics Power. Compare page / RTX 5070 family page (https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/): '12 GB GDDR7', 192-bit, Total Graphics Power '250'. GeForce news 'RTX 5070 out now' dated March 05, 2025. Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/black....

NVIDIA GeForce RTX 3080 Ti

Ada whitepaper Table 3 / Appendix B (p.32-35, 'RTX 3080 Ti' column): 384-bit, 'Memory Clock (Data Rate) 19 Gbps', 'Memory Bandwidth 912 GB/sec', TGP 350 W. RTX 4080 SUPER launch spec image (https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/news/geforce-rtx-4080-4070-ti-4070-super-gpu/nvidia-geforce-rtx-ces-2024-4080-super-specifications.png) also lists RTX 3080 Ti '12GB G6X', '6MB L2 / 912 GB/sec', TGP '350W'. Power: Graphics Card Power. Compare page: 12 GB GDDR6X, 384-bit, Graphics Card Power '350'. Release year: unveiled at COMPUTEX 2021 per GeForce news (https://www.nvidia.com/en-us/geforce/news/rtx-3080-ti-3070-ti-graphics-cards/), seen via search result only. Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/ada/n....

NVIDIA GeForce RTX 4060 Ti (16 GB)

Same NVIDIA RTX 4060 Ti spec image (VRAM '8GB / 16GB', '288 GB/s'); GeForce news (https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4060-4060ti/): the 16GB model has 'additional graphics memory but otherwise identical specifications'. Per-pin data rate not stated. Power: Total Graphics Power. Compare page / RTX 4060 family page: Total Graphics Power '165 or 160' listed against '16 GB GDDR6 or 8 GB GDDR6', so 16 GB = 165 W by position; the launch spec image shows only 160 W for the 4060 Ti column. Arrived July 2023 per GeForce news dated May 18, 2023. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 5060 Ti (16 GB)

Compare page (50 Series): RTX 5060 Ti column '16 GB / 8 GB GDDR7', 128-bit, 'Memory Bandwidth 448 GB/sec'. Per-pin data rate not stated. Power: Total Graphics Power. Total Graphics Power '180'. Available April 16, 2025 per NVIDIA newsroom (April 15, 2025). Second source: https://www.nvidia.com/en-us/geforce/graphics-cards/50-se....

NVIDIA GeForce RTX 4070 Ti SUPER

NVIDIA CES 2024 spec image for RTX 4070 Ti SUPER (on https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4080-4070-ti-4070-super-gpu/): Frame Buffer '16GB G6X', Memory Subsystem '48MB L2 / 672 GB/sec', TGP '285 W'. Per-pin data rate not stated. Power: Total Graphics Power. Compare page: 16 GB GDDR6X, 256-bit, Total Graphics Power '285'. Release year from GeForce news dated January 08, 2024. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 4080

Ada whitepaper Table 3 / Appendix B (p.32-35, 'RTX 4080 16 GB'): 256-bit, 'Memory Clock (Data Rate) 22.4 Gbps', 'Memory Bandwidth 716.8 GB/sec', TGP 320 W. Same in RTX Blackwell whitepaper Table 4. Power: Total Graphics Power. Compare page / RTX 4080 family page (https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4080-family/): 16 GB GDDR6X, 256-bit, Total Graphics Power '320'. release_year 2022 not verified against a dated NVIDIA page. Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/ada/n....

NVIDIA GeForce RTX 4080 SUPER

NVIDIA CES 2024 spec image for RTX 4080 SUPER (on https://www.nvidia.com/en-us/geforce/news/geforce-rtx-4080-4070-ti-4070-super-gpu/): Frame Buffer '16GB G6X', Memory Subsystem '64MB L2 / 736 GB/sec', TGP '320 W'. Article text: 'GDDR6X video memory (VRAM) running at 23 Gbps' (23 x 256 / 8 = 736, consistent). Power: Total Graphics Power. Compare page: 16 GB GDDR6X, 256-bit, Total Graphics Power '320'. GeForce news dated January 08, 2024: 'arrives January 31st'. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/geforc....

NVIDIA GeForce RTX 5070 Ti

Compare page 'Memory Bandwidth' row: '896 GB/sec'. RTX Blackwell whitepaper Table 5 (p.51-52): 'Memory Clock (Data Rate) 28 Gbps', 'Memory Bandwidth 896 GB/sec', TGP 300 W. Power: Total Graphics Power. Compare page / RTX 5070 family page: '16 GB GDDR7', 256-bit, Total Graphics Power '300'. GeForce news 'RTX 5070 Ti out now' dated February 20, 2025. Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/black....

NVIDIA GeForce RTX 5080

Compare page 'Memory Bandwidth' row: '960 GB/sec'. RTX Blackwell whitepaper Table 4 (p.49-50): 'Memory Clock (Data Rate) 30 Gbps', 'Memory Bandwidth 960 GB/sec'; text: 'RTX 5080 ships with 30 Gbps GDDR7 memory, delivering 960 GB/sec'. Power: Total Graphics Power. Compare page / product page (https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/): '16 GB GDDR7', 256-bit, Total Graphics Power '360'. GeForce news 'RTX 5090 & 5080 out now' dated January 30, 2025. Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/black....

NVIDIA GeForce RTX 3090

Ampere GA102 whitepaper Table 9 (p.44-45, 'RTX 3090 FE'): 384-bit, 'Memory Clock (Data Rate) 19.5 Gbps', 'Memory Bandwidth (GB/sec) 936 GB/sec', 'Total Graphics Power (TGP) 350 Watts'. Also RTX Blackwell whitepaper Table 1/3. Power: Graphics Card Power. Compare page: 24 GB GDDR6X, 384-bit, Graphics Card Power '350'. Product page: https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/. Release year: RTX 30 Series GeForce news dated September 01, 2020. Second source: https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-g....

NVIDIA GeForce RTX 3090 Ti

Ada whitepaper Table 1 (p.13) and Table 2 / Appendix A (p.29-31, 'GeForce RTX 3090 Ti'): 384-bit, 'Memory Clock (Data Rate) 21 Gbps', 'Memory Bandwidth 1008 GB/sec', 'TGP (Total Graphics Power) 450 W'. GeForce news (March 29, 2022): '24GB of the fastest 21Gbps GDDR6X memory'. Power: Graphics Card Power. Compare page: 24 GB GDDR6X, 384-bit, Graphics Card Power '450'. Release year from GeForce news 'GeForce RTX 3090 Ti out now' dated March 29, 2022 (https://www.nvidia.com/en-us/geforce/news/geforce-rtx-3090-ti-out-now/). Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/ada/n....

NVIDIA GeForce RTX 4090

Ada whitepaper Table 1 (p.13) and Table 2 / Appendix A (p.29-31): 384-bit, 'Memory Clock (Data Rate) 21 Gbps', 'Memory Bandwidth 1008 GB/sec', 'TGP (Total Graphics Power) 450 W'. Same in RTX Blackwell whitepaper Table 1/3. Power: Total Graphics Power. Compare page / product page (https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/): 24 GB GDDR6X, 384-bit, Total Graphics Power '450'. Available October 12, 2022 per GeForce news 'GeForce RTX 4090 Out Now' (seen via search result only). Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/ada/n....

NVIDIA GeForce RTX 5090

Compare page 'Memory Bandwidth' row: '1792 GB/sec'. RTX Blackwell whitepaper Table 1 (p.14) and Table 3 (p.46-47): 512-bit, 'Memory Clock (Data Rate) 28 Gbps', 'Memory Bandwidth 1792 GB/sec', TGP 575 W. Power: Total Graphics Power. Compare page / product page (https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/): '32 GB GDDR7', 512-bit, Total Graphics Power '575'. GeForce news dated January 30, 2025. Second source: https://images.nvidia.com/aem-dam/Solutions/geforce/black....

NVIDIA RTX A4000

RTX A4000 datasheet (scanned PDF, read visually): 'Memory interface 256-bit', 'Memory bandwidth 448 GB/s', 'Power consumption Total board power: 140 W'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: GPU Memory '16GB GDDR6 with error-correction code (ECC)', Max Power Consumption '140 W' (datasheet calls it 'Total board power'). Announced April 12, 2021 per NVIDIA newsroom (seen via search result). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX 4000 Ada Generation

RTX 4000 Ada datasheet: 'GPU Memory 20GB GDDR6', 'Memory Interface 160 bit', 'Memory Bandwidth 360GB/s', 'Total board power: 130W'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: '20GB GDDR6 with error-correction code (ECC)', Max Power Consumption '130W', single slot. Announced at SIGGRAPH (Aug) 2023, available fall 2023 per NVIDIA newsroom (seen via search result). The separate RTX 4000 SFF Ada is not included. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX 4500 Ada Generation

RTX 4500 Ada datasheet: 'GPU Memory 24GB GDDR6', 'Memory Interface 192 bit', 'Memory Bandwidth 432GB/s', 'Total board power: 210W'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: '24GB GDDR6 with error-correction code (ECC)', Max Power Consumption '210W'. Available fall 2023 per NVIDIA newsroom (seen via search result). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX PRO 4000 Blackwell

Datasheet: 'GPU Memory 24 GB GDDR7 with ECC', 'Memory Interface 192-bit', 'Memory Bandwidth 672 GB/s', 'Total board power: 145 W'. Product page also: 'Memory Bandwidth 672 GB/sec'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: Memory Configuration '24GB GDDR7 with error-correcting code (ECC)', Max Power Consumption '145 W'. NVIDIA newsroom (March 18, 2025): RTX PRO 5000/4500/4000 'available in the summer' (2025). The RTX PRO 4000 Blackwell SFF Edition is a separate product, not included. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX A5000

RTX A5000 datasheet: 'Memory interface 384-bit', 'Memory bandwidth 768 GB/s', 'Power consumption Total board power: 230 W'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: '24GB GDDR6 with error-correction code (ECC)', Max Power Consumption '230 W'; NVLink 2-way. Announced April 12, 2021 per NVIDIA newsroom (seen via search result). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX 5000 Ada Generation

RTX 5000 Ada datasheet: 'GPU Memory 32GB GDDR6', 'Memory Interface 256 bit', 'Memory Bandwidth 576GB/s', 'Total board power: 250W'. The datasheet is the one linked from the product page (resources.nvidia.com viewer); the PDF file itself is served from cdn.pathfactory.com/assets/10412/contents/1095460/b57d512c-c1b8-4fea-b9e5-cb95bcb33c44.pdf. Per-pin data rate not stated. Power: Max Power Consumption. Product page: '32GB GDDR6 with error-correction code (ECC)', Max Power Consumption '250W'. Shipping from August 2023 (SIGGRAPH) per NVIDIA newsroom (seen via search result). Second source: https://resources.nvidia.com/en-us-briefcase-for-datashee....

NVIDIA RTX PRO 4500 Blackwell Workstation Edition

Datasheet: 'GPU memory 32 GB GDDR7 with ECC', 'Memory interface 256-bit', 'Memory bandwidth 896 GB/s', 'Total board power: 200 W'. Product page: 'Memory Bandwidth 896 GB/s'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: '32GB GDDR7 with error-correcting code (ECC)', Max Power Consumption '200 W'. Summer 2025 availability per NVIDIA newsroom (March 18, 2025). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/data-c....

NVIDIA RTX A6000

RTX A6000 datasheet: 'Memory interface 384-bit', 'Memory bandwidth 768 GB/s', 'Total board power: 300 W'. Ampere GA102 whitepaper Table 3 (p.15-16): 'Memory Clock (Data Rate) 16 Gbps', 'Memory Bandwidth 768 GB/sec', 'TGP 300W'. Power: Max Power Consumption. Product page: '48 GB GDDR6 with error-correcting code (ECC)', Max Power Consumption '300 W'. Data rate from https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.pdf. release_year 2020 not verified against a dated NVIDIA page. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX 6000 Ada Generation

RTX 6000 Ada datasheet (nvidia.com copy; identical values in dam-cdn.nvd.orangelogic.com/AssetLink/83uo5f1t66fkm0dh16mi5uq1412l7ux5.pdf): 'GPU memory 48GB GDDR6', 'Memory interface 384-bit', 'Memory bandwidth 960 GB/s', 'Total board power: 300 W'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: '48GB GDDR6 with error-correcting code (ECC)', Max Power Consumption '300 W'. Announced September 20, 2022, available from December 2022 per NVIDIA newsroom (seen via search result). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/design....

NVIDIA RTX PRO 5000 Blackwell (48 GB)

Datasheet: 'GPU Memory 48 GB GDDR7 with ECC | 72 GB GDDR7 with ECC', 'Memory Interface 384-bit', 'Memory Bandwidth 1,344 GB/s', 'Total board power: 300 W'. Product page: 'Memory Bandwidth 1,344 GB/sec'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: Max Power Consumption '300 W'. Summer 2025 availability per NVIDIA newsroom (March 18, 2025). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX PRO 5000 72GB Blackwell

Same datasheet and product-page spec table as the 48 GB model: one column lists '48 GB GDDR7 with ECC' and '72 GB GDDR7 with ECC', 'Memory Interface 384-bit', 'Memory Bandwidth 1,344 GB/s' (one value for both). Per-pin data rate not stated. Power: Max Power Consumption. Max Power Consumption '300 W' (same column). NVIDIA blog dated December 18, 2025 (https://blogs.nvidia.com/blog/rtx-pro-5000-72gb-blackwell-gpu/): 'now generally available'. Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition

Max-Q datasheet: 'GPU Memory 96 GB GDDR7 with ECC', 'Memory Interface 512-bit', 'Memory Bandwidth 1792 GB/s', 'Total board power: 300 W'. Product page: 'Memory Bandwidth 1792 GB/sec'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: Max Power Consumption '300 W'; RTX PRO 6000 series page lists Power Consumption '300 W' for Max-Q. Available from April 2025 per NVIDIA newsroom (March 18, 2025). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/produc....

NVIDIA RTX PRO 6000 Blackwell Workstation Edition

Datasheet: 'GPU Memory 96 GB GDDR7 with ECC', 'Memory Interface 512-bit', 'Memory Bandwidth 1792 GB/s', 'Total board power: 600 W'. Product page: 'Memory Bandwidth 1792 GB/sec'. Per-pin data rate not stated. Power: Max Power Consumption. Product page: Max Power Consumption '600 W', 5.4" x 12" dual slot, double flow-through. NVIDIA newsroom (March 18, 2025): available from distributors starting April (2025). Second source: https://www.nvidia.com/content/dam/en-zz/Solutions/data-c....

AMD Radeon RX 7600

AMD product page: 'Memory Bandwidth Up to 288 GB/s'; 'Effective Memory Bandwidth Up to 477 GB/s' (includes 32 MB Infinity Cache) - not used. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1102; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); NOT in the ROCm 7.14.0 Linux GPU table (newly listed in the 10.x stream). ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2; no WSL2 option for this GPU); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: AMD press release May 24, 2023 (available from May 25). Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://ir.amd.com/news-events/press-releases/detail/1133....

AMD Radeon RX 9050

AMD product page: 'Up to 288 GB/s'. 32 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1200; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); NOT in the ROCm 7.14.0 Linux GPU table (newly listed in the 10.x stream). ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2; no WSL2 option for this GPU); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Extra 9000-series desktop card found on AMD spec page. No launch date found on AMD pages; release_year null. Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://www.amd.com/en/products/specifications/graphics.html.

AMD Radeon RX 9060

AMD product page: 'Up to 288 GB/s'. 32 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1200; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Extra 9000-series desktop card found on AMD spec page. No launch date found on AMD pages; release_year null. Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://www.amd.com/en/products/specifications/graphics.html.

AMD Radeon RX 9060 XT (8GB)

AMD product page: 'Up to 320 GB/s'. 32 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1200; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: Computex press release May 20, 2025 (TBP 'Starting at 150W', availability 'later this year'). ROCm docs list 'AMD Radeon RX 9060 XT' without distinguishing 8 GB vs 16 GB. Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://ir.amd.com/news-events/press-releases/detail/1253....

AMD Radeon RX 7700 XT

AMD product page: 'Up to 432 GB/s'; effective (with 48 MB Infinity Cache) 'Up to 1995 GB/s' - not used. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1101; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2; no WSL2 option for this GPU); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: AMD press release Aug 25, 2023 (board-partner availability from Sept 6, 2023). Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://ir.amd.com/news-events/press-releases/detail/1153....

AMD Radeon RX 9070 GRE

AMD product page: 'Up to 432 GB/s'. 48 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1201; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: AMD Computex 2026 blog (May 31, 2026) says the card launches 'globally' with availability 'starting June 2'; product page has no launch date and any earlier regional launch was not verified from an AMD source. CONFLICT: ROCm GPU specs page lists 'VRAM (GiB) 16' for RX 9070 GRE; AMD product page and the Computex 2026 blog say 12 GB - 12 used. Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://www.amd.com/en/blogs/2026/amd-computex-2026-10-ye....

AMD Radeon RX 7600 XT

AMD product page: 'Up to 288 GB/s'; effective (with 32 MB Infinity Cache) 'Up to 477 GB/s' - not used. Power: Typical Board Power (Desktop). ROCm on Linux: not listed (absent from ROCm 10.1.0 compatibility matrix and ROCm 7.14.0 Linux GPU table; same gfx1102 target as the listed RX 7600). ROCm on Windows: not listed in ROCm 10.1.0 matrix; listed as supported (Runtime and HIP SDK) in HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: AMD press release Jan 8, 2024 (available from Jan 24, 2024); press release lists TBP 'Starting at 190W'. 16 GB on a 128-bit bus. ROCm caveat: not on the ROCm 10.1.0 matrix even though RX 7600 (same gfx1102) is. Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://ir.amd.com/news-events/press-releases/detail/1176....

AMD Radeon RX 9060 XT (16GB)

AMD product page: 'Up to 320 GB/s'. 32 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1200; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: Computex press release May 20, 2025 (TBP 'Starting at 160W'). ROCm docs list 'AMD Radeon RX 9060 XT' without memory variant. Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://ir.amd.com/news-events/press-releases/detail/1253....

AMD Radeon RX 9060 XT LP

AMD product page: 'Up to 320 GB/s'. 32 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1200; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2; no WSL2 option for this GPU); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Extra 9000-series desktop card (low-power 16 GB variant). Product page power reads 'Up to 140W'. No launch date found; release_year null. Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://www.amd.com/en/products/specifications/graphics.html.

AMD Radeon RX 7900 GRE

AMD product page: 'Up to 576 GB/s'; effective (with 64 MB Infinity Cache) 'Up to 2250 GB/s' - not used. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1100; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year from AMD product page 'Launch Date 7/27/2023'. ROCm caveat: in ROCm 10.1.0 matrix (incl. Windows) but not in the HIP SDK for Windows 7.2.0 GPU table. Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://rocm.docs.amd.com/en/latest/compatibility/compati....

AMD Radeon RX 7700

AMD product page: 'Up to 624 GB/s'; effective (with 40 MB Infinity Cache) 1927 GB/s - not used. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1101; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2; no WSL2 option for this GPU); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Extra card (not in the requested list) included because it is a 16 GB RX 7000 desktop card on AMD's spec page. No launch date on AMD page; release_year null. Values from AMD graphics specifications data (product page URL as listed there). Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://www.amd.com/en/products/specifications/graphics.html.

AMD Radeon RX 7800 XT

AMD product page: 'Up to 624 GB/s'; effective (with 64 MB Infinity Cache) 'Up to 2708 GB/s' - not used. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1101; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: AMD press release Aug 25, 2023 (available from Sept 6, 2023). Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://ir.amd.com/news-events/press-releases/detail/1153....

AMD Radeon RX 9070

AMD product page: 'Up to 640 GB/s'. 64 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1201; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: AMD press release Feb 28, 2025 (available from March 6, 2025). Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://ir.amd.com/news-events/press-releases/detail/1238....

AMD Radeon RX 9070 XT

AMD product page: 'Up to 640 GB/s'. 64 MB Infinity Cache. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1201; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Release year: AMD press release Feb 28, 2025 (available from March 6, 2025). ROCm 10.1.0 release notes list as resolved a PyTorch training/fine-tuning GPU-reset issue that affected RX 9070 Series and Radeon AI PRO R9700. Architecture per ROCm compatibility matrix / GPU specs page and AMD RDNA 4 press releases. Second source: https://ir.amd.com/news-events/press-releases/detail/1238....

AMD Radeon RX 7900 XT

AMD product page: 'Up to 800 GB/s'; effective (with 80 MB Infinity Cache) 'Up to 2900 GB/s' - not used. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1100; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Product page 'Launch Date 12/13/2022'. Bus width 320-bit from AMD press release Nov 3, 2022 (product page omits Memory Interface; 320 x 20 / 8 = 800 matches stated bandwidth). CONFLICT: product page 'Typical Board Power (Desktop) 315 W' vs press release table 'TBP 300W'; product page value used. Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://ir.amd.com/news-events/press-releases/detail/1099....

AMD Radeon RX 7900 XTX

AMD product page: 'Memory Bandwidth Up to 960 GB/s'; 'Effective Memory Bandwidth Up to 3500 GB/s' (includes 96 MB Infinity Cache) - not used. Power: Typical Board Power (Desktop). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1100; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). Product page 'Launch Date 12/13/2022'. Bus width 384-bit from AMD press release Nov 3, 2022 ('24GB of high-speed GDDR6 memory running at 20Gbps over a 384-bit memory bus'); product page omits Memory Interface. Architecture per ROCm compatibility matrix / GPU specs page (RDNA 3). Second source: https://ir.amd.com/news-events/press-releases/detail/1099....

AMD Radeon PRO W7800

AMD product page: 'Peak Memory Bandwidth 576 GB/s'. 64 MB Infinity Cache. Power: Total Board Power (TBP). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1100; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). ECC memory supported per AMD spec. Memory speed: product page lists none; AMD press release Apr 13, 2023 table header reads 'GDDR6 ECC Memory (18GB/s)' (apparently 18 Gbps; 256 x 18 / 8 = 576 matches). Release year: press release Apr 13, 2023 (retail availability 'starting in Q2, 2023'). Architecture from product page 'GPU Architecture AMD RDNA 3'. Second source: https://ir.amd.com/news-events/press-releases/detail/1123....

AMD Radeon AI PRO R9600

AMD product page: 'Peak Memory Bandwidth 640 GB/s'. Memory speed not stated. Power: Total Board Power (TBP). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1201; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); NOT in the ROCm 7.14.0 Linux GPU table (newly listed in the 10.x stream). ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2; no WSL2 option for this GPU); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). ECC 'Yes (Linux Only)'. Newer Radeon AI PRO card: active cooling, single slot, 48 CUs (ROCm GPU specs). ROCm 10.1.0 release notes: 'ROCm 10.1.0 adds support for AMD Radeon AI PRO R9600 GPUs.' No launch date found; release_year null. Second source: https://rocm.docs.amd.com/en/latest/about/release-notes.html.

AMD Radeon AI PRO R9600D

AMD product page: 'Peak Memory Bandwidth 640 GB/s'. Memory speed not stated. Power: Total Board Power (TBP). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1201; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2; no WSL2 option for this GPU); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). ECC 'Yes (Linux Only)'. Newer Radeon AI PRO card: passive cooling, single slot (server/chassis-airflow design). No launch date found; release_year null. Second source: https://www.amd.com/en/products/specifications/profession....

AMD Radeon AI PRO R9700

AMD product page: 'Peak Memory Bandwidth 640 GB/s'. 64 MB Infinity Cache. Memory speed not stated (640 x 8 / 256 implies 20 Gbps). Power: Total Board Power (TBP). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1201; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). ECC 'Yes (Linux Only)' per AMD spec. Release year: Computex press release May 20, 2025 ('expected to be available from leading board partners starting in July 2025'). Active cooling, double slot, 12V-2x6 connector. ROCm 10.1.0 release notes list as resolved a PyTorch training GPU-reset issue affecting R9700. Second source: https://ir.amd.com/news-events/press-releases/detail/1253....

AMD Radeon AI PRO R9700S

AMD product page: 'Peak Memory Bandwidth 640 GB/s'. Memory speed not stated. Power: Total Board Power (TBP). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1201; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); NOT in the ROCm 7.14.0 Linux GPU table (newly listed in the 10.x stream). ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). ECC 'Yes (Linux Only)'. Newer Radeon AI PRO card: passive cooling, double slot (server/chassis-airflow design, not a typical home card). No launch date found; release_year null. Not in ROCm 7.14.0 Linux table or HIP SDK Windows 7.2.0 table; listed in ROCm 10.1.0 matrix. Second source: https://www.amd.com/en/products/specifications/profession....

AMD Radeon PRO W7800 48GB

AMD product page: 'Peak Memory Bandwidth 864 GB/s'. 96 MB Infinity Cache. Memory speed not stated (864 x 8 / 384 implies 18 Gbps). Power: Total Board Power (TBP). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1100; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); not listed in the HIP SDK for Windows 7.2.0 GPU table. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). ECC memory supported per AMD spec. Extra card (48 GB variant of W7800). No launch date found on AMD pages; release_year null. Architecture from product page 'AMD RDNA 3'. ROCm caveat: not in HIP SDK for Windows 7.2.0 GPU table. Second source: https://www.amd.com/en/products/specifications/profession....

AMD Radeon PRO W7900

AMD product page: 'Peak Memory Bandwidth 864 GB/s'. 96 MB Infinity Cache. Power: Total Board Power (TBP). ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1100; Ubuntu 26.04.1/24.04.5/22.04.5, RHEL 10.2/9.8); also listed in ROCm 7.14.0 Linux GPU table. ROCm on Windows: supported (ROCm 10.1.0 matrix: Windows 11 25H2 and WSL2); HIP SDK for Windows 7.2.0 table: Runtime and HIP SDK supported. Vulkan: llama.cpp Vulkan backend (README backend table: 'Vulkan | GPU'). llama.cpp HIP backend (needs ROCm; build.md). ECC memory supported per AMD spec. Memory speed: product page lists none; AMD press release Apr 13, 2023 header 'GDDR6 ECC Memory (18GB/s)' (apparently 18 Gbps; 384 x 18 / 8 = 864 matches). Release year: press release Apr 13, 2023. Triple-slot board; AMD also lists a 'Radeon PRO W7900 Dual Slot' with identical memory specs (48 GB, 384-bit, 864 GB/s, 295 W TBP) and the same ROCm status - not given a separate record. Second source: https://ir.amd.com/news-events/press-releases/detail/1123....

Intel Arc B570 Graphics

Intel ark: 'Graphics Memory Bandwidth 380 GB/s'. Power: Total Board Power (TBP). Vulkan: Intel spec: Vulkan 1.3; llama.cpp Vulkan backend. llama.cpp SYCL backend (docs/backend/SYCL.md 'Verified devices': 'Intel Arc B-Series | Support | Arc B580'; Linux and Windows 11 supported). SYCL.md notes oneAPI 2026.1 build speed-up 'measured ... on Arc B570'. Intel OpenVINO 2025.2 release notes cite optimizations for 'Intel Arc B Series Graphics'. Intel B-series marketing cites AI Playground. Intel ark 'Launch Date Q4'24' (release_year 2024 from ark); Intel press release Dec 3, 2024 gives retail availability 'starting Jan. 16, 2025'. PCIe 4.0 x8. Second source: https://www.intel.com/content/www/us/en/newsroom/news/int....

Intel Arc B580 Graphics

Intel ark: 'Graphics Memory Bandwidth 456 GB/s'. Power: Total Board Power (TBP). Vulkan: Intel spec: Vulkan 1.3; llama.cpp Vulkan backend. llama.cpp SYCL backend (docs/backend/SYCL.md 'Verified devices': 'Intel Arc B-Series | Support | Arc B580'; Linux and Windows 11 supported). Intel OpenVINO 2025.2 release notes cite optimizations for 'Intel Arc B Series Graphics'. Intel ark 'Launch Date Q4'24'; press release Dec 3, 2024: available 'starting Dec. 13'. PCIe 4.0 x8. Only B-series card named in llama.cpp SYCL verified-device table. Second source: https://www.intel.com/content/www/us/en/newsroom/news/int....

Intel Arc Pro B50 Graphics

Intel ark: 'Graphics Memory Bandwidth 224 GB/s'. Power: Total Board Power (TBP). Vulkan: Intel spec: Vulkan 1.3; llama.cpp Vulkan backend. llama.cpp SYCL backend (docs/backend/SYCL.md 'Verified devices': 'Intel Arc B-Series | Support | Arc B580'; Linux and Windows 11 supported). Intel datasheet 'API Support: DirectX 12 Ultimate, oneAPI, OpenCL 3.0, OpenGL 4.6, OpenVINO, Vulkan 1.3'; OS support Windows 11/10 and Linux Ubuntu. Intel ark 'Launch Date Q3'25'. Datasheet: '70W Total Board Power', 'No Connector Required', dual-slot low profile. PCIe 5.0 x8. Second source: https://download.intel.com/newsroom/2025/client-computing....

Intel Arc Pro B60 Graphics

Intel ark: 'Graphics Memory Bandwidth 456 GB/s'. Power: Total Board Power (TBP). Vulkan: Intel spec: Vulkan 1.3; llama.cpp Vulkan backend. llama.cpp SYCL backend (docs/backend/SYCL.md 'Verified devices': 'Intel Arc B-Series | Support | Arc B580'; Linux and Windows 11 supported). Intel datasheet 'API Support: DirectX 12 Ultimate, oneAPI, OpenCL 3.0, OpenGL 4.6, OpenVINO, Vulkan 1.3'; 'Multi-GPU Linux Ready'; OS support Windows 11/10 and Linux Ubuntu. Intel ark 'Launch Date Q2'25'; ark 'TBP 200 W' and 'TBP - LP 120 - 200 W'; 2025 datasheet 'Consumption 120-200W Total Board Power' (partner boards vary). Sold via board partners. PCIe 5.0 x8. Second source: https://download.intel.com/newsroom/2025/client-computing....

Intel Arc Pro B65 Graphics

Intel ark and datasheet: '608 GB/s'. Memory speed not stated (608 x 8 / 256 implies 19 Gbps). Power: Total Board Power (TBP). Vulkan: Intel spec: Vulkan 1.3; llama.cpp Vulkan backend. Not named in llama.cpp SYCL verified-device table (only B580 is). Intel datasheet: 'Leverage Linux multi-GPU compatibility and OneAPI integration'; OS support Windows 11/10 and Linux Ubuntu. Additional Intel B-series card. Intel ark 'Launch Date Q1'26'; datasheet '200W Total Board Power', 'Available as Partner Branded Card'. PCIe 5.0 x16. Second source: https://www.intel.com/content/dam/www/central-libraries/u....

Intel Arc Pro B70 Graphics

Intel ark and datasheet: '608 GB/s'. Memory speed not stated (608 x 8 / 256 implies 19 Gbps). Power: Total Board Power (TBP). Vulkan: Intel spec: Vulkan 1.3; llama.cpp Vulkan backend. Not named in llama.cpp SYCL verified-device table. Intel datasheet 'API Support: DirectX 12 Ultimate, oneAPI, OpenCL 3.0, OpenGL 4.6, OpenVINO, Vulkan 1.3'; 'Scalable multiple-GPU LLM Linux support'. Intel Arc Pro B-series page benchmarks use container 'intel/llm-scaler-vllm:1.3' (vLLM). Additional Intel B-series card (32 Xe-cores). Intel ark 'Launch Date Q1'26'; ark 'TBP 230 W' and 'TBP - LP 160-290'; datasheet 'Consumption 160W-290W, 230W for Intel Branded Card'. PCIe 5.0 x16. Second source: https://www.intel.com/content/dam/www/central-libraries/u....

Apple M1 Pro

Tech specs: "200GB/s memory bandwidth"; newsroom (Oct 18, 2021): "M1 Pro offers up to 200GB/s of memory bandwidth with support for up to 32GB of unified memory." MacBook Pro 14-inch (2021) and 16-inch (2021): 16GB, configurable to 32GB (support.apple.com/en-us/111902, 111901). Laptop-only chip. GPU variants: 14-core or 16-core (CPU 8- or 10-core); same 200GB/s. gpu_cores = top variant. Shipped only in MacBook Pro, so no desktop power figure. Second source: https://support.apple.com/en-us/111901.

Apple M1 Max

Mac Studio (2022) tech specs: "400GB/s memory bandwidth" (M1 Max 10-core CPU/24-core GPU base, configurable to 32-core GPU); newsroom: "M1 Max delivers up to 400GB/s of memory bandwidth". Power: Mac Studio (2022): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14/16-inch (2021): 32GB or 64GB (support.apple.com/en-us/111902, 111901); Mac Studio (2022): 32GB, configurable to 64GB (support.apple.com/en-us/111900). GPU variants: 24-core or 32-core, both 400GB/s. gpu_cores = top variant. Apple's power figure was measured on: M1 Max 10-Core CPU & 32-Core GPU, 32GB, 2TB SSD (idle 11 W). Second source: https://support.apple.com/en-us/102027.

Apple M1 Ultra

Mac Studio (2022) tech specs: "800GB/s memory bandwidth"; newsroom (Mar 8, 2022): "Memory bandwidth is increased to 800GB/s". Power: Mac Studio (2022): Apple's maximum power for the whole computer, at the wall. Mac Studio (2022) only: 64GB, configurable to 128GB (support.apple.com/en-us/111900). GPU variants: 48-core or 64-core, both 800GB/s. gpu_cores = top variant; note Apple's power figure is for the 48-core GPU config. Apple's power figure was measured on: M1 Ultra 20-Core CPU & 48-Core GPU, 64GB, 1TB SSD (idle 13 W). Second source: https://support.apple.com/en-us/102027.

Apple M2 Pro

Tech specs: "200GB/s memory bandwidth"; newsroom (Jan 17, 2023): "M2 Pro features 40 billion transistors, 200GB/s of unified memory bandwidth, and up to 32GB of fast, low-latency unified memory." Power: Mac mini (2023): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14/16-inch (2023): 16GB, configurable to 32GB (support.apple.com/en-us/111340, 111838); Mac mini (2023): 16GB, configurable to 32GB (support.apple.com/en-us/111837). GPU variants: 16-core (10-core CPU) or 19-core (12-core CPU), both 200GB/s. gpu_cores = top variant. Apple's power table does not say which GPU variant was measured. Apple's power figure was measured on: M2 Pro, 32GB, 8TB SSD (idle 7 W). Second source: https://support.apple.com/en-us/103253.

Apple M3 Pro

Tech specs: "150GB/s memory bandwidth" (the Oct 30, 2023 newsroom gives no bandwidth figure). MacBook Pro 14-inch and 16-inch (Nov 2023): 18GB, configurable to 36GB (support.apple.com/en-us/117736, 117737). Laptop-only chip. GPU variants: 14-core (11-core CPU) or 18-core (12-core CPU), both 150GB/s. gpu_cores = top variant. Laptop-only, so no desktop power figure. Second source: https://support.apple.com/en-us/117737.

Apple M3 Max (14-core CPU, 30-core GPU)

Tech specs: "M3 Max with 14-core CPU and 30-core GPU (300GB/s memory bandwidth)". MacBook Pro 14/16-inch (Nov 2023): 36GB, or 96GB ('96GB (M3 Max with 14-core CPU)') (support.apple.com/en-us/117736, 117737). Laptop-only. Laptop-only chip, so no desktop power figure. Second source: https://support.apple.com/en-us/117737.

Apple M2 Max

Mac Studio (2023) tech specs: "400GB/s memory bandwidth"; newsroom: "M2 Max features 67 billion transistors, 400GB/s of unified memory bandwidth, and up to 96GB of fast, low-latency unified memory." Power: Mac Studio (2023): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14/16-inch (2023): 32GB, 64GB, or 96GB, where 96GB requires the 38-core GPU (support.apple.com/en-us/111340, 111838); Mac Studio (2023): 32GB, configurable to 64GB or 96GB, 96GB with the 38-core GPU (support.apple.com/en-us/111835). GPU variants: 30-core or 38-core, both 400GB/s. gpu_cores = top variant; Apple's power figure is for the 30-core config. Apple's power figure was measured on: M2 Max 12-Core CPU & 30-Core GPU, 32GB, 512GB SSD (idle 9 W). Second source: https://support.apple.com/en-us/102027.

Apple M3 Max (16-core CPU, 40-core GPU)

Tech specs: "M3 Max with 16-core CPU and 40-core GPU (400GB/s memory bandwidth)". MacBook Pro 14/16-inch (Nov 2023): 48GB, 64GB, or 128GB ('48GB, 64GB, or 128GB (M3 Max with 16-core CPU)') (support.apple.com/en-us/117736, 117737). Laptop-only. Laptop-only chip, so no desktop power figure. Second source: https://support.apple.com/en-us/117737.

Apple M2 Ultra

Tech specs: "800GB/s memory bandwidth"; newsroom (Jun 5, 2023): "M2 Ultra features 800GB/s of system memory bandwidth". Power: Mac Studio (2023): Apple's maximum power for the whole computer, at the wall. Mac Studio (2023): 64GB, configurable to 128GB or 192GB (support.apple.com/en-us/111835); Mac Pro (2023): 64GB, configurable to 128GB or 192GB (support.apple.com/en-us/111343). GPU variants: 60-core or 76-core, both 800GB/s. gpu_cores = top variant. Apple's power figure was measured on: M2 Ultra 24-Core CPU & 76-Core GPU, 192GB, 8TB SSD (idle 10 W). Mac Pro (2023) is listed at 330 W (76-core GPU, 192GB) and 300 W (60-core GPU, 64GB) at support.apple.com/en-us/102839. Second source: https://support.apple.com/en-us/102027.

Apple M4

Tech specs: "120GB/s memory bandwidth"; newsroom (Oct 30, 2024): "M4 supports up to 32GB of unified memory and has higher memory bandwidth of 120GB/s." Power: Mac mini (2024): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14-inch (M4, 2024): 16/24/32GB (support.apple.com/en-us/121552); Mac mini (2024): 16GB, configurable to 24 or 32GB (121555); iMac (24-inch, 2024): 16/24/32GB on the 10-core-GPU model, 16/24GB on the 8-core-GPU model (121557, 121556); MacBook Air (M4, 2025): 16/24/32GB (122209, 122210). iPad Pro is not included. GPU variants: 8-core or 10-core (both 120GB/s). First shipped in iPad Pro (May 2024); Macs from Oct 2024. Apple's power figure was measured on: M4, 16GB, 256GB SSD (idle 4 W). Second source: https://support.apple.com/en-us/103253.

Apple M4 Max (14-core CPU, 32-core GPU)

Tech specs: "410GB/s memory bandwidth" (M4 Max 14-core CPU, 32-core GPU). Power: Mac Studio (2025): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14/16-inch (2024): 36GB ('36GB (M4 Max with 14-core CPU)') (support.apple.com/en-us/121553, 121554); Mac Studio (2025): 36GB (support.apple.com/en-us/122211). Apple's power figure was measured on: M4 Max 14-Core CPU & 32-Core GPU, 36GB, 512GB SSD (idle 6 W). Second source: https://support.apple.com/en-us/102027.

Apple M4 Pro

Tech specs: "273GB/s memory bandwidth"; newsroom (Oct 30, 2024): "M4 Pro supports up to 64GB of fast unified memory and 273GB/s of memory bandwidth". Power: Mac mini (2024): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14/16-inch (2024): 24GB or 48GB (support.apple.com/en-us/121553, 121554); Mac mini (2024): 24GB, configurable to 48GB or 64GB (support.apple.com/en-us/121555). GPU variants: 16-core (12-core CPU) or 20-core (14-core CPU), both 273GB/s. gpu_cores = top variant. Apple's power figure was measured on: M4 Pro, 64GB, 8TB SSD (idle 5 W); Apple does not say which CPU/GPU variant. Second source: https://support.apple.com/en-us/103253.

Apple M4 Max (16-core CPU, 40-core GPU)

Tech specs: "M4 Max with 16-core CPU and 40-core GPU (546GB/s memory bandwidth)"; newsroom: "up to 546GB/s of memory bandwidth". MacBook Pro 14/16-inch (2024): 48GB, 64GB, or 128GB (M4 Max with 16-core CPU) (support.apple.com/en-us/121553, 121554); Mac Studio (2025): 48GB, 64GB, or 128GB (M4 Max with 16-core CPU and 40-core GPU) (support.apple.com/en-us/122211). power_w is null because Apple's Mac Studio (2025) power table (support.apple.com/en-us/102027) lists only the 14-core CPU/32-core GPU, 36GB config (145 W max). It gives no figure for the 40-core GPU config. Second source: https://support.apple.com/en-us/121554.

Apple M5

Tech specs: "153GB/s memory bandwidth"; newsroom (Oct 15, 2025): "M5 offers unified memory bandwidth of 153GB/s". MacBook Pro 14-inch (M5): 16/24/32GB (support.apple.com/en-us/125405); MacBook Air 13/15-inch (M5): 16GB, configurable to 24 or 32GB (126320, 126321). iPad Pro and Apple Vision Pro are not included. No desktop Mac uses M5 (Mac mini moved to M6). GPU variants: 8-core (MacBook Air 13-inch base) or 10-core. Ships only in laptops among Macs, so no desktop power figure. Apple now calls the M5 CPU '4 super cores and 6 efficiency cores'. Second source: https://www.apple.com/newsroom/2025/10/apple-unleashes-m5....

Apple M3 Ultra

Mac Studio (2025) tech specs: "819GB/s memory bandwidth" (both 28-core CPU/60-core GPU and 32-core CPU/80-core GPU); newsroom: "over 800GB/s of memory bandwidth". Power: Mac Studio (2025): Apple's maximum power for the whole computer, at the wall. Mac Studio (2025) only. The current tech specs page (support.apple.com/en-us/122211, read 2026-10-07) lists '96GB unified memory, Configurable to: 256GB'. The launch newsroom (Mar 5, 2025) says it 'starts with 96GB of unified memory, which can be configured up to 512GB'. The 512GB option no longer appears on the specs page, but Apple's power table still lists a 512GB config. GPU variants: 60-core or 80-core, both 819GB/s. gpu_cores = top variant. Apple has not announced an M4 Ultra: Mac Studio (2025) pairs M4 Max with M3 Ultra, and the 2026 Mac Studio moved to M5 Ultra. Apple's power figure was measured on: M3 Ultra 32-Core CPU & 80-Core GPU, 512GB, 16TB SSD (idle 9 W). Second source: https://support.apple.com/en-us/102027.

Apple M6

Tech specs: "153GB/s memory bandwidth, 170GB/s memory bandwidth" and memory "Configurable to: 24GB or 32GB (170GB/s memory bandwidth)". The 16GB config is 153GB/s and the 24/32GB configs are 170GB/s. Newsroom (Aug 25, 2026): "up to 170GB/s of unified memory bandwidth". Power: Mac mini (M6, 2026): Apple's maximum power for the whole computer, at the wall. Mac mini (M6, 2026): 16GB, configurable to 24GB or 32GB (support.apple.com/en-us/128108). It is the only Mac with M6 as of 2026-10-07. Not in the requested list; added because Apple announced it on Aug 25, 2026 (Mac mini). Bandwidth depends on the memory config, so 170 is the 'up to' value and 16GB configs are 153GB/s. Apple's power figure was measured on: M6, 16GB, 256GB SSD (idle 4 W). Second source: https://support.apple.com/en-us/103253.

Apple M5 Max (18-core CPU, 32-core GPU)

Tech specs: "M5 Max with 18-core CPU and 32-core GPU (460GB/s memory bandwidth)". Power: Mac Studio (M5 Max, 2026): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14/16-inch (M5 Max): 36GB ('36GB (M5 Max with 32-core GPU)') (support.apple.com/en-us/126318, 126319); Mac Studio (M5 Max, 2026): 36GB (support.apple.com/en-us/128107). Announced Mar 3, 2026 (MacBook Pro); Mac Studio with M5 Max announced Aug 25, 2026. Apple's power figure was measured on: M5 Max 18-Core CPU & 32-Core GPU, 36GB, 512GB SSD (idle 7 W). Second source: https://support.apple.com/en-us/102027.

Apple M5 Pro

Tech specs: "307GB/s memory bandwidth" (both 15-core CPU/16-core GPU and 18-core CPU/20-core GPU); newsroom (Mar 3, 2026): "M5 Pro supports up to 64GB of unified memory with higher unified memory bandwidth up to 307GB/s." Power: Mac mini (M5 Pro, 2026): Apple's maximum power for the whole computer, at the wall. MacBook Pro 14/16-inch (M5 Pro or M5 Max): 24GB, configurable to 48GB (any M5 Pro) or 64GB ('M5 Pro with 20-core GPU') (support.apple.com/en-us/126318, 126319); Mac mini (M5 Pro, 2026): 24GB, configurable to 48GB or 64GB (support.apple.com/en-us/128108). GPU variants: 16-core (15-core CPU) or 20-core (18-core CPU), both 307GB/s. gpu_cores = top variant. Announced Mar 3, 2026 (MacBook Pro); Mac mini with M5 Pro announced Aug 25, 2026 and available Sep 22, 2026. Apple's power figure was measured on: M5 Pro, 64GB, 8TB SSD (idle 6 W); Apple does not say which CPU/GPU variant. Second source: https://support.apple.com/en-us/103253.

Apple M5 Max (18-core CPU, 40-core GPU)

Tech specs: "M5 Max with 18-core CPU and 40-core GPU (614GB/s memory bandwidth)"; newsroom (Mar 3, 2026): "M5 Max supports up to 128GB of unified memory with higher unified memory bandwidth up to 614GB/s." MacBook Pro 14/16-inch (M5 Max): 48GB, 64GB, or 128GB (M5 Max with 40-core GPU) (support.apple.com/en-us/126318, 126319); Mac Studio (M5 Max, 2026): 48GB, 64GB, or 128GB (M5 Max with 18-core CPU and 40-core GPU) (support.apple.com/en-us/128107). power_w is null because Apple's Mac Studio power table (support.apple.com/en-us/102027) lists only the M5 Max 18-core CPU/32-core GPU, 36GB config (200 W max). Second source: https://support.apple.com/en-us/126319.

Apple M5 Ultra

Tech specs: "1.2TB/s memory bandwidth" (both 30-core CPU/64-core GPU and 36-core CPU/80-core GPU); newsroom (Aug 25, 2026): "1.2TB/s of unified memory bandwidth, 50 percent more than M3 Ultra". Apple states 1.2TB/s, a rounded figure, recorded here as 1200 GB/s. Power: Mac Studio (M5 Ultra, 2026): Apple's maximum power for the whole computer, at the wall. Mac Studio (M5 Ultra, 2026) only: 96GB, configurable to 256GB or 512GB, with 512GB requiring the 36-core CPU/80-core GPU (support.apple.com/en-us/128107; the configure-to-order list shows '256GB unified memory' and '512GB unified memory (M5 Ultra with 36-core CPU and 80-core GPU)'). GPU variants: 64-core (30-core CPU) or 80-core (36-core CPU). gpu_cores = top variant. Announced Aug 25, 2026; available Sep 22, 2026. Apple's power figure was measured on: M5 Ultra 36-Core CPU & 80-Core GPU, 512GB, 16TB SSD (idle 9 W). Second source: https://support.apple.com/en-us/102027.

NVIDIA RTX Spark (N1X) desktop PCs

NVIDIA states neither memory bandwidth nor bus width nor data rate for RTX Spark on the product page or the May 31, 2026 press release, so bandwidth is null. Power: Chip TDP as NVIDIA lists it (the chip only, not the whole PC). OEM Windows desktop PCs, not an NVIDIA-built box. Product page desktop spec column: 'RTX Spark N1X', GPU '6144-core NVIDIA Blackwell RTX GPU', CPU '20-core NVIDIA Grace CPU', Memory Type 'Unified', 'Power (TDP) 140 W' (superchip), Memory Configuration 'Up to 128 GB LPDDR5X'. Laptop versions are 45-80 W TDP, with a 5120-core / up to 64 GB variant. Press release (May 31, 2026): 'available this fall' from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI; product page shows 'Preorder Now'. Second source: https://nvidianews.nvidia.com/news/nvidia-microsoft-windo....

AMD Ryzen AI Max 385 (Radeon 8050S)

Derived: LPDDR5x-8000 (8000 MT/s) x 256-bit / 8 / 1000 = 256 GB/s, both inputs from the AMD product page. Power: Top of AMD's configurable TDP range (45-120 W; default 55 W): the chip only, and the computer maker sets it. ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1151; Ubuntu 26.04.1 and 24.04.5 (HWE 7.0) with inbox kernel driver; RHEL not offered for APUs). Vulkan: llama.cpp Vulkan backend. llama.cpp HIP backend; llama.cpp build.md: on Linux set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 to use UMA with an integrated GPU. AMD blogs show LM Studio (llama.cpp runtime) on Windows with Variable Graphics Memory. Product page: 256-bit LPDDR5x, Max. Memory 128 GB, LPDDR5x-8000, Radeon 8050S (32 graphics cores), 8 CPU cores, Default TDP 55W, cTDP 45-120W, Strix Halo. gpu_memory_max_gb 96 from CES 2025 press release (Ryzen AI Max series, 'up to 96GB available for graphics'). Release year: CES press release Jan 6, 2025. Linux: ROCm RDNA3.5 optimization doc says GTT (GPU-mappable system RAM) defaults to 'approximately 50 percent of total system RAM' and can be raised via the TTM page limit (example sets 100 GB on a ~128 GB system); it recommends a small BIOS VRAM reservation (e.g. 0.5 GB) plus a larger GTT limit. AMD states a range of memory sizes up to 128 GB, not a list, so only the maximum is offered here; for a smaller system, enter it as an 'Other' unified-memory device with 256 GB/s. Second source: https://ir.amd.com/news-events/press-releases/detail/1232....

AMD Ryzen AI Max 390 (Radeon 8050S)

Derived: LPDDR5x-8000 (8000 MT/s, product page 'Max Memory Speed') x 256-bit (product page 'System Memory Type 256-bit LPDDR5x') / 8 / 1000 = 256 GB/s. AMD's VGM FAQ states 256 GB/s for the Strix Halo platform generally. Power: Top of AMD's configurable TDP range (45-120 W; default 55 W): the chip only, and the computer maker sets it. ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1151; Ubuntu 26.04.1 and 24.04.5 (HWE 7.0) with inbox kernel driver; RHEL not offered for APUs). Vulkan: llama.cpp Vulkan backend. llama.cpp HIP backend; llama.cpp build.md: on Linux set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 to use UMA with an integrated GPU. AMD blogs show LM Studio (llama.cpp runtime) on Windows with Variable Graphics Memory. Product page: 256-bit LPDDR5x, Max. Memory 128 GB, LPDDR5x-8000, Radeon 8050S (32 graphics cores), 12 CPU cores, Default TDP 55W, cTDP 45-120W, Strix Halo. gpu_memory_max_gb 96 from CES 2025 press release for the Ryzen AI Max series: 'Featuring up to 128GB of unified memory with up to 96GB available for graphics'. Release year: CES press release Jan 6, 2025. Linux: ROCm RDNA3.5 optimization doc says GTT (GPU-mappable system RAM) defaults to 'approximately 50 percent of total system RAM' and can be raised via the TTM page limit (example sets 100 GB on a ~128 GB system); it recommends a small BIOS VRAM reservation (e.g. 0.5 GB) plus a larger GTT limit. AMD states a range of memory sizes up to 128 GB, not a list, so only the maximum is offered here; for a smaller system, enter it as an 'Other' unified-memory device with 256 GB/s. Second source: https://ir.amd.com/news-events/press-releases/detail/1232....

AMD Ryzen AI Max+ 395 (Radeon 8060S)

AMD blogs state 256 GB/s for Strix Halo ('This platform, codenamed "Strix Halo", doubles the bandwidth to 256 GB/s' - VGM FAQ; 'taking full advantage of the 256 GB/s bandwidth' - Max+ 395 blog). Cross-check from product page: LPDDR5x-8000 x 256-bit = 8000 x 256 / 8 / 1000 = 256 GB/s. Power: Top of AMD's configurable TDP range (45-120 W; default 55 W): the chip only, and the computer maker sets it. ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1151; Ubuntu 26.04.1 and 24.04.5 (HWE 7.0) with inbox kernel driver; RHEL not offered for APUs). Vulkan: llama.cpp Vulkan backend. llama.cpp HIP backend; llama.cpp build.md: on Linux set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 to use UMA with an integrated GPU. AMD blogs show LM Studio (llama.cpp runtime) on Windows with Variable Graphics Memory. Product page: 'System Memory Type 256-bit LPDDR5x', 'Max. Memory 128 GB', 'Max Memory Speed LPDDR5x-8000', Radeon 8060S (40 graphics cores), Default TDP 55W, cTDP 45-120W, codename Strix Halo, form factor Laptops and Desktops. Memory options: AMD blog says 'system memory options ranging from 32GB all the way up to 128GB of unified memory - out of which up to 96GB can be converted to VRAM through AMD Variable Graphics Memory' (no explicit list, so memory_options_gb null). GPU memory: CES 2025 press release 'up to 128GB of unified memory with up to 96GB available for graphics'; VGM FAQ: 'the iGPU (set to 96GB VGM) can technically access a total graphics memory size of 112GB on AMD Ryzen AI MAX+ powered 128GB systems'; on 64 GB systems the FAQ cites 'up to 48GB dedicated graphics memory through VGM'. Release year: CES press release Jan 6, 2025 (systems 'starting in Q1 2025'). AMD also sells the Ryzen AI Halo developer mini-PC with this chip (128GB). AMD additionally lists Ryzen AI Max+ 392 and 388 with the same Radeon 8060S, 128 GB max and LPDDR5x-8000 (not given separate records). Linux: ROCm RDNA3.5 optimization doc says GTT (GPU-mappable system RAM) defaults to 'approximately 50 percent of total system RAM' and can be raised via the TTM page limit (example sets 100 GB on a ~128 GB system); it recommends a small BIOS VRAM reservation (e.g. 0.5 GB) plus a larger GTT limit. AMD states a range of memory sizes up to 128 GB, not a list, so only the maximum is offered here; for a smaller system, enter it as an 'Other' unified-memory device with 256 GB/s. Second source: https://www.amd.com/en/blogs/2025/amd-ryzen-ai-max-395-pr....

NVIDIA DGX Spark

Product page spec table: 'System Memory 64 GB LPDDR5X* or 128 GB LPDDR5x, coherent unified system memory', 'Memory Interface 256-bit', 'Memory Bandwidth 273 GB/s'. Datasheet: 'Memory Bandwidth Up to 273 GB/s'. Per-pin data rate not stated. Power: Power supply / power consumption of the whole DGX Spark, as NVIDIA lists it. Unified CPU+GPU memory. Product page: 'Power Supply 240 Watts', 'GB10 TDP 140 W' (TDP of the GB10 chip incl. CPU and GPU); datasheet: 'Power Consumption 240 W'. 128 GB model shipped Oct 15, 2025 (NVIDIA newsroom, seen via search result). '*64 GB memory configuration is available exclusively through participating OEM partners' - NVIDIA blog dated October 2, 2026 (https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/): 64GB available from Friday, Oct. 23 (2026). Second source: https://dam-cdn.nvd.orangelogic.com/AssetLink/es6d60li4v5....

AMD Ryzen AI Max+ PRO 495 (Radeon 8065S)

Derived: LPDDR5x-8533 (8533 MT/s, product page 'Max Memory Speed') x 256-bit (product page) / 8 / 1000 = 273.056 GB/s. No AMD-stated bandwidth found. Power: Top of AMD's configurable TDP range (45-120 W; default 55 W): the chip only, and the computer maker sets it. ROCm on Linux: supported (ROCm 10.1.0 compatibility matrix, gfx1151; Ubuntu 26.04.1 and 24.04.5 (HWE 7.0) with inbox kernel driver; RHEL not offered for APUs). Vulkan: llama.cpp Vulkan backend. llama.cpp HIP backend; llama.cpp build.md: on Linux set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 to use UMA with an integrated GPU. AMD blogs show LM Studio (llama.cpp runtime) on Windows with Variable Graphics Memory. Newer large-memory APU (codename Gorgon Halo). Product page: 'System Memory Type 256-bit LPDDR5x', 'Max. Memory 192 GB', 'Max Memory Speed LPDDR5x-8533', Radeon 8065S (40 graphics cores), Default TDP 55W, cTDP 45-120W. AMD blog May 20, 2026: Ryzen AI Max PRO 400 Series feature 'up to 192GB of unified memory and 160GB of VRAM'; footnote GRHP-01: 'As of 5/11/2026, the AMD Ryzen AI Max+ 495 PRO processor supports up to 160 GB dedicated graphics memory'. release_year 2026 = unveiling date of that blog; product page shows no launch date, and the AMD Ryzen AI Halo page lists the 495 version as 'Available Soon'. Linux: ROCm RDNA3.5 optimization doc says GTT (GPU-mappable system RAM) defaults to 'approximately 50 percent of total system RAM' and can be raised via the TTM page limit (example sets 100 GB on a ~128 GB system); it recommends a small BIOS VRAM reservation (e.g. 0.5 GB) plus a larger GTT limit. AMD states a range of memory sizes up to 192 GB, not a list, so only the maximum is offered here; for a smaller system, enter it as an 'Other' unified-memory device with 273.1 GB/s. Second source: https://www.amd.com/en/blogs/2026/amd-powers-next-generat....

NVIDIA cards run every major engine through CUDA. AMD cards run llama.cpp through ROCm (HIP) where AMD lists the card, or through Vulkan; Intel Arc cards through SYCL or Vulkan; Macs through Metal (llama.cpp) and MLX. What AMD's and llama.cpp's own documents say:

Vulkan

README 'Supported backends' table: 'Vulkan | GPU' (vendor-neutral). Description: 'Vulkan and SYCL backend support'. Build instructions in docs/build.md#vulkan for Windows and Linux. Source.

HIP (ROCm)

README table: 'HIP | AMD GPU'. build.md: 'This provides GPU acceleration on HIP-supported AMD GPUs. Make sure to have ROCm installed.' Linux and Windows CMake examples; 'gfx1100 ... corresponds to Radeon RX 7900XTX/XT/GRE'. On Linux, GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 enables UMA for integrated GPUs. Source.

SYCL

README table: 'SYCL | Intel GPU'. SYCL.md: supports Intel Data Center Max, Flex, Arc series, built-in Arc GPU and iGPUs; 'Verified devices' table: 'Intel Arc B-Series | Support | Arc B580'; OS: Linux (Ubuntu 22.04, Fedora Silverblue 39, Arch) and Windows 11 supported. Source.

OpenVINO

README table: 'OpenVINO [In Progress] | Intel CPUs, GPUs, and NPUs'. OPENVINO.md: work in progress; 'validated specifically on AI PCs such as the Intel Core Ultra Series 1 and Series 2' (Arc B-series not specifically validated). Source.

ROCm

ROCm Core SDK 10.1.0 (released October 5, 2026) is the current ROCm release; its compatibility matrix 'Applies to Linux and Windows' and lists Radeon GPUs: RX 9070 XT/GRE/9070, RX 9060 XT LP/9060 XT/9060, RX 9050 and RX 9050 (4GB), RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT, RX 7700, RX 7600, PRO W7900 (and Dual Slot), W7800 48GB, W7800, W7700, AI PRO R9700S/R9700/R9600D/R9600, plus Ryzen AI Max/Max+ (gfx1151). RX 7600 XT is not listed. Per its OS selector, R9700S/R9700/RX 9070-series/9060 XT/9060/W7900/W7800/RX 7900-series/7800 XT offer Ubuntu, RHEL, Windows and WSL2; R9600D/R9600/9060 XT LP/9050/7700 XT/7700/7600 offer Ubuntu, RHEL and Windows (no WSL2); Ryzen AI Max offers Ubuntu, Windows, WSL2. Source.

ROCm

ROCm release history: 10.1.0 Oct 5, 2026; 10.0.0 Aug 26, 2026; 7.14.1 Sept 2, 2026; 7.14.0 July 15, 2026; 7.2.4 May 29, 2026. Two version lines are current, so a 'ROCm version' needs the stream stated. Source.

ROCm

ROCm 10.1.0 release notes: 'Since ROCm 7.14, ROCm uses TheRock as its build and release system.' 'ROCm 10.1.0 adds support for AMD Radeon AI PRO R9600 GPUs.' 'If your GPU is not listed, it might be community-enabled through TheRock nightly builds.' Resolved issue: PyTorch training/fine-tuning (Llama-Factory, Unsloth) could cause GPU resets or crashes on 'Radeon RX 9070 Series and Radeon AI PRO R9700'. Source.

ROCm

ROCm 7.14.0 Linux system requirements (page dated 2026-07-15, now marked 'This page has moved'): 'If a GPU is not listed on this table, it's not officially supported by AMD.' Radeon list excludes RX 7600, RX 7600 XT, RX 9050, AI PRO R9700S and R9600. Footnote: listed Radeon/Radeon PRO GPUs 'only support Ubuntu 24.04.4, Ubuntu 22.04.5, RHEL 10.1, and RHEL 9.7.' Source.

ROCm

HIP SDK for Windows 7.2.0 system requirements (page dated 2026-07-31): 'If a GPU is not listed on this table, it is not officially supported by AMD.' Lists Runtime+HIP SDK support for RX 9070 XT, 9070, 9070 GRE, 9060 XT, 9060, 7900 XTX, 7900 XT, 7800 XT, 7700 XT, 7650 GRE, 7600 XT, 7600, PRO W7900 (and Dual Slot), W7800, W7700, AI PRO R9700, and Ryzen AI Max+/Max 395/390/385 (and PRO), plus 'AMD Ryzen AI MAX+ 495'. RX 7900 GRE, W7800 48GB, R9700S, R9600/R9600D, RX 9060 XT LP and RX 9050 are absent. ROCm Debugger is unsupported on the RDNA 3 entries. Source.

ROCm

ROCm GPU specifications page lists VRAM 'Dynamic + carveout' for Ryzen APUs and gives RX 9070 GRE VRAM as 16 GiB, which conflicts with AMD's product page (12 GB). Source.

ROCm

ROCm RDNA3.5 system optimization (Strix Halo): GTT defaults to approximately 50 percent of system RAM; the limit can be raised via the kernel TTM page limit (amd-ttm --set 100 example); a small BIOS VRAM reservation plus a larger GTT is recommended. Source.

ROCm

llama.cpp build.md (HIP): 'If your GPU is not officially supported you can use the environment variable HSA_OVERRIDE_GFX_VERSION set to a similar GPU ... 11.0.0 on RDNA3. Note that HSA_OVERRIDE_GFX_VERSION is not supported on Windows'. That is a community workaround, not AMD support. Source.