GPU specs for AI / ML
VRAM, memory bandwidth, and tensor/matrix-core compute for GPUs actually used to train and run AI models — cross-linked to which CUDA and ROCm versions each one supports.
| GPU | Released | Vendor | Category | Architecture | CUs / SMs | VRAM | Bandwidth | FP16/BF16 TFLOPS | INT8 TOPS | TDP |
|---|---|---|---|---|---|---|---|---|---|---|
| Instinct MI455X | 2026 | AMD | datacenter | CDNA 5 | 256 | 432 GB | 23,300 GB/s | 5,000 | 5,000 | — |
| Instinct MI350X | 2025 | AMD | datacenter | CDNA 4 | 256 | 288 GB | 8,000 GB/s | 2,309.6 | 4,600 | 1000 W |
| Instinct MI355X | 2025 | AMD | datacenter | CDNA 4 | 256 | 288 GB | 8,000 GB/s | 2,516.6 | 5,000 | 1400 W |
| Radeon AI PRO R9700 | 2025 | AMD | workstation | RDNA 4 | 64 | 32 GB | 640 GB/s | 191 | 383 | 300 W |
| Radeon RX 9060 XT (8GB) | 2025 | AMD | consumer | RDNA 4 | 32 | 8 GB | 320 GB/s | 103 | 205 | 150 W |
| Radeon RX 9060 XT (16GB) | 2025 | AMD | consumer | RDNA 4 | 32 | 16 GB | 320 GB/s | 103 | 205 | 160 W |
| Radeon RX 9070 | 2025 | AMD | consumer | RDNA 4 | 56 | 16 GB | 640 GB/s | 145 | 289 | 220 W |
| Radeon RX 9070 XT | 2025 | AMD | consumer | RDNA 4 | 64 | 16 GB | 640 GB/s | 195 | 389 | 304 W |
| Instinct MI325X | 2024 | AMD | datacenter | CDNA 3 | 304 | 256 GB | 6,000 GB/s | 1,307.4 | 2,600 | 1000 W |
| Radeon RX 7600 XT | 2024 | AMD | consumer | RDNA 3 | 32 | 16 GB | 288 GB/s | 45.1 | 45.1 | 190 W |
| Instinct MI300A | 2023 | AMD | datacenter | CDNA 3 | 228 | 128 GB | 5,300 GB/s | 980.6 | 1,961 | 550 W |
| Instinct MI300X | 2023 | AMD | datacenter | CDNA 3 | 304 | 192 GB | 5,325 GB/s | 1,307.4 | 2,600 | 750 W |
| Radeon RX 7700 XT | 2023 | AMD | consumer | RDNA 3 | 54 | 12 GB | 432 GB/s | 70.3 | 70.3 | 245 W |
| Radeon RX 7800 XT | 2023 | AMD | consumer | RDNA 3 | 60 | 16 GB | 624 GB/s | 74.6 | 74.6 | 263 W |
| Radeon RX 7900 GRE | 2023 | AMD | consumer | RDNA 3 | 80 | 16 GB | 576 GB/s | 92 | 92 | 260 W |
| Instinct MI210 | 2022 | AMD | datacenter | CDNA 2 | 104 | 64 GB | 1,600 GB/s | 181 | 181 | 300 W |
| Radeon RX 6650 XT | 2022 | AMD | consumer | RDNA 2 | 32 | 8 GB | 280 GB/s | — | — | 180 W |
| Radeon RX 6750 XT | 2022 | AMD | consumer | RDNA 2 | 40 | 12 GB | 432 GB/s | — | — | 250 W |
| Radeon RX 6950 XT | 2022 | AMD | consumer | RDNA 2 | 80 | 16 GB | 576 GB/s | — | — | 335 W |
| Radeon RX 7900 XT | 2022 | AMD | consumer | RDNA 3 | 84 | 20 GB | 800 GB/s | 103 | 103 | 315 W |
| Radeon RX 7900 XTX | 2022 | AMD | consumer | RDNA 3 | 96 | 24 GB | 960 GB/s | 123 | 123 | 355 W |
| Instinct MI250X | 2021 | AMD | datacenter | CDNA 2 | 220 | 128 GB | 3,276.8 GB/s | 383 | 383 | 500 W |
| Radeon RX 6600 | 2021 | AMD | consumer | RDNA 2 | 28 | 8 GB | 224 GB/s | — | — | 132 W |
| Radeon RX 6700 XT | 2021 | AMD | consumer | RDNA 2 | 40 | 12 GB | 384 GB/s | — | — | 230 W |
| Instinct MI100 | 2020 | AMD | datacenter | CDNA 1 | 120 | 32 GB | 1,200 GB/s | 184.6 | 92.3 | 300 W |
| Radeon RX 6800 | 2020 | AMD | consumer | RDNA 2 | 60 | 16 GB | 512 GB/s | — | — | 250 W |
| Radeon RX 6800 XT | 2020 | AMD | consumer | RDNA 2 | 72 | 16 GB | 512 GB/s | — | — | 300 W |
| Radeon RX 6900 XT | 2020 | AMD | consumer | RDNA 2 | 80 | 16 GB | 512 GB/s | — | — | 300 W |
| Radeon VII | 2019 | AMD | consumer | GCN 5.1 (Vega 20) | 60 | 16 GB | 1,024 GB/s | — | — | 300 W |
| Radeon Instinct MI50 | 2018 | AMD | datacenter | Vega 7nm (GCN 5.1) | 60 | 32 GB | 1,000 GB/s | — | 53.6 | 300 W |
| Radeon Instinct MI25 | 2017 | AMD | datacenter | Vega10 (GCN 5.0) | 64 | 16 GB | 484 GB/s | — | — | 300 W |
| Radeon Instinct MI6 | 2017 | AMD | datacenter | Polaris (GCN 4) | 36 | 16 GB | 224 GB/s | — | — | 150 W |
| Radeon Instinct MI8 | 2017 | AMD | datacenter | Fiji (GCN 3) | 64 | 4 GB | 512 GB/s | — | — | 175 W |
| Arc B580 | 2024 | Intel | consumer | Xe2-HPG (Battlemage) | 20 | 12 GB | 456 GB/s | — | 233 | 190 W |
| Data Center GPU Max 1550 | 2023 | Intel | datacenter | Xe-HPC (Ponte Vecchio) | 128 | 128 GB | 3,276.8 GB/s | — | — | 600 W |
| Arc A380 | 2022 | Intel | consumer | Xe-HPG (Alchemist) | 8 | 6 GB | 186 GB/s | — | 66 | 75 W |
| Arc A750 | 2022 | Intel | consumer | Xe-HPG (Alchemist) | 28 | 8 GB | 512 GB/s | — | 229 | 225 W |
| Arc A770 (16GB) | 2022 | Intel | consumer | Xe-HPG (Alchemist) | 32 | 16 GB | 560 GB/s | — | 262 | 225 W |
| B300 (Blackwell Ultra) | 2026 | NVIDIA | datacenter | Blackwell Ultra | 160 | 288 GB | 8,000 GB/s | 2,250 | — | 1400 W |
| RTX 5090 | 2025 | NVIDIA | consumer | Blackwell | 170 | 32 GB | 1,792 GB/s | 209.5 | 838 | 575 W |
| RTX PRO 6000 Blackwell | 2025 | NVIDIA | workstation | Blackwell | 188 | 96 GB | 1,792 GB/s | 250 | 1,000 | 600 W |
| B200 | 2024 | NVIDIA | datacenter | Blackwell | 148 | 180 GB | 7,700 GB/s | 2,250 | 4,500 | 1000 W |
| H200 SXM 141GB | 2024 | NVIDIA | datacenter | Hopper | 132 | 141 GB | 4,800 GB/s | 989.5 | 1,979 | 700 W |
| L40S | 2023 | NVIDIA | datacenter | Ada Lovelace | 142 | 48 GB | 864 GB/s | 362.05 | — | 350 W |
| H100 SXM5 80GB | 2022 | NVIDIA | datacenter | Hopper | 132 | 80 GB | 3,350 GB/s | 989.5 | 1,979 | 700 W |
| RTX 4090 | 2022 | NVIDIA | consumer | Ada Lovelace | 128 | 24 GB | 1,008 GB/s | 165.2 | 660.6 | 450 W |
| RTX 6000 Ada Generation | 2022 | NVIDIA | workstation | Ada Lovelace | 142 | 48 GB | 960 GB/s | 364.25 | — | 300 W |
| A100 SXM4 80GB | 2020 | NVIDIA | datacenter | Ampere | 108 | 80 GB | 2,039 GB/s | 312 | 624 | 400 W |
| RTX 3090 | 2020 | NVIDIA | consumer | Ampere | 82 | 24 GB | 936 GB/s | 71.2 | 284.7 | 350 W |
| T4 | 2018 | NVIDIA | datacenter | Turing | 40 | 16 GB | 300 GB/s | 65 | 130 | 70 W |
| TITAN RTX | 2018 | NVIDIA | consumer | Turing | 72 | 24 GB | 672 GB/s | 130 | — | 280 W |
| GTX 1080 Ti | 2017 | NVIDIA | consumer | Pascal | 28 | 11 GB | 484 GB/s | — | — | 250 W |
| TITAN V | 2017 | NVIDIA | consumer | Volta | 80 | 12 GB | 653 GB/s | 110 | — | 250 W |
| V100 SXM2 32GB | 2017 | NVIDIA | datacenter | Volta | 80 | 32 GB | 900 GB/s | 125 | — | 300 W |
| P100 SXM2 16GB | 2016 | NVIDIA | datacenter | Pascal | 56 | 16 GB | 732 GB/s | — | — | 300 W |
| TITAN X (Pascal) | 2016 | NVIDIA | consumer | Pascal | 28 | 12 GB | 480 GB/s | — | — | 250 W |
| M40 24GB | 2015 | NVIDIA | datacenter | Maxwell | 24 | 24 GB | 288 GB/s | — | — | 250 W |
| K80 | 2014 | NVIDIA | datacenter | Kepler | 26 | 24 GB | 480 GB/s | — | — | 300 W |
FP16/BF16 TFLOPS figures are dense (non-sparse) tensor/matrix-core throughput from each vendor's own spec sheets — real-world throughput depends on kernel, framework, and batch size. The CUs / SMs column counts each vendor's own top-level processing block (AMD Compute Units, NVIDIA Streaming Multiprocessors, Intel Xe-cores), which is not a like-for-like unit across vendors or generations. See each GPU's page for sourcing notes.