16 accelerators from 4 vendors, with specifications taken from vendor product pages and datasheets. Pick any two to compare them side by side.
NVIDIA datacenter GPUs are fabbed via Samsung Electronics, GlobalFoundries and Taiwan Semiconductor Mfg (ADR) — see the supply-chain graph for foundry → NVDA edges.
| CMP | SKU | VENDOR | ARCH | MEM | BW | TDP | FP16 | DOCS |
|---|---|---|---|---|---|---|---|---|
AMD Radeon RX 7900 XTX RX-7900-XTX Consumer GPU reference; not datacenter AI but competes with GeForce. | amd | RDNA3 | 24 GB GDDR6 | 960 GB/s | 355 W | — | [PAGE] | |
AMD Instinct MI300X MI300X-OAM-192GB CDNA3 accelerator; 192GB HBM3; primary NVDA H100 competitor. | amd | CDNA3 | 192 GB HBM3 | 5,300 GB/s | 750 W | — | [PAGE] | |
AMD Instinct MI325X MI325X-OAM-256GB HBM3e refresh of MI300X; 256GB capacity. | amd | CDNA3 | 256 GB HBM3e | 6,000 GB/s | 1000 W | — | [PAGE] | |
AMD Instinct MI355X MI355X-OAM-288GB CDNA4 generation; targets Blackwell B200 class workloads. | amd | CDNA4 | 288 GB HBM3e | 8,000 GB/s | 1400 W | — | [PAGE] | |
Huawei Ascend 910B ASCEND-910B Primary China-market AI training chip; specs from Huawei Ascend product documentation. | huawei | Ascend | 64 GB HBM2e | — | 400 W | 320 TFLOPS | [PAGE][ARCH] | |
Huawei Ascend 910C ASCEND-910C Successor Ascend generation; public spec sheet limited outside China — verify before production use. | huawei | Ascend | — | — | — | — | [PAGE][ARCH] | |
NVIDIA GeForce RTX 5090 RTX-5090 Consumer Blackwell flagship; 32GB GDDR7. | nvidia | Blackwell | 32 GB GDDR7 | — | 575 W | — | [PAGE][ARCH] | |
NVIDIA B200 SXM B200-SXM-192GB Blackwell datacenter GPU; FP4 native; 5th-gen NVLink. | nvidia | Blackwell | 192 GB HBM3e | 8,000 GB/s | 1000 W | — | [PAGE][ARCH] | |
NVIDIA GB200 Grace Blackwell Superchip GB200-SUPERCHIP 2x B200 GPU + Grace CPU; aggregate 384GB HBM3e per superchip module. | nvidia | Blackwell | 384 GB HBM3e | — | 2700 W | — | [PAGE] | |
NVIDIA H100 PCIe H100-PCIe-80GB PCIe FHFL dual-slot; 350W default TGP mode. | nvidia | Hopper | 80 GB HBM2e | 2,000 GB/s | 350 W | — | [PAGE][PDF] | |
NVIDIA H100 SXM5 H100-SXM5-80GB SXM5 variant; FP8/FP16 from NVIDIA product spec table (sparse peak). | nvidia | Hopper | 80 GB HBM3 | 3,350 GB/s | 700 W | 1,979 TFLOPS | [PAGE][PDF] | |
NVIDIA H200 SXM H200-SXM-141GB Hopper refresh with HBM3e; ~1.4x memory capacity vs H100. | nvidia | Hopper | 141 GB HBM3e | 4,800 GB/s | 700 W | — | [PAGE] | |
NVIDIA L40S L40S-48GB Inference/visualization GPU; Ada architecture, not Hopper/Blackwell. | nvidia | Ada Lovelace | 48 GB GDDR6 | 864 GB/s | 350 W | — | [PAGE] | |
NVIDIA RTX PRO 6000 Blackwell RTX-PRO-6000-BLACKWELL Workstation Blackwell; 96GB GDDR7 ECC. | nvidia | Blackwell | 96 GB GDDR7 | 1,792 GB/s | 600 W | — | [PAGE] | |
Samsung HBM3E 12-Hi Stack SAMSUNG-HBM3E-12H Memory component used by NVDA/AMD accelerators — supply-chain competition, not a compute GPU. | samsung | HBM3E | 36 GB HBM3E | — | — | — | [PAGE] | |
Samsung Mach-1 AI Accelerator MACH-1-AI Inference-focused logic chip roadmap; Samsung also competes as HBM supplier (see supply chain). Not a direct GeForce/datacenter GPU SKU. | samsung | Mach-1 | — | — | — | — | [PAGE][ARCH] |
Memory capacity and bandwidth usually decide which model fits on a single accelerator, and TDP decides what a rack can hold — which is why those are the columns shown first rather than peak throughput. Where a vendor doesn't publish a figure it is shown as absent rather than estimated.