加速卡规格库

显存

44 GB

带宽

43200 TB/s

FP16

250,000

FP8

—

Announced August 18, 2026. Same 46,225 mm^2 wafer, 4 trillion transistors and 900,000 cores as WSE-3, but Cerebras doubles AI compute to 250 PFLOPS and memory bandwidth to 43.2 PB/s (shown here as 43,200 TB/s). A CS-4 rack holds three wafers for 750 PFLOPS. Compute figure is on the same basis as the 125 PFLOPS FP16 rating of WSE-3.

显存

44 GB

带宽

21000 TB/s

FP16

125,000

FP8

—

Single 46,225 mm^2 wafer. Holds entire activations on-chip; eliminates many distributed-training pains. Cerebras rates aggregate memory bandwidth at 21 PB/s (shown here as 21,000 TB/s).

显存

432 GB

带宽

23.3 TB/s

FP16

5,030

FP8

20,130

AMD's MI400-series flagship, launched at Advancing AI on July 23, 2026 with Helios racks already in production. 432GB HBM4. Per-GPU figures are MXFP4, MXFP8 and dense BF16 peaks; AMD publishes the rack numbers of 2.9 exaflops FP4, 1.4 exaflops FP8, 31TB HBM4 and 1.7 PB/s across 72 GPUs. OpenAI expects to bring Helios online starting in Q4 2026.

显存

372 GB

带宽

16 TB/s

FP16

5,000

FP8

10,000

Blackwell flagship: dual B200 + Grace CPU. FP4 native. The chip every major lab is buying for 2025-2026 frontier training.

显存

288 GB

带宽

19.2 TB/s

FP16

4,000

FP8

17,500

Blackwell successor. 288GB HBM4 per GPU. NVIDIA began production shipments in August 2026. FLOPS are NVIDIA dense figures (FP8 is FP8/FP6 training, FP4 is NVFP4 training); the 50 PFLOPS NVFP4 inference headline is a sparse figure and is not used here.

显存

288 GB

带宽

8 TB/s

FP16

2,500

FP8

5,000

Mid-cycle Blackwell refresh: 208 billion transistors across two dies, with NVIDIA quoting 1.5x the HBM3e capacity and 1.5x the dense NVFP4 rate of Blackwell. Dense figures shown; NVIDIA rates the NVL72 rack at 1,440 PFLOPS NVFP4 with sparsity and 1.1 exaFLOPS dense.

显存

180 GB

带宽

8 TB/s

FP16

2,500

FP8

5,000

Single-die Blackwell. 180GB HBM3e per GPU (1,440GB across an 8-GPU DGX B200). Inference workhorse for serving 100B+ MoE models on a single GPU.

显存

288 GB

带宽

8 TB/s

FP16

2,500

FP8

5,000

AMD's MI350-series flagship. 288GB HBM3e with native MXFP4 and MXFP6 support. 1400W typical board power. The step between MI325X and the Helios-rack MI400 series.

显存

192 GB

带宽

7.38 TB/s

FP16

2,307

FP8

4,614

Seventh-generation TPU and the first in the Ironwood family. 192 GiB HBM per chip, six times Trillium. Google positions it for large-scale training and inference of LLMs, MoE, and diffusion models.

显存

256 GB

带宽

6 TB/s

FP16

1,300

FP8

2,600

AMD's answer to H200. 256GB HBM3e (highest in the catalog at single-die). Strong inference value when supply is constrained on NVIDIA.

显存

192 GB

带宽

5.3 TB/s

FP16

1,300

FP8

2,600

Wider availability than NVIDIA flagship; cheaper to rent. ROCm software stack matures; vLLM and PyTorch first-class on MI300X in 2026.

显存

144 GB

带宽

4 TB/s

FP16

1,150

FP8

2,300

CDNA 4 in a standard PCIe slot: 128 compute units, roughly half an MI350X, with 144GB HBM3e and a configurable 450W power mode. Peak dense figures shown; AMD doubles its 8-bit and 16-bit ratings with structured sparsity. Built to drop inference capacity into existing air-cooled enterprise servers instead of liquid-cooled racks.

显存

141 GB

带宽

4.8 TB/s

FP16

989

FP8

1,979

Hopper refresh with HBM3e. The sweet spot for production inference in 2025. Extra memory means 70B-class models fit single-GPU at FP8.

显存

80 GB

带宽

3.35 TB/s

FP16

989

FP8

1,979

The chip that built the 2023-2024 LLM boom. Still the most-deployed AI GPU. FP8 native. 80GB HBM3.

显存

32 GB

带宽

1.64 TB/s

FP16

918

FP8

918

Sixth-generation TPU. Twice the per-chip BF16 compute of v5p with a third of the memory; Google used it to train Gemini 2.0.

显存

144 GB

带宽

4.9 TB/s

FP16

671

FP8

2,517

Third-generation AWS training and inference silicon. 144 GB HBM3e, 4.9 TB/s bandwidth. 8-bit and 4-bit figures are MXFP8 and MXFP4 per AWS; BF16 is roughly flat with Trainium 2. Roughly 30-50% lower cost per hour than H100/H200 on committed AWS workloads.

显存

96 GB

带宽

2.9 TB/s

FP16

667

FP8

1,299

AWS custom training/inference silicon. Dense figures shown; AWS rates it at 2,563 TFLOPS with sparsity. Anthropic uses these for Claude training. Cheaper TCO on AWS than NVIDIA flagship for committed workloads.

显存

95 GB

带宽

2.77 TB/s

FP16

459

FP8

459

Google's previous-generation training TPU, now behind Trillium and Ironwood. Used for Gemini training. Only available on GCP; pod-scale (8960 chips) is a different beast than per-chip rentals.

显存

96 GB

带宽

1.79 TB/s

FP16

360

FP8

720

Workstation Blackwell. 96GB GDDR7. Good fit for on-prem agent dev environments where multi-H100 rentals are overkill.

显存

80 GB

带宽

2 TB/s

FP16

312

FP8

—

Workhorse of the late-2010s ML wave. Still widely used for training smaller models and inference where H100/H200 supply is tight.

显存

32 GB

带宽

1.79 TB/s

FP16

209

FP8

838

Consumer Blackwell flagship. 32GB GDDR7. FP4 native. Strong pick for local agent dev with current-frontier features.

显存

16 GB

带宽

0.8 TB/s

FP16

197

FP8

393

TPU inference tier. The 8-bit figure is Google's INT8 rating (393 TOPS). Cheaper than v5p; 256-chip pods. Best fit for Google Cloud customers serving Gemini-style inference at scale.

显存

32 GB

带宽

0.82 TB/s

FP16

190

FP8

—

AWS inference silicon. Lower compute than Trainium 2 but cheaper per-token for high-volume serving workloads on AWS.

显存

0.23 GB

带宽

80 TB/s

FP16

188

FP8

—

Deterministic single-thread inference silicon. SRAM-only memory model means small per-chip capacity but massive bandwidth. Behind Groq's 700+ tokens/sec Llama 4 Scout serving.

显存

24 GB

带宽

1.01 TB/s

FP16

165

FP8

660

Best price/perf for local agent dev. 24GB VRAM fits Q4-quantized 70B models. The default consumer-tier choice for self-hosted Ollama/llama.cpp.

显存

128 GB

带宽

0.546 TB/s

FP16

38

FP8

—

128GB unified memory means a 2024-2025 Mac Studio or MacBook Pro runs 70B-class models at Q4 quantization with usable speed. The dark-horse local-inference platform. Apple publishes memory and bandwidth for Apple Silicon but no TFLOPS rating for the M5 generation, so no M5 row is listed here.