AI 加速卡规格库

NVIDIA Rubin (Vera Rubin)：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.

NVIDIA B300 (Blackwell Ultra)：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.

NVIDIA GB200 (Blackwell)：Blackwell flagship: dual B200 + Grace CPU. FP4 native. The chip every major lab is buying for 2025-2026 frontier training.

NVIDIA B200：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.

NVIDIA H200：Hopper refresh with HBM3e. The sweet spot for production inference in 2025. Extra memory means 70B-class models fit single-GPU at FP8.

NVIDIA H100：The chip that built the 2023-2024 LLM boom. Still the most-deployed AI GPU. FP8 native. 80GB HBM3.

NVIDIA A100 80GB：Workhorse of the late-2010s ML wave. Still widely used for training smaller models and inference where H100/H200 supply is tight.

NVIDIA RTX PRO 6000 Blackwell：Workstation Blackwell. 96GB GDDR7. Good fit for on-prem agent dev environments where multi-H100 rentals are overkill.

NVIDIA RTX 4090：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.

NVIDIA RTX 5090：Consumer Blackwell flagship. 32GB GDDR7. FP4 native. Strong pick for local agent dev with current-frontier features.

AMD MI455X：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.

AMD MI355X：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.

AMD Instinct MI350P：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.

AMD MI325X：AMD's answer to H200. 256GB HBM3e (highest in the catalog at single-die). Strong inference value when supply is constrained on NVIDIA.

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

Google TPU7x (Ironwood)：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.

Google TPU v6e (Trillium)：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.

Google TPU v5p：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.

Google TPU v5e：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.

AWS Trainium 2：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.

AWS Trainium3：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.

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

Apple M4 Max：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.

Cerebras WSE-3 Turbo：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.

Cerebras WSE-3：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).

Groq LPU：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.