AI 加速卡规格库
本页收录 26 款 AI 加速卡的公开规格,按厂商与发布顺序排列。FP16/FP8/FP4 均为厂商公布的密集(dense)算力,未采用稀疏(sparse)宣传值;标价为空表示厂商未公开官方定价,需向渠道询价。如需按算力挑选,可关注 Cerebras WSE-3 Turbo(FP16 250,000 TFLOPS)。
| # | 型号 | 厂商 | 家族 | 制程 | 发布 | 显存 | 带宽 | FP16 (TFLOPS) | FP8 (TFLOPS) | FP4 (TFLOPS) | 功耗 | 互联 | 标价 | 官网 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | NVIDIA Rubin (Vera Rubin) | NVIDIA | Rubin | Not disclosed by NVIDIA | 2026-08 | 288 GB | 19.2 TB/s | 4,000 | 17,500 | 35,000 | — | NVLink 6 (3 TB/s per GPU) + Vera CPU NVLink-C2C (1.8 TB/s) | 未公开 | 规格页 |
| 2 | NVIDIA B300 (Blackwell Ultra) | NVIDIA | Blackwell Ultra | TSMC 4NP | 2025 | 288 GB | 8 TB/s | 2,500 | 5,000 | 15,000 | 1400 W | NVLink 5 (1.8 TB/s) | 未公开 | 规格页 |
| 3 | NVIDIA GB200 (Blackwell) | NVIDIA | Blackwell | TSMC 4NP | 2025 | 372 GB | 16 TB/s | 5,000 | 10,000 | 20,000 | 2700 W | NVLink 5 (1.8 TB/s) + Grace CPU C2C (900 GB/s) | 未公开 | 规格页 |
| 4 | NVIDIA B200 | NVIDIA | Blackwell | TSMC 4NP | 2025 | 180 GB | 8 TB/s | 2,500 | 5,000 | 10,000 | 1000 W | NVLink 5 (1.8 TB/s) | 未公开 | 规格页 |
| 5 | NVIDIA H200 | NVIDIA | Hopper | TSMC 4N | 2024 | 141 GB | 4.8 TB/s | 989 | 1,979 | — | 700 W | NVLink 4 (900 GB/s) | 未公开 | 规格页 |
| 6 | NVIDIA H100 | NVIDIA | Hopper | TSMC 4N | 2022 | 80 GB | 3.35 TB/s | 989 | 1,979 | — | 700 W | NVLink 4 (900 GB/s) | $30000 | 规格页 |
| 7 | NVIDIA A100 80GB | NVIDIA | Ampere | TSMC 7nm | 2020 | 80 GB | 2 TB/s | 312 | — | — | 400 W | NVLink 3 (600 GB/s) | $18000 | 规格页 |
| 8 | NVIDIA RTX PRO 6000 Blackwell | NVIDIA | Blackwell | TSMC 4NP | 2025 | 96 GB | 1.79 TB/s | 360 | 720 | 1,440 | 600 W | PCIe 5.0 | $16000 | 规格页 |
| 9 | NVIDIA RTX 4090 | NVIDIA | Ada Lovelace | TSMC 4N | 2022 | 24 GB | 1.01 TB/s | 165 | 660 | — | 450 W | PCIe 4.0 | $1600 | 规格页 |
| 10 | NVIDIA RTX 5090 | NVIDIA | Blackwell | TSMC 4NP | 2025 | 32 GB | 1.79 TB/s | 209 | 838 | 3,352 | 575 W | PCIe 5.0 | $2000 | 规格页 |
| 11 | AMD MI455X | AMD | Instinct | Not disclosed by AMD | 2026-07 | 432 GB | 23.3 TB/s | 5,030 | 20,130 | 40,260 | — | UALink over Ethernet (3.6 TB/s bidirectional scale-up per GPU) plus 2,400 Gb/s scale-out | 未公开 | 规格页 |
| 12 | AMD MI355X | AMD | Instinct | TSMC N3 / N6 | 2025-06 | 288 GB | 8 TB/s | 2,500 | 5,000 | 10,100 | 1400 W | Infinity Fabric (7x 153 GB/s scale-up + 1x 128 GB/s scale-out) | 未公开 | 规格页 |
| 13 | AMD Instinct MI350P | AMD | Instinct | TSMC N3 / N6 | 2026-05 | 144 GB | 4 TB/s | 1,150 | 2,300 | 4,600 | 600 W | PCIe 5.0 x16 (no GPU-to-GPU Infinity Fabric links) | 未公开 | 规格页 |
| 14 | AMD MI325X | AMD | Instinct | TSMC N5 / N6 | 2024-Q4 | 256 GB | 6 TB/s | 1,300 | 2,600 | — | 1000 W | Infinity Fabric (8x 153 GB/s) | 未公开 | 规格页 |
| 15 | AMD MI300X | AMD | Instinct | TSMC N5 / N6 | 2023-Q4 | 192 GB | 5.3 TB/s | 1,300 | 2,600 | — | 750 W | Infinity Fabric | $15000 | 规格页 |
| 16 | Google TPU7x (Ironwood) | TPU | Not disclosed by Google | 2026-03 | 192 GB | 7.38 TB/s | 2,307 | 4,614 | — | — | ICI (1.2 TB/s bidirectional per chip), 9,216-chip pods | 未公开 | 规格页 | |
| 17 | Google TPU v6e (Trillium) | TPU | Not disclosed by Google | 2024-12 | 32 GB | 1.64 TB/s | 918 | 918 | — | — | ICI (800 GB/s bidirectional per chip), 256-chip pods | 未公开 | 规格页 | |
| 18 | Google TPU v5p | TPU | TSMC 3nm | 2023-12 | 95 GB | 2.77 TB/s | 459 | 459 | — | 700 W | ICI (Inter-Chip Interconnect, 1.2 TB/s bidirectional per chip) | 未公开 | 规格页 | |
| 19 | Google TPU v5e | TPU | TSMC 5nm | 2023-08 | 16 GB | 0.8 TB/s | 197 | 393 | — | 170 W | ICI (400 GB/s bidirectional per chip) | 未公开 | 规格页 | |
| 20 | AWS Trainium 2 | AWS | Trainium | TSMC 5nm | 2024-12 | 96 GB | 2.9 TB/s | 667 | 1,299 | — | 500 W | NeuronLink-v3 (1.28 TB/s per chip) | 未公开 | 规格页 |
| 21 | AWS Trainium3 | AWS | Trainium | TSMC 3nm | 2025-12 | 144 GB | 4.9 TB/s | 671 | 2,517 | 2,517 | — | NeuronLink-v4 (2.56 TB/s per chip) | 未公开 | 规格页 |
| 22 | AWS Inferentia 2 | AWS | Inferentia | TSMC 5nm | 2023-04 | 32 GB | 0.82 TB/s | 190 | — | — | — | NeuronLink | 未公开 | 规格页 |
| 23 | Apple M4 Max | Apple | Apple Silicon | TSMC 3nm | 2024-10 | 128 GB | 0.55 TB/s | 38 | — | — | 60 W | On-die (unified memory) | $4699 | 规格页 |
| 24 | Cerebras WSE-3 Turbo | Cerebras | WSE | Not disclosed by Cerebras | 2026-08 | 44 GB | 43200 TB/s | 250,000 | — | — | — | On-wafer fabric (53.5 PB/s), 2.4 Tb/s off-wafer I/O | 未公开 | 规格页 |
| 25 | Cerebras WSE-3 | Cerebras | WSE | TSMC 5nm | 2024 | 44 GB | 21000 TB/s | 125,000 | — | — | 23000 W | On-wafer (no interconnect needed) | 未公开 | 规格页 |
| 26 | Groq LPU | Groq | LPU | GlobalFoundries 14nm (gen 1) | 2023 | 0 GB | 80 TB/s | 188 | — | — | 350 W | GroqLink | 未公开 | 规格页 |
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 at 1,792 GB/s, 600W max board power per NVIDIA. Good fit for on-prem agent dev environments where multi-H100 rentals are overkill. NVIDIA listed the Workstation Edition at about 8,565 dollars for single units at its March 2025 launch, raised that to 13,250 dollars in June 2026 and to 16,000 dollars in August 2026 as GDDR7 memory costs climbed; NVIDIA has not explained the increases. The prior URL on this entry returned HTTP 404 and was replaced with the current NVIDIA product page.
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. Still the newest TPU listed here: Google announced the eighth generation (TPU 8t for training, TPU 8i for inference) at Cloud Next on April 22, 2026 and added TPU 8tx and TPU 8i Max on November 9, 2026, but its TPU page still lists them as coming soon and it publishes no per-chip BF16 rating for any of the four, so none has a row yet. Google states 288 GB of HBM, 384 MB of SRAM and 19.2 Tb/s of ICI bandwidth for TPU 8i, says TPU 8i Max merges two 8i chips to double memory, compute and bandwidth, and says TPU 8tx was designed with AMD for confidential training and inference.
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. NVIDIA bought Groq's inference technology and assets in December 2025 (CNBC reported about 20 billion dollars; neither company confirmed the price) and said on August 24, 2026 that the successor NVIDIA Groq 3 LPX rack is in full production, with Nebius the first named deployer. NVIDIA publishes per-LPU figures for that part of 500 MB SRAM, 150 TB/s SRAM bandwidth and 2.5 TB/s scale-up bandwidth, plus 256 LPUs, 128 GB SRAM, 40 PB/s and 315 PFLOPS FP8 per rack, but no FP16 or BF16 rating, so it has no row of its own here yet. The prior URL on this entry returned HTTP 404 and was replaced with Groq's platform page.
同步频率:每 3 小时同步一次,规格变更后自动更新,无需人工维护。
口径说明:显存带宽单位 TB/s,算力单位 TFLOPS;TPU/LPU 等非 GPU 架构的算力口径与 NVIDIA 不完全可比,跨厂商对比时请以显存容量与带宽为主。
免责:本页为公开资料汇总,仅供选型参考,最终规格以厂商官方文档为准。