| NVIDIA Blackwell B200Flagship |
NVIDIA (Blackwell) |
TSMC 4NP Dual-die |
192GB HBM3e |
8.0 TB/s |
FP4: 9,000 TFLOPS FP8: 4,500 TFLOPS |
NVLink 5 (1.8 TB/s) |
1,000W Liquid |
实测 |
| NVIDIA Blackwell B100Dual-Die |
NVIDIA (Blackwell) |
TSMC 4NP Dual-die |
192GB HBM3e |
8.0 TB/s |
FP4: 7,000 TFLOPS FP8: 3,500 TFLOPS |
NVLink 5 (1.8 TB/s) |
700W Air / Liquid |
规格 |
| NVIDIA GB200 SuperchipNVL72 |
Grace + 2x B200 |
TSMC 4NP + 4N |
384GB HBM3e |
16.0 TB/s |
FP4: 40 PFLOPS FP8: 20 PFLOPS |
NVLink-C2C (900 GB/s) |
2,700W Liquid |
集群 |
| NVIDIA Hopper H200141GB HBM3e |
NVIDIA (Hopper) |
TSMC 4N |
141GB HBM3e |
4.8 TB/s |
FP8: 1,979 TFLOPS FP16: 989 TFLOPS |
NVLink 4 (900 GB/s) |
700W Air / Liquid |
对比 |
| NVIDIA Hopper H100 |
NVIDIA (Hopper) |
TSMC 4N |
80GB HBM3 |
3.35 TB/s |
FP8: 1,979 TFLOPS FP16: 989 TFLOPS |
NVLink 4 (900 GB/s) |
700W Air / Liquid |
分析 |
| NVIDIA H20 (特供版)合规版 |
NVIDIA (Hopper Cut) |
TSMC 4N |
96GB HBM3 |
4.0 TB/s |
FP8: 296 TFLOPS FP16: 148 TFLOPS |
NVLink 4 (900 GB/s) |
400W Air |
实测 |
| NVIDIA L40S |
NVIDIA (Ada Lovelace) |
TSMC 4N |
48GB GDDR6 (ECC) |
864 GB/s |
FP8: 733 TFLOPS FP16: 366 TFLOPS |
PCIe Gen4 x16 (64 GB/s) |
350W Air |
规格 |
| AMD Instinct MI325X256GB HBM3e |
AMD (CDNA 3) |
TSMC 5nm + 6nm |
256GB HBM3e |
6.0 TB/s |
FP8: 2,614 TFLOPS FP16: 1,307 TFLOPS |
Infinity Fabric (896 GB/s) |
750W Liquid / Air |
实测 |
| AMD Instinct MI300XCDNA 3 |
AMD (CDNA 3) |
TSMC 5nm + 6nm |
192GB HBM3 |
5.3 TB/s |
FP8: 2,614 TFLOPS FP16: 1,307 TFLOPS |
Infinity Fabric (896 GB/s) |
750W Air / Liquid |
ROCm |
| AMD Instinct MI300AAPU |
AMD (CDNA 3 + Zen 4) |
TSMC 5nm + 6nm |
128GB Unified HBM3 |
5.3 TB/s |
FP32: 122 TFLOPS FP64: 245 TFLOPS |
Infinity Fabric Unified |
550W Liquid |
解析 |
| Huawei Ascend 910C (华为昇腾)DaVinci v3 |
Huawei (DaVinci v3) |
Domestic Advanced Node |
64GB / 128GB HBM |
1.8 TB/s |
FP16: ~320 TFLOPS INT8: ~640 TOPS |
HCCS (800 GB/s) |
600W Liquid |
全景实测 |
| Huawei Ascend 910B (华为昇腾)DaVinci v2 |
Huawei (DaVinci v2) |
Domestic Advanced Node |
64GB HBM2e |
1.2 TB/s |
FP16: 280 TFLOPS INT8: 560 TOPS |
HCCS (392 GB/s) |
400W Air / Liquid |
组网实测 |
| Cambricon Siyuan 590 (寒武纪)MLU Core |
Cambricon (MLU Core) |
7nm Node |
64GB HBM2e |
1.6 TB/s |
FP16: 320 TFLOPS FP8: 640 TFLOPS |
MLU-Link 3.0 (400 GB/s) |
350W Air / Liquid |
生态评测 |
| Biren BR100 (壁仞科技)Chiplet |
Biren (BR100 Arch) |
TSMC 7nm Chiplet |
64GB HBM2e |
2.3 TB/s |
16-bit: 256 TFLOPS 8-bit: 1,024 TOPS |
BLink (640 GB/s) |
550W Air / Liquid |
架构实测 |
| Moore Threads MTT S4000 (摩尔线程)MUSA 3.0 |
MUSA 3.0 Core |
Domestic Node |
48GB GDDR6 |
768 GB/s |
FP32: 25 TFLOPS INT8: 200 TOPS |
MT-Link (240 GB/s) |
255W Air |
转译实测 |
| Kunlunxin XPU-P800 (昆仑芯)XPU 3.0 |
Kunlun XPU 3.0 |
7nm Node |
64GB HBM |
1.8 TB/s |
FP16: 256 TFLOPS INT8: 512 TOPS |
XPU-Link (300 GB/s) |
350W Air |
TCO 测算 |
| Iluvatar Tiangai 100 (天数智芯)GPGPU |
Iluvatar CoreX GPGPU |
7nm Node |
32GB HBM2 |
1.2 TB/s |
FP16: 147 TFLOPS INT8: 295 TOPS |
PCIe 4.0 x16 |
300W Air |
选型实测 |
| Google TPU v6 (Trillium)3nm |
Google DeepMind |
TSMC 3nm |
32GB HBM |
1.6 TB/s |
BF16: 959 TFLOPS FP8: 1,918 TFLOPS |
ICI 3.2 Tbps Direct Pod |
350W Liquid / Air |
对比 |
| Google TPU v5pPod |
Google |
TSMC 5nm |
95GB HBM2e |
2.76 TB/s |
BF16: 459 TFLOPS FP8: 918 TFLOPS |
ICI 4.8 Tbps (OCS 3D Torus) |
450W Liquid |
组网 |
| AWS Trainium 2Annapurna |
Annapurna Labs |
TSMC 4nm / 5nm |
96GB HBM |
3.2 TB/s |
3x Trainium 1 Compute BF16 / FP8 Tensor |
NeuronLink-v2 (800 Gbps) |
400W Air / Liquid |
实测 |
| Microsoft Maia 100Azure |
Microsoft (Maia) |
TSMC 5nm |
64GB HBM2e |
1.8 TB/s |
BF16: ~400 TFLOPS |
Custom Backend RoCE |
500W Liquid |
液冷 |
| Groq LPU (LP-1)SRAM LPU |
Groq (TSP Arch) |
GlobalFoundries 14nm |
230MB On-chip SRAM |
80 TB/s (SRAM) |
FP16: 188 TFLOPS INT8: 750 TOPS |
Direct Mesh (Chip-to-Chip) |
300W Air |
架构 |
| Intel Gaudi 324x RoCE |
Intel (Habana Labs) |
TSMC 5nm |
128GB HBM2e |
3.7 TB/s |
BF16: 1,835 TFLOPS FP8: 1,835 TFLOPS |
24x 200GbE RoCE On-die |
900W Liquid / 600W Air |
组网 |
| Cerebras CS-3 (WSE-3)Wafer Scale |
Cerebras Systems |
TSMC 5nm (Full Wafer) |
44GB On-wafer SRAM |
21 PB/s (On-wafer) |
FP16: 125 PFLOPS 900,000 AI Cores |
1.2 Pbps Swarm Fabric |
23,000W Liquid |
解析 |
| Tenstorrent WormholeRISC-V |
Tenstorrent (Jim Keller) |
TSMC 12nm / 6nm |
24GB GDDR6 |
288 GB/s |
16-bit: 350 TFLOPS INT8: 700 TOPS |
16x 100GbE Direct Mesh |
300W Air |
速查 |