LLM Hardware Spec Sheet: GPUs, Apple Silicon & Domestic Accelerators
Memory Capacity (GB) determines maximum model weight size; Memory Bandwidth (GB/s) governs single-stream token generation speed.
Consumer GPUs (NVIDIA / AMD)
| Hardware | VRAM | Bandwidth | FP16 | TDP | Spec Source | Models |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 5090 32GB | 32 GB | 1792 GB/s | 419 TFLOPS | 575W | NVIDIA Blackwell GeForce Official Specs (GDDR7 512-bit) | View → |
| NVIDIA GeForce RTX 5080 16GB | 16 GB | 960 GB/s | 225 TFLOPS | 360W | NVIDIA Blackwell GeForce Official Specs (GDDR7 256-bit) | View → |
| NVIDIA GeForce RTX 5070 Ti 16GB | 16 GB | 896 GB/s | 175 TFLOPS | 300W | NVIDIA Blackwell GeForce Official Specs (GDDR7 256-bit) | View → |
| NVIDIA GeForce RTX 5070 12GB | 12 GB | 672 GB/s | 123 TFLOPS | 250W | NVIDIA Blackwell GeForce Official Specs (GDDR7 192-bit) | View → |
| NVIDIA GeForce RTX 4090 24GB | 24 GB | 1008 GB/s | 330 TFLOPS | 450W | NVIDIA Ada Lovelace Official Specs (GDDR6X 384-bit) | View → |
| NVIDIA GeForce RTX 4070 Ti Super 16GB | 16 GB | 672 GB/s | 176 TFLOPS | 285W | NVIDIA Official Specs | View → |
| NVIDIA GeForce RTX 4060 Ti 16GB | 16 GB | 288 GB/s | 88 TFLOPS | 165W | NVIDIA Official Specs (GDDR6 128-bit) | View → |
| NVIDIA GeForce RTX 4060 8GB | 8 GB | 272 GB/s | 60 TFLOPS | 115W | NVIDIA Official Specs | View → |
| NVIDIA GeForce RTX 3090 24GB | 24 GB | 936 GB/s | 142 TFLOPS | 350W | NVIDIA Ampere Official Specs (二手参考均价) | View → |
| NVIDIA GeForce RTX 3060 12GB | 12 GB | 360 GB/s | 51 TFLOPS | 170W | NVIDIA Ampere Official Specs | View → |
| AMD Radeon RX 7900 XTX 24GB | 24 GB | 960 GB/s | 123 TFLOPS | 355W | AMD Official Specs | View → |
| AMD Radeon RX 9070 XT 16GB | 16 GB | 640 GB/s | 97 TFLOPS | 304W | AMD RDNA 4 Official Specs | View → |
Apple Silicon (M5 / M4 / M3)
| Hardware | VRAM | Bandwidth | FP16 | TDP | Spec Source | Models |
|---|---|---|---|---|---|---|
| Apple Mac mini M5 Pro (64GB Unified) | 64 GB | 307 GB/s | 22 TFLOPS | 65W | Apple Official Tech Specs (307 GB/s Unified Memory) | View → |
| Apple Mac Studio M5 Max (128GB Unified) | 128 GB | 614 GB/s | 42 TFLOPS | 120W | Apple Official Tech Specs (40-core GPU, 614 GB/s) | View → |
| Apple Mac Studio M5 Ultra (512GB Unified) | 512 GB | 1200 GB/s | 84 TFLOPS | 190W | Apple Official Tech Specs (1.2 TB/s UltraFusion) | View → |
| Apple Mac mini M4 (16GB Unified) | 16 GB | 120 GB/s | 8.5 TFLOPS | 40W | Apple Official Tech Specs (120 GB/s) | View → |
| Apple Mac mini M4 (24GB Unified) | 24 GB | 120 GB/s | 8.5 TFLOPS | 40W | Apple Official Tech Specs (120 GB/s) | View → |
| Apple Mac mini M4 Pro (48GB Unified) | 48 GB | 273 GB/s | 18.4 TFLOPS | 65W | Apple Official Tech Specs (273 GB/s) | View → |
| Apple Mac Studio M4 Max (128GB Unified) | 128 GB | 546 GB/s | 36.8 TFLOPS | 115W | Apple Official Tech Specs (546 GB/s) | View → |
| Apple Mac Studio M3 Ultra (512GB Unified) | 512 GB | 819 GB/s | 54.2 TFLOPS | 170W | Apple Official Tech Specs (819 GB/s) | View → |
Chinese Domestic AI Accelerators
| Hardware | VRAM | Bandwidth | FP16 | TDP | Spec Source | Models |
|---|---|---|---|---|---|---|
| 华为昇腾 Ascend 910B 64GB (HBM2e) | 64 GB | 1200 GB/s | 376 TFLOPS | 310W | 昇腾 910B2 主流规格(64GB HBM2e,约 1.2 TB/s 带宽) | View → |
| 海光深算 DCU K100-AI 64GB (HBM2e) | 64 GB | 896 GB/s | 196 TFLOPS | 350W | 海光深算 K100-AI 公开规格(64GB HBM2e,896 GB/s,兼容 ROCm) | View → |
| 摩尔线程 MTT S4000 48GB (GDDR6) | 48 GB | 768 GB/s | 100 TFLOPS | 450W | 摩尔线程官网规格(48GB GDDR6 384-bit,768 GB/s,MUSA 架构) | View → |
| 摩尔线程 MTT S80 16GB | 16 GB | 448 GB/s | 28.8 TFLOPS | 255W | 摩尔线程官网消费级显卡规格(16GB GDDR6,448 GB/s) | View → |
Unified-Memory AI Mini PCs
| Hardware | VRAM | Bandwidth | FP16 | TDP | Spec Source | Models |
|---|---|---|---|---|---|---|
| AMD Ryzen AI Max+ 395 Mini PC (128GB Unified) | 128 GB | 256 GB/s | 59 TFLOPS | 120W | AMD Strix Halo LPDDR5X-8000 256-bit Official Specs | View → |
| NVIDIA DGX Spark GB10 (128GB Unified) | 128 GB | 273 GB/s | 250 TFLOPS | 170W | NVIDIA DGX Spark Official Specs | View → |
Datacenter & Workstation GPUs
| Hardware | VRAM | Bandwidth | FP16 | TDP | Spec Source | Models |
|---|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | 1792 GB/s | 500 TFLOPS | 600W | NVIDIA Workstation Blackwell Official Specs | View → |
| NVIDIA L40S 48GB | 48 GB | 864 GB/s | 362 TFLOPS | 350W | NVIDIA Datacenter Official Datasheet | View → |
| NVIDIA A100 SXM 80GB | 80 GB | 2039 GB/s | 312 TFLOPS | 400W | NVIDIA A100 Tensor Core Official Datasheet | View → |
| NVIDIA H100 SXM 80GB | 80 GB | 3350 GB/s | 989 TFLOPS | 700W | NVIDIA H100 Hopper Official Datasheet | View → |
| NVIDIA H20 96GB (HBM3) | 96 GB | 4000 GB/s | 148 TFLOPS | 400W | NVIDIA H20 96GB HBM3 (4.0 TB/s) Specs | View → |