| 规格 | AMD Instinct MI300A | AMD Instinct MI325X |
|---|
| 架构 | CDNA3 APU融合封装 (Zen4 EPYC 24核CPU@峰值3700MHz + CDNA3 GPU, TSMC 5nm|6nm FinFET, 1460亿晶体管) | CDNA3 |
| 显存 | 128GB HBM3 (CPU+GPU统一共享内存, Full-Chip ECC, Infinity Cache Yes) | 256 GB |
| 显存带宽 | 5.3 TB/s | 6 TB/s |
| 显存位宽 | 8192-bit (Memory Clock 5.2 GHz) | 8192-bit |
| 算力 | FP64 61.3 / FP64 Matrix 122.6 / FP32 122.6 / TF32 Matrix 490.3 / FP16·BF16 980.6 / FP8 1.96 PF / INT8 1.96 POPs (TFLOPS, Peak Engine Clock 2100MHz; FP8/FP16/INT8稀疏翻倍) | — |
| 计算单元 | GPU 228 CU / 14592 流处理器 / 912 Matrix Cores + CPU 24核 Zen4 (任务brief预设304CU系MI300X数字, MI300A官方页228为准) | — |
| 功耗 | 550W (TDP) | 760W Peak | — |
| 互联 | Infinity Fabric 8链 (峰值128GB/s/链) / PCIe 5.0 x16 | PCIe® 5.0 x16 |
| 发布年 | 2023 (官方Launch Date 12/06/2023, Advancing AI大会与MI300X同场) | — |
| 形态 | APU SH5 Socket (非MI300X的OAM模组, 冷却Passive & Active) | OAM Module |
| 显存类型 | — | HBM3E |
| FP8算力 | — | 2.61 PFLOPs |
| FP16算力 | — | 1.3 PFLOPs |
| FP32算力 | — | 163.4 TFLOPs |
| FP64算力 | — | 81.7 TFLOPs |
| INT8算力 | — | 2.6 POPs |
| 功耗TBP | — | 1000W Peak |
| Infinity Fabric Links | — | 8 |
| Infinity Fabric带宽 | — | 128 GB/s |
| 流处理器 | — | 19,456 |
| 末级缓存LLC | — | 256 MB |
| 制程 | — | TSMC 5nm | 6nm FinFET |