| 规格 | AMD Instinct MI300A | AMD Instinct MI100 |
|---|
| 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) | CDNA / gfx908 (CDNA1 世代开山, TSMC 7nm FinFET) |
| 显存 | 128GB HBM3 (CPU+GPU统一共享内存, Full-Chip ECC, Infinity Cache Yes) | 32GB HBM2 (Full-Chip ECC, 4 stacks) |
| 显存带宽 | 5.3 TB/s | 1.2 TB/s (官方页口径; ROCm 架构文档 1.228 TB/s @1.2GHz) |
| 显存位宽 | 8192-bit (Memory Clock 5.2 GHz) | 4096bit (4 stacks × 1024bit) |
| 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) | FP64 11.5 / FP32 23.1 / FP32 Matrix 46.1 / FP16 184.6 / BF16 92.3 / INT8·INT4 92.3 TOPS (TFLOPS, Peak Engine Clock 1502MHz) |
| 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) | 120 CU / 7,680 流处理器 (8 Shader Engine × 15 CU, Matrix Core) |
| 功耗 | 550W (TDP) | 760W Peak | 300W Peak (TDP, 被动散热) |
| 互联 | Infinity Fabric 8链 (峰值128GB/s/链) / PCIe 5.0 x16 | Infinity Fabric 3链 (峰值92GB/s/链, 2-4卡hive) / PCIe 4.0 x16 (兼容 3.0 x16) |
| 发布年 | 2023 (官方Launch Date 12/06/2023, Advancing AI大会与MI300X同场) | 2020 (官方 Launch Date 11/16/2020) |
| 形态 | APU SH5 Socket (非MI300X的OAM模组, 冷却Passive & Active) | PCIe Add-in Card (全高全长双槽 10.5"/267mm, 被动散热) |