A AMD Last verified

AMD Instinct MI355X

OAM 在售 发布于 2025 cdna4
BF16
TFLOP/s
2300 厂商声称
FP8
TFLOP/s
4600 厂商声称
FP4
TFLOP/s
9200 厂商声称
Memory
GB
288 厂商声称
Mem BW
GB/s
8000 厂商声称
TDP
W
1400 厂商声称

完整规格

算力

FP4 TFLOPS
9200
FP8 TFLOPS
4600
BF16 TFLOPS
2300
FP16 TFLOPS
2300
INT8 TOPS
4600

显存

容量
288 GB
带宽
8000 GB/s
类型
HBM3e

芯片架构 🟢 vendor floorplan

CU count
256
L2 cache
256 MB
HBM stacks
8
制程
3 nm
PCIe
Gen 5 ×16

Scale-Up (节点内)

协议
Infinity-Fabric
单链带宽
1075 GB/s
World size
8
拓扑
fully-connected
交换机

Scale-Out (节点间)

单卡出口
400 Gbps
协议
RoCEv2
NIC

拓扑示意

拓扑结构 · Topology
8 卡 scale-up domain
芯片内部 / Die-level architecture
HBM HBM HBM HBM HBM HBM HBM HBM AMD Instinct MI355X L2 / shared cache · NoC L1$ / register file (per CU) 256 CUs · darker block = tensor / matrix engine 2300 TFLOPS BF16 · 4600 FP8 · 288 GB HBM3e @ 8.0 TB/s · 1400 W TDP

🟢 vendor floorplan 256 CUs · 8× HBM · 256 MB L2 · 3 nm


集群拓扑 / Cluster topology · Infinity-Fabric @ 1075 GB/s
Infinity-Fabric switch 1075 GB/s/link · all-to-all GPU 0 288GB GPU 1 288GB GPU 2 288GB GPU 3 288GB GPU 4 288GB GPU 5 288GB GPU 6 288GB GPU 7 288GB 8 cards · fully-connected topology · scale-out: 400 Gbps/card
Scale-Up · 域内
Infinity-Fabric
1075 GB/s · 拓扑: fully-connected
world_size = 8
Scale-Out · 跨域
RoCEv2
400 Gbps/卡 NIC

能跑哪些模型?

Quick estimates · decode tok/s/card 上界

TP=8 · FP4 · batch=16 · prefill=1024 · decode=256 · 已应用 efficiency 校准

在计算器中调整 →
模型 参数 (active) Decode tok/s/card 瓶颈
DeepSeek V4 Pro
deepseek
49B 489,796 内存带宽
DeepSeek V4 Flash
deepseek
13B 682 内存带宽
Mistral Small 4
mistral
22B 311 内存带宽
GLM-5 Reasoning
zhipu
32B 257 内存带宽
GLM-5.1
zhipu
32B 175 内存带宽
Qwen3.6 Plus
alibaba
35B 167 内存带宽
Kimi K2.6
moonshot
32B 143 内存带宽
MiniMax M2.7
minimax
46B 114 内存带宽

算子级 fit · 任意模型瓶颈类型 + 上界

算子级 fit · operator-level fit (per-token roofline)

基于每个模型 operator_decomposition + 本卡 BF16 2,300 TFLOPS / 8,000 GB/s 计算 · ridge point ≈ 288 FLOPs/byte

上界 = min(计算屋顶, 内存带宽屋顶) · efficiency 未应用
模型 domain 主导算子 AI · F/B 瓶颈 tok/s 上界
DeepSeek V4 Pro llm matmul 245.5 💾 内存带宽 327k
GraphCast scientific graph-message-passing 0.9 💾 内存带宽 15k
AlphaFold 3 scientific pair-bias-attention 2.3 💾 内存带宽 4435
GPT-OSS llm matmul 0.7 💾 内存带宽 647
Gemma 4 26B llm matmul 0.7 💾 内存带宽 481
DeepSeek V4 Flash llm matmul 0.8 💾 内存带宽 455
Mistral Small 4 llm matmul 0.6 💾 内存带宽 207
Llama 4 Maverick llm matmul 0.8 💾 内存带宽 205
需要 efficiency 校准 + concurrency 扫描 + TCO 估算 → 在计算器中评估 →

算子支持 & 优化空间

算子支持 & 优化空间 / Operator support & headroom

Per-operator support derived from software_support.engines + scale-up topology. Optimization headroom from measured efficiency factor.

Optimization headroom
+-50 pp
saturated

Near saturation at 150% of roofline. Further gains require workload restructure (disaggregated, speculative, smaller batch).

Communication (collective)
All-to-All 🟢 mature
all-to-all via Infinity-Fabric world_size=8
AllReduce 🟢 mature
Infinity-Fabric ring all-reduce
Attention
Multi-Head Attention 🟢 mature
paged-attention via vLLM/SGLang/MindIE
FlashAttention-3 🟢 mature
FA-3 on modern engine + tensor cores
Matrix multiply (GEMM)
Matrix Multiplication 🟢 mature
GEMM supported on all inference engines
MoE routing
MoE Routing 🟢 mature
MoE gating supported via vLLM ≥0.4 / SGLang
Normalization
RMSNorm 🟢 mature
fused into engine kernels
Embedding
fused into engine kernels
Activation
SiLU / Swish 🟢 mature
fused into engine kernels
Softmax 🟢 mature
fused into engine kernels

软件栈支持

引擎 状态 BF16FP16FP4FP8 E4M3FP8 E5M2INT4 AWQ
HanGuangAI 未确认
LMDeploy 未确认
MindIE 未确认
MoRI 未确认
SGLang 官方
TensorRT-LLM (Dynamo) 未确认
vLLM 官方
实测校准 efficiency factor

基于 1 个该硬件的实测案例计算得出, 计算器使用此值替代默认 0.5。

1.50
measured / theoretical (n=1)

已有部署案例 (1)

引证

  1. [1] AMD MI355X announcement (vendor-claimed) — https://www.amd.com/en/products/accelerators/instinct/mi355x.html · 访问于 2026-04-28 厂商声称
  2. [2] CDNA 4 architecture: 256 CUs (4 XCDs × 64 CU configurable), 256 MB Infinity Cache, 8× HBM3e stacks @ 36 GB ⇒ 288 GB, FP4 native @ 9.2 PFLOPS, TSMC 3nm — https://www.amd.com/en/products/accelerators/instinct/mi355x.html · 访问于 2026-04-28 社区估算
⚠ All performance figures are vendor-claimed unless tier=measured.