Intel Gaudi 3
OAM 在售 发布于 2024 gaudi-gen3
BF16
TFLOP/s
1835 厂商声称
FP8
TFLOP/s
1835 厂商声称
FP4
TFLOP/s
不支持
Memory
GB
128 厂商声称
Mem BW
GB/s
3700 厂商声称
TDP
W
900 厂商声称
完整规格
算力
FP4 TFLOPS
不支持
FP8 TFLOPS
1835
BF16 TFLOPS
1835
FP16 TFLOPS
1835
INT8 TOPS
1835
显存
容量
128 GB
带宽
3700 GB/s
类型
HBM2e
芯片架构 🟢 vendor floorplan
Cluster count
64
HBM stacks
8
制程
5 nm
PCIe
Gen 5 ×16
Scale-Up (节点内)
协议
RoCE-v2-200GbE
单链带宽
1200 GB/s
World size
8
拓扑
all-to-all
交换机
—
Scale-Out (节点间)
单卡出口
200 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
🟢 vendor floorplan 64 Clusters · 8× HBM · 5 nm
集群拓扑 / Cluster topology · RoCE-v2-200GbE @ 1200 GB/s
Scale-Up · 域内
RoCE-v2-200GbE
1200 GB/s · 拓扑: all-to-all
world_size = 8
Scale-Out · 跨域
RoCEv2
200 Gbps/卡 NIC
能跑哪些模型?
Quick estimates · decode tok/s/card 上界
TP=8 · FP8 · batch=16 · prefill=1024 · decode=256 · 已应用 efficiency 校准
| 模型 | 参数 (active) | Decode tok/s/card | 瓶颈 |
|---|---|---|---|
| DeepSeek V4 Pro deepseek | 49B | — | 显存不足 |
| DeepSeek V4 Flash deepseek | 13B | 255 | 内存带宽 |
| Mistral Small 4 mistral | 22B | 116 | 内存带宽 |
| GLM-5 Reasoning zhipu | 32B | 96 | 内存带宽 |
| GLM-5.1 zhipu | 32B | 65 | 内存带宽 |
| Qwen3.6 Plus alibaba | 35B | 63 | 内存带宽 |
| Kimi K2.6 moonshot | 32B | 54 | 内存带宽 |
| MiniMax M2.7 minimax | 46B | 42 | 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 1,835 TFLOPS / 3,700 GB/s 计算 · ridge point ≈ 496 FLOPs/byte
| 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 | 151k |
| GraphCast | scientific | graph-message-passing | 0.9 | 💾 内存带宽 | 6827 |
| AlphaFold 3 | scientific | pair-bias-attention | 2.3 | 💾 内存带宽 | 2051 |
| GPT-OSS | llm | matmul | 0.7 | 💾 内存带宽 | 299 |
| Gemma 4 26B | llm | matmul | 0.7 | 💾 内存带宽 | 222 |
| DeepSeek V4 Flash | llm | matmul | 0.8 | 💾 内存带宽 | 210 |
| Mistral Small 4 | llm | matmul | 0.6 | 💾 内存带宽 | 96 |
| Llama 4 Maverick | llm | matmul | 0.8 | 💾 内存带宽 | 95 |
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
+-21 pp
saturated
Near saturation at 121% of roofline. Further gains require workload restructure (disaggregated, speculative, smaller batch).
Communication (collective)
All-to-All 🟢 mature
all-to-all via RoCE-v2-200GbE world_size=8
AllReduce 🟢 mature
RoCE-v2-200GbE 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
Rotary Position Embedding 🟢 mature
fused into engine kernels
Activation
SiLU / Swish 🟢 mature
fused into engine kernels
Softmax 🟢 mature
fused into engine kernels
最接近的替代卡 (按规格相似度)
基于 BF16 算力 / 显存 / 显存带宽 / FP8 加权欧氏距离。供选型决策参考。
软件栈支持
| 引擎 | 状态 | BF16 | FP16 | FP4 | FP8 E4M3 | FP8 E5M2 | INT4 AWQ |
|---|---|---|---|---|---|---|---|
| HanGuangAI | 未确认 | — | — | — | — | — | — |
| LMDeploy | 未确认 | — | — | — | — | — | — |
| MindIE | 未确认 | — | — | — | — | — | — |
| MoRI | 未确认 | — | — | — | — | — | — |
| SGLang | 未确认 | — | — | — | — | — | — |
| TensorRT-LLM (Dynamo) | 未确认 | — | — | — | — | — | — |
| vLLM | 社区 | ✓ | ✓ | — | ✓ | — | — |
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。
1.21
measured / theoretical (n=1)
已有部署案例 (1)
引证
- [1] Intel Gaudi 3 product page — https://www.intel.com/content/www/us/en/products/details/processors/ai-accelerators/gaudi3.html · 访问于 2026-04-28 厂商声称
- [2] Gaudi 3: dual-die package, 64 TPCs total + 8 MMEs, 8× HBM2e ⇒ 128 GB; 24× 200 GbE on-chip RoCE; TSMC 5nm — https://habana.ai/products/gaudi3/ · 访问于 2026-04-28 厂商声称
⚠ All performance figures are vendor-claimed unless tier=measured.