I Intel Last verified

Intel Gaudi 2

OAM In production Released 2022 gaudi-gen2
BF16
TFLOP/s
432 厂商声称
FP8
TFLOP/s
865 厂商声称
FP4
TFLOP/s
unsupported
Memory
GB
96 厂商声称
Mem BW
GB/s
2450 厂商声称
TDP
W
600 厂商声称

Full specs

Compute

FP4 TFLOPS
unsupported
FP8 TFLOPS
865
BF16 TFLOPS
432
FP16 TFLOPS
432
INT8 TOPS
865

Memory

Capacity
96 GB
Bandwidth
2450 GB/s
Type
HBM2e

Die architecture 🟢 vendor floorplan

Cluster count
24
HBM stacks
6
Process
7 nm
PCIe
Gen 4 ×16

Scale-Up (intra-node)

Protocol
RoCE-v2-100GbE
Per-link BW
600 GB/s
World size
8
Topology
all-to-all
Switch
—

Scale-Out (inter-node)

Per-card NIC
100 Gbps
Protocol
RoCEv2
NIC
—

Topology

拓扑结构 · Topology
8 卡 scale-up domain
芯片内部 / Die-level architecture
HBM HBM HBM HBM HBM HBM Intel Gaudi 2 L2 / shared cache · NoC L1$ / register file (per Cluster) 24 Clusters · darker block = tensor / matrix engine 432 TFLOPS BF16 · 865 FP8 · 96 GB HBM2e @ 2.5 TB/s · 600 W TDP

🟢 vendor floorplan 24 Clusters · 6× HBM · 7 nm

集群拓扑 / Cluster topology · RoCE-v2-100GbE @ 600 GB/s
RoCE-v2-100GbE switch 600 GB/s/link · all-to-all GPU 0 96GB GPU 1 96GB GPU 2 96GB GPU 3 96GB GPU 4 96GB GPU 5 96GB GPU 6 96GB GPU 7 96GB 8 cards · all-to-all topology · scale-out: 100 Gbps/card
Scale-Up · 域内
RoCE-v2-100GbE
600 GB/s · 拓扑: all-to-all
world_size = 8
Scale-Out · 跨域
RoCEv2
100 Gbps/卡 NIC

Which models can it run?

Quick estimates · decode tok/s/card 上界

TP=8 · FP8 · 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 — 显存不足
DeepSeek V4 Flash
deepseek
 13B 70 内存带宽
Mistral Small 4
mistral
 22B 32 内存带宽
GLM-5 Reasoning
zhipu
 32B 26 内存带宽
GLM-5.1
zhipu
 32B — 显存不足
Qwen3.6 Plus
alibaba
 35B 17 内存带宽
Kimi K2.6
moonshot
 32B — 显存不足
MiniMax M2.7
minimax
 46B 12 内存带宽

Operator-level fit · per-model bottleneck + upper bound

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

基于每个模型 operator_decomposition + 本卡 BF16 432 TFLOPS / 2,450 GB/s 计算 · ridge point ≈ 176 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 🔥 计算 72k
GraphCast scientific graph-message-passing 0.9 💾 内存带宽 4520
AlphaFold 3 scientific pair-bias-attention 2.3 💾 内存带宽 1358
GPT-OSS llm matmul 0.7 💾 内存带宽 198
Gemma 4 26B llm matmul 0.7 💾 内存带宽 147
DeepSeek V4 Flash llm matmul 0.8 💾 内存带宽 139
Mistral Small 4 llm matmul 0.6 💾 内存带宽 63
Llama 4 Maverick llm matmul 0.8 💾 内存带宽 63
需要 efficiency 校准 + concurrency 扫描 + TCO 估算 → 在计算器中评估 →

Operator support & optimization headroom

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
moderate

No cases yet — using default 0.5 efficiency. Real headroom unknown until first measurement lands.

Communication (collective)
All-to-All 🟢 mature
all-to-all via RoCE-v2-100GbE world_size=8
AllReduce 🟢 mature
RoCE-v2-100GbE 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。

AWS Trainium 2
AWS Annapurna Labs
100%
相似度
BF16
650
Mem
96
BW
2.9
Google TPU v5p
Google
100%
相似度
BF16
459
Mem
95
BW
2.8
海光 DCU K100
海光信息
100%
相似度
BF16
192
Mem
96
BW
2.4
⇄ 国产 替代
NVIDIA A100 SXM4 80GB
NVIDIA
99%
相似度
BF16
312
Mem
80
BW
2.0

Software-stack support

Engine Status BF16FP16FP4FP8 E4M3FP8 E5M2INT4 AWQ
HanGuangAI unconfirmed — — — — — —
LMDeploy unconfirmed — — — — — —
MindIE unconfirmed — — — — — —
MoRI unconfirmed — — — — — —
SGLang unconfirmed — — — — — —
TensorRT-LLM (Dynamo) unconfirmed — — — — — —
vLLM community ✓ ✓ — ✓ — —

Existing deployment cases (0)

No measured cases yet for this card. Be the first contributor?

Citations

  1. [1] Intel Gaudi 2 product page — https://www.intel.com/content/www/us/en/products/details/processors/ai-accelerators/gaudi2.html · accessed 2026-04-28 厂商声称
  2. [2] Gaudi 2: 24 Tensor Processor Cores (TPCs) + matrix engines, 6× HBM2e ⇒ 96 GB; 24× 100 GbE on-chip RoCE NICs; TSMC 7nm — https://habana.ai/products/gaudi2/ · accessed 2026-04-28 厂商声称
⚠ All performance figures are vendor-claimed unless tier=measured.