Infrastructure atlas
Intel/Gaudi accelerator

Gaudi 3

An AI accelerator designed around HBM capacity, integrated RoCE networking, and an open software stack for training and inference.

ASICGauditraininginferenceRoCE
Gaudi 3 hardware

01 / Memory

128 GB HBM2e

Per accelerator

02 / Bandwidth

3.67 TB/s

HBM bandwidth

03 / Compute

1,678 TFLOPS

FP8 / BF16 matrix, Intel specification

Precision & peak compute

PrecisionPublished peakBasis
FP8 matrix1,678 TFLOPSIntel Gaudi 3 specification
BF16 matrix1,678 TFLOPSIntel Gaudi 3 specification

Peak FLOPS and TOPS are theoretical hardware figures. Sparse and dense modes, chip and system totals, and form factors are not interchangeable.

Intel Gaudi 3 product page

Technical specification

Silicon & memory
Architecture
Gaudi 3
HBM
128 GB HBM2e · 3.67 TB/s
FP8 matrix
1,678 TFLOPS
BF16 matrix
1,678 TFLOPS
Memory hierarchy
HBM plus 96 MB local SRAM
System & software
Networking
Integrated high-speed Ethernet / RoCE fabric
Software
Intel Gaudi software, PyTorch, SynapseAI
Deployment
PCIe cards and OEM server platforms
Best fit
Training and inference with open networking