Intel Gaudi 3 HL-338

Gaudi 3 PCIe 5nm

Overview

229.0

FP32 TFLOPS

128 GB

Memory HBM2E

3.7 TB/s

Memory Bandwidth

600 W

TDP

Performance Metrics

Precision	Peak (dense)	Bits	Type
FP64	—	—	Standard
FP32	229.0	32	Standard
TF32	459.0	32	Standard
BF16	1678.0	16	Standard
FP16	459.0	16	Standard
FP8	1678.0	8	Standard
FP6	—	—	Standard
FP4	—	—	Standard
INT8	—	—	Standard

Memory Specifications

Capacity: 128 GB
Type: HBM2E
Bandwidth: 3.7 TB/s
Interface Width: —

Power & Efficiency

TDP: 600 W
Max power (Flopper estimate): ~690 W est.
FP32 dense TFLOPS per W of TDP: 0.38 TFLOPS/W
FP16 dense TFLOPS per W of TDP: 0.77 TFLOPS/W

Flopper estimate: 600 W TDP x 1.15. The vendor publishes no maximum board power for this part. Per W of TDP figures divide dense peak by accelerator TDP. Board power only: excludes host CPUs, networking, cooling and facility overhead.

Hardware Details

Vendor: Intel

Architecture: Gaudi 3

Process Node: 5nm

Form Factor: PCIe

Launch Year: —

Variant: HL-338

Compute Engine

Tensor Processor Cores: 64

Interconnect

GPU-to-GPU: RoCE v2 RDMA over top bridge (HLTB-304A/304B), 900 GB/s per card in a 4-card quad

Bandwidth: 900 GB/s direction and scope not stated by vendor

PCIe: Gen 5 x16

Enterprise Features

Sparsity: No