Intel Gaudi 3 HL-338

The Intel Gaudi 3 HL-338 is the PCIe add-in-card form of Gaudi 3, putting the same 5nm processor as the HL-325L OAM mezzanine into a full-height dual-slot card 10.5 inches long at a 600 W card-level TDP, two thirds the OAM card's 900 W. It carries the same eight MME engines and 64 programmable Tensor Processor Cores, the same 128 GB of HBM2E at 3.7 TB/s and the same 96 MB of on-die L2, and connects to the host over PCIe 5.0 x16 at up to 128 GB/s. Intel positions it for inference and fine-tuning rather than large-scale training, and it is the variant that fits existing servers without an OAM baseboard. Scale-up is by top bridge rather than a baseboard: the HLTB-304A and HLTB-304B boards aggregate four cards into a quad at 900 GB/s per card, and Intel documents 2x4, 1x4 and 4x1 topologies, with two quads providing 512 GB of pooled HBM2E memory. Intel publishes no dense or sparse labelling for Gaudi 3 and no sparse figure; the matrix throughput figures come from Table 5 of Intel's technical paper, which footnotes them as projected maximums.