Brand: Intel | Category: GPUs
SKU: HLDK-325B | Part #: HLDK-325B | MPN: HLDK-325B
Contact for Pricing — Request a Quote
The Intel Gaudi 3 HL-338 (part number HLDK-325B) is an AI accelerator in a full-height, dual-slot PCIe add-in card form factor, 10.5 in. long. Built on 5 nm process technology, it carries 128 GB of HBM2e memory with 3.7 TB/s of memory bandwidth and 96 MB of on-die SRAM. Compute is provided by 8 matrix engines (MME) and 64 Tensor Processor Cores (TPC), with support for FP8, BF16, FP16, TF32 and FP32 data types.
The card connects to the host over PCIe 5.0 x16 (up to 128 GB/s) and has a card TDP of 600 W. For scale-up, RoCE v2 RDMA links up to 4 cards at 900 GB/s via the HLTB-304A / HLTB-304B top bridges. An integrated media processor handles image and video decoding and pre-processing.
Supporting PyTorch, vLLM and Hugging Face, the HL-338 suits AI inference and fine-tuning servers in data centre environments.
Send the part number through our RFQ form and our team will reply within 48 hours.
| Brand | Intel |
| Category | GPUs |
| SKU | HLDK-325B |
| Part Number | HLDK-325B |
| Condition | New |
| Model | Intel Gaudi 3 HL-338 PCIe Add-In Card |
| Manufacturer Part Number | HLDK-325B |
| Process technology | 5 nm |
| Memory | 128 GB HBM2e |
| Memory bandwidth | 3.7 TB/s |
| On-die SRAM | 96 MB |
| Matrix engines (MME) | 8 |
| Tensor Processor Cores (TPC) | 64 |
| Host interface | PCIe 5.0 x16 (up to 128 GB/s) |
| Form factor | Full-height, dual-slot PCIe card, 10.5 in. long |
| Card TDP | 600 W |
| Scale-up | RoCE v2 RDMA via top bridges HLTB-304A / HLTB-304B (4 cards, 900 GB/s) |
| Data types | FP8, BF16, FP16, TF32, FP32 |
| Media | Integrated media processor for image and video decoding and pre-processing |
| Software | PyTorch, vLLM, Hugging Face |
Reference servers include Dell PowerEdge XE9680 / XE9712, HPE Cray XD670, Lenovo ThinkSystem SR685a / SR675 V3, Supermicro AS-A21GE / SYS-821GE, Gigabyte G593 / G894, ASUS ESC. Share your target platform in the RFQ and we will confirm chassis-to-GPU compatibility and recommended NIC pairing.
Yes — Omnixon supplies the full NVIDIA networking lineup (Quantum-2 / Quantum-X InfiniBand, Spectrum-X Ethernet, ConnectX NICs, BlueField DPUs) so we can quote a complete training-cluster BOM, not just the GPUs.
