← All hardware
An open, cost-effective AI accelerator priced well below NVIDIA.
Pros
- Roughly half the price of H100-class GPUs
- Open Ethernet networking, no proprietary fabric
- 60% more memory than H100 SXM
Cons
- Lower raw performance than H100/Blackwell
- Software ecosystem trails CUDA and ROCm
- Smaller deployment base and tooling
✓ Where it shines / best for
- Cost-sensitive enterprise AI training/inference
- Organizations avoiding NVIDIA lock-in
- Ethernet-based AI clusters (IBM Cloud, Dell)
✕ Not the best fit for
- Absolute peak-performance frontier training
- CUDA-dependent software stacks
- Teams needing the largest software ecosystem
Features
- ✓ 128 GB HBM2e with 3.7 TB/s bandwidth
- ✓ 1.8 PFLOPS FP8 and BF16 compute (1,835 TFLOPS BF16)
- ✓ Two chiplets, 64 tensor cores, 8 MME engines
- ✓ 96 MB on-die SRAM at 19.2 TB/s
- ✓ 24x 200 GbE RoCE ports (standard Ethernet scale-out)
- ✓ 600W TDP
- ✓ Open software (PyTorch, Hugging Face) via SynapseAI
- ✓ Available on IBM Cloud and Dell AI platforms
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| 8-chip baseboard kit (HL-325L UBB) | ~$125,000 | one-time | Intel-disclosed list price for the OAM Universal Baseboard with 8 Gaudi 3 accelerators (disclosed at Computex 2024) |
| Gaudi 3 accelerator (implied per-chip) | ~$15,625 | one-time | Implied per-chip from the $125,000 8-chip baseboard kit; ~half the price of an H100 |
| Cloud rental | Varies | per hour | Available on IBM Cloud |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| sram | 96 MB on-die, 19.2 TB/s |
| power | 600W TDP |
| memory | 128 GB HBM2e, 3.7 TB/s |
| networking | 24x 200 GbE RoCE |
| architecture | Gaudi 3 (Intel), dual-chiplet |
| fp8_bf16_performance | 1.8 PFLOPS (1,835 TFLOPS BF16) |
Sponsored
A full review is being generated for this product and will appear here shortly.