Intel Gaudi 3 is Intel's third-generation AI accelerator, with 128 GB of HBM2e, 3.7 TB/s of memory bandwidth and 1,835 TFLOPS of dense FP8 and BF16. Against the NVIDIA H100 it has more memory but slightly lower dense peak compute, and Intel's headline speedups are its own projections, not independent tests. In 2026 it is a mature but older part, and Intel's newest announced data center GPU is a different design.
TL;DR
- Specs. Gaudi 3 has 128 GB of HBM2e at 3.7 TB/s, against 80 GB at 3.35 TB/s for the H100 SXM and 141 GB at 4.8 TB/s for the H200.
- Compute. Dense FP8 is 1,835 TFLOPS on Gaudi 3 and 1,979 TFLOPS on the H100. Gaudi 3 has no sparsity support, so NVIDIA's headline 3,958 is a sparse number.
- Networking. Gaudi 3 ships with 24 x 200GbE ports on every chip and uses standard Ethernet (RoCE), not a proprietary link like NVLink.
- Claims are Intel's. Intel projected up to 1.7x faster training and about 1.3x faster inference than H100. We found no independent head-to-head.
- Verdict. Gaudi 3 is credible for teams that buy hardware, value Ethernet and have PyTorch workloads. For renting by the hour and for the widest software support, NVIDIA H100 and H200 remain the practical choice.
Spec table
Gaudi 3 figures are from The Register's launch coverage of Intel's briefing and Intel's product page. NVIDIA figures are from NVIDIA's H100 and H200 pages, where the FP8 values are published with sparsity and the dense figure is half, per NVIDIA's footnote.
| Intel Gaudi 3 | NVIDIA H100 SXM | NVIDIA H200 SXM | |
|---|---|---|---|
| FP8, dense | 1,835 TFLOPS | 1,979 TFLOPS (3,958 with sparsity) | 1,979 TFLOPS (3,958 with sparsity) |
| BF16, dense | 1,835 TFLOPS | 989.5 TFLOPS (1,979 with sparsity) | 989.5 TFLOPS (1,979 with sparsity) |
| Memory | 128 GB HBM2e | 80 GB | 141 GB HBM3e |
| Memory bandwidth | 3.7 TB/s | 3.35 TB/s | 4.8 TB/s |
| Scale-out | 24 x 200GbE RoCE per chip | NVLink 900 GB/s plus network | NVLink 900 GB/s plus network |
| Max power | 900 W (OAM), 600 W (PCIe card) | up to 700 W | up to 700 W |
| Form factors | OAM, 8-chip baseboard, PCIe card | SXM, NVL (PCIe) | SXM |
The dense BF16 values for NVIDIA are computed by halving NVIDIA's sparse figures per its footnote; The Register reports the same 989 TFLOPS number. One thing stands out: Gaudi 3 matches its FP8 and BF16 peak (1,835 TFLOPS for both), which is unusual, and it has no sparsity mode.
Architecture
Intel's product page lists three form factors: the HL-338 PCIe card (PCIe Gen5, for existing servers), the HL-325L OAM mezzanine card, and the HLB-325 baseboard that holds eight accelerators, similar in role to NVIDIA's HGX. The silicon is the same across all three.
The notable design choice is networking. Per The Register's coverage of Intel's briefing, each Gaudi 3 has 24 links of 200GbE, with 21 used for chip-to-chip traffic inside a node (about 1 TB/s) and 3 for off-node traffic. Intel's product page footnotes this as 1,200 GB/s of open-standard RoCE versus 900 GB/s of "closed" NVLink on the H100. Intel's point is that you build a Gaudi cluster from commodity Ethernet switches. NVIDIA's counter is that NVLink is faster inside a box and its stack is better tested. For the NVIDIA picture, see our guide to what NVLink is.
Intel's page names OEM partners Dell, HPE and Supermicro, a 32-node cluster reference design, and cloud availability through IBM Cloud and Denvr Dataworks.
Performance: Intel's claims
Everything here is Intel's. According to The Register's report of Intel's launch briefing:
- Training: up to 1.7x faster than H100 (the range given is 1.4x to 1.7x), strongest on smaller models like Llama2 7B and 13B.
- Inference: about 1.3x faster than H100 on average, with the lead on larger models like Llama2 70B.
- Efficiency: 1.2x to 2.3x more tokens per second per watt per card.
- Against H200, Intel expected Gaudi 3 to match or slightly trail on smaller models and lead on larger ones like Falcon 180B.
The Register notes the H100 comparisons are "projections" built on NVIDIA's published methodology and not independent benchmarks. Those claims also date from April 2024, before newer NVIDIA software releases. We did not find an independent, current, same-model comparison, so we make no claim about who is faster today. See MLCommons for submitted results.
A fair reading of the spec sheet: Gaudi 3's extra memory (128 GB vs 80 GB) helps models that barely fit on an H100, and its dense FP8 is within about 7% of H100's. The H200's 141 GB and 4.8 TB/s close most of that memory gap.
Software
Intel lists PyTorch integration, Hugging Face resources and tools for porting GPU-based models. Intel's product page does not mention vLLM, so check current support before committing. As with every non-CUDA chip, custom CUDA kernels do not carry over, and you will rely on Intel's runtime and graph compiler for performance. If your model is a standard transformer in PyTorch, porting is realistic. If your stack depends on GPU-only libraries, it is not.
Power and infrastructure
The OAM module is rated at 900 W and the PCIe card at 600 W, per The Register's account of Intel's briefing, with Intel saying real inference draw is lower. An eight-chip baseboard therefore sits in the same planning range as an 8-GPU H100 HGX system. Because networking is Ethernet, you need 200GbE switches and RoCE tuning rather than InfiniBand, which some operators prefer.
When to choose which
Gaudi 3 makes sense if
- You are buying servers from Dell, HPE or Supermicro and want an Ethernet-based cluster.
- Your workload is PyTorch, and the model is one Intel's tooling supports.
- The extra memory over H100 keeps a model on fewer cards.
H100 or H200 makes sense if
- You rent GPUs by the hour rather than buy.
- You need CUDA, or any library with a GPU-only backend.
- You want the lowest integration risk and the broadest community knowledge.
- You need a card with more memory bandwidth than Gaudi 3 (H200).
If you are unsure between the two NVIDIA parts, our guide to H100, H200, SXM, NVL and PCIe walks through the variants.
Cost
We do not publish Gaudi 3 pricing because Intel sells through OEMs and partners and prices vary by deal. For a fair comparison, benchmark your model on both systems, take measured tokens per second per accelerator, and divide by the hourly or amortized cost. For NVIDIA GPUs on Aquanode, multiply your measured throughput by the live hourly price below.
Rent today
We do not rent Gaudi 3. The NVIDIA parts it targets are live below.
See the H100 and H200 pages for specs and history.
What's next
Intel's October 2025 announcement (October 14, 2025, at the OCP Global Summit) of "Crescent Island" points to its next data center GPU: an inference-focused part on the Xe3P architecture with 160 GB of LPDDR5X memory, aimed at power- and cost-optimized air-cooled servers, with customer sampling expected in the second half of 2026 (Intel newsroom). It uses LPDDR5X, not HBM, and it is a GPU architecture, not a Gaudi successor. We did not find an Intel announcement of a Gaudi 4.
For every datacenter chip on one page, with status, memory and links to each guide, see the datacenter GPU guide.
FAQ
Is Intel Gaudi 3 better than the H100?
On paper it has more memory (128 GB vs 80 GB) and similar dense compute. Intel projects up to 1.7x faster training and 1.3x faster inference, but those are Intel's projections and we found no independent test.
How much memory does Gaudi 3 have?
128 GB of HBM2e with 3.7 TB/s of bandwidth.
Does Gaudi 3 use NVLink?
No. It uses 24 ports of 200GbE with RoCE, built on standard Ethernet.
Where can I get Gaudi 3?
Through OEMs such as Dell, HPE and Supermicro, and per Intel through cloud providers including IBM Cloud and Denvr Dataworks. Aquanode does not rent it.
Does Gaudi 3 support sparsity?
No. The Register reports Intel has no immediate plans to add it, so its 1,835 TFLOPS is a dense figure.
