Intel Gaudi 3: The Complete Guide
The first chip in the line to carry Intel’s own name — made by its biggest foundry rival, TSMC. Simple to expert, every number links to its source.
Stuck at any point? Ask an AI about this page →The first chip in the line to carry Intel’s own name — made by its biggest foundry rival, TSMC. Simple to expert, every number links to its source.
Stuck at any point? Ask an AI about this page →Gaudi 3 is Intel’s current AI accelerator, unveiled 9 April 2024 with systems reaching customers from late 2024 1 2. It comes as an OAM module and a PCIe card, with 128 GB of HBM2e memory — and unlike NVIDIA’s and AMD’s flagship accelerators, it is made by Intel’s own biggest foundry rival, TSMC, on a 5 nm process 3.
Intel’s own 2024 claims, stated as projections rather than measured results: Gaudi 3 is “projected to deliver 50% faster time-to-train on average” across specific named models (Llama 2 7B/13B and GPT-3 175B), and inference throughput is “projected to outperform the H100 by 50% on average and 40% for inference power-efficiency”, averaged across Llama 7B/70B and Falcon 180B 1; separately, Intel claims up to twice the price-performance on Llama 2 70B inference 2. These are comparisons to the H100, NVIDIA’s 2022 chip, not to its newer Blackwell or Rubin generations (see our NVIDIA guide on this site).
| Spec | Gaudi 3 (OAM) | Gaudi 3 PCIe (HL-338) |
|---|---|---|
| Memory | 128 GB HBM2e 3 | 128 GB HBM2e 3 |
| Memory bandwidth | 3.7 TB/s 3 | 3.7 TB/s |
| AI compute | 1,678 TFLOPS BF16/FP8 (white paper) 3; 1,835 TFLOPS (Hot Chips slides) 6 | Not stated separately |
| Power | 900 W air 3; 1,200 W liquid, per Intel’s Hot Chips slides 6 | 600 W 3 |
| Made on | TSMC 5 nm 3 | TSMC 5 nm |
Intel publishes two peak figures: 1,678 TFLOPS in its white paper 3, and 1,835 TFLOPS in its Hot Chips 2024 slides, shown next to the 1,200 W liquid-cooled rating 6. Neither says whether the figure is dense or sparse.
Designed by Intel’s Habana Labs team in Israel; manufactured by TSMC on a 5 nm process — Intel’s own AI accelerator, made by the foundry it competes against 3.
Intel has not published which of TSMC’s fabs makes Gaudi 3. This is a contrast with Intel’s Panther Lake and Xeon 6+ processors, which Intel makes in its own Arizona fab on its Intel 18A process (see the Intel brand guide on this site).
Unlike NVIDIA and AMD, Intel has published a list price for its AI accelerator. In June 2024 it said a kit of eight Gaudi 3 chips on a baseboard “will list at $125,000” — about $15,625 per chip by our arithmetic 14.
Rental prices are inconsistent across providers. Denvr Dataworks was “pre-selling now” with no published per-chip rate as of our check 15; IBM Cloud offers a quote tool only 16; Intel’s own Tiber AI Cloud pricing page lists no Gaudi hourly rate 17. India’s IndiaAI Mission compute portal lists one Gaudi 3 card (128 GB) at ₹153.00 an hour and a full 8-card server at ₹1,224, before GST — the same rate the portal lists for one NVIDIA H100 SXM card 18.
Intel’s own 2024 claims, stated as projections: on average 50% faster time-to-train and 50% better inference throughput than NVIDIA’s H100, 40% better inference power efficiency (each figure averaged across specific named models, not a general claim), and up to twice the price-performance on Llama 2 70B inference 1 2.
These are Intel’s own comparisons, against NVIDIA’s H100 — a 2022 chip, not NVIDIA’s newest — and Intel’s own release describes them as “projected” figures, not measured benchmark results. The 50%/50%/40% figures are averages across named models only (Llama 2 7B/13B and GPT-3 175B for training; Llama 7B/70B and Falcon 180B for inference), not a blanket claim across all workloads 1. We found no equivalent Gaudi 3 comparison against NVIDIA’s newer Blackwell (B200/B300) or Rubin chips in the sources checked for this page (see the NVIDIA guide on this site for those chips’ own figures).
Intel said in January 2025 that Falcon Shores, the chip once expected to follow Gaudi 3, would be used only as an internal test chip “without bringing it to market,” shifting focus to Jaguar Shores, a rack-scale system, instead 19. Unlike Gaudi 3 itself — which went from an April 2024 unveiling to shipping systems within months — no successor accelerator had reached general availability as of this page’s last check, making Gaudi 3 Intel’s current accelerator for longer than any earlier Gaudi generation held that position.
19 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.
Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.
Nothing is sent from here. The link carries only this page’s title and address.