Chips · accelerators · Intel
Intel Gaudi 2 · Gaudi
720 gigaflops of dense BF16 compute per watt, at 600 W
Shipped the second quarter of 2022; shipped by the first date promised. Epoch AI names it at no AI data center and 1 GPU cluster. Intel across the site.
Embed
The frame credits Intel, Epoch AI, AI data centers and Epoch AI, Data on GPU clusters and links back to its page. Preview it.
Cite
Each figure in Intel’s own words
one chip, dense compute; memory is high-bandwidth memory- BF16, dense
- 432 TFLOPS
BF16 MME TFLOPS 432 1678
Intel Gaudi 3 AI Accelerator technical paper (July 2025, V1 Rev. 3) (read from a table with a column for each product, whose head reads “Feature/Product Intel® Gaudi® 2 AI Accelerator Intel® Gaudi® 3 AI Accelerator”) - FP8, dense
- 865 TFLOPS
FP8 MME TFLOPS 865 1678
Intel Gaudi 3 AI Accelerator technical paper (July 2025, V1 Rev. 3) (read from a table with a column for each product, whose head reads “Feature/Product Intel® Gaudi® 2 AI Accelerator Intel® Gaudi® 3 AI Accelerator”) - Memory
- 96 GB
MEMORY 96GBHBM2E TDP 600W
Intel Gaudi2 AI Accelerator HL-225H Mezzanine Card datasheet - Memory bandwidth
- 2.4 TB/s
supporting extremely high memory capacity of 96GB and total throughput of 2.4TB/s
Intel Gaudi2 AI Accelerator HL-225H Mezzanine Card datasheet - Maximum power
- 600 W
supports card level TDP of 600 watts
Intel Gaudi2 AI Accelerator HL-225H Mezzanine Card datasheet - Compute per watt
- 720 GFLOPS/Wdense BF16 over maximum power; 1,442 GFLOPS/W at FP8
The BF16 and FP8 figures come from the Gaudi 2 column of the table in Intel's Gaudi 3 paper; no Gaudi 2 page states them. Intel gives HBM bandwidth as 2.4 TB/s (datasheet), 2.45 TB/s (docs.habana.ai) and 2.46 TB/s (Gaudi 3 paper). Intel's own newsroom answers 403, so 'launched' in the 10-K is the only delivery statement found.
1 dated promise and the shipment
each date links to the page it is quoted from- 10 May 2022 Promised Supermicro Gaudi2 training server to market in the second half of 2022.
Habana is partnering with Supermicro to bring the Supermicro Gaudi®2 Training Server to market in 2H 2022
Habana Labs Launches Second-generation AI Deep Learning Processors - the second quarter of 2022 Shipped: launched in Q2 2022 (Intel's word; no readable Intel page gives a first-shipment date).
In Q2 2022, we launched our second-generation deep learning processors for training and inference: Habana Gaudi2* and Habana Greco*
Intel 2022 annual report, Form 10-K
Named at no AI data center Epoch AI tracks
H100-equivalents and power are the whole site’s today, with every chip it holds4,000 chips in 1 GPU cluster
Epoch AI’s list of AI supercomputers, which stops in July 2025| Cluster | Owner | Country | Chips | First operational |
|---|---|---|---|---|
| Intel Stability Gaudi 2 | Intel | United States of America | 4,000 | 2024-03-11 |
Epoch AI’s names for this chip: Intel Habana Gaudi2.
Other Intel chips
- Gaudi 3 1,864 gigaflops of dense BF16 compute per watt, at 900 W
- Intel’s own pages, each linked beside the figure or date it gives. Each page was read and each quote checked on 23 Oct 2026.
- Where it runs: Epoch AI, AI data centers and Epoch AI, GPU clusters, CC BY 4.0.