The Habana® team is excited to be in Dallas at SuperComputing 2022. We look forward to sharing the latest performance advances for Gaudi®2, expanded software support and partner solutions from Supermicro, Inspur and Wiwynn. Additionally, in our booth #1838 we’re displaying an expanding range of deep learning use cases. We invite you to stop by and also visit our partners’ booths at the Kay Bailey Hutchison Convention Center. Come check out the demos, presentations and tutorials of Gaudi2 Deep Learning (DL) accelerators in action.
Hot off the press, Gaudi2 DL accelerators make another performance leap on the latest MLPerf benchmark. Launched in May, Habana Gaudi2 is our second-generation DL processor, which is produced in 7nm process and features 24 tensor processor cores, 96 GB on-board HBM2E memory and 24 100 integrated Gigabit Ethernet ports. Once again Gaudi2 has shown leading eight-card server performance on the benchmark. See more details about performance results here and below.
- 10% faster vs. June Gaudi2 submission
- Added PyTorch for BERT & ResNet-50
- Maturing and expanding software driving performance advances
The MLPerf results also showed that Gaudi2 delivered the top performance in the “available” category vs. A100. Gaudi2 continued to outperform the Nvidia A100 for both BERT and ResNet-50, as it did in the previous MLPerf submission and shown below. In addition, it’s notable that Nvidia’s H100, reported in “preview,” posted ResNet-50 TTT that’s only 11% faster than the Gaudi2 performance and though the H100 is 59% faster than Gaudi2 on BERT they reported BERT TTT in the FP8 datatype, while Gaudi2 BERT TTT is using the standard, verified BF16 datatype (FP8 enablement for Gaudi2 is in our software plans). In either case, Gaudi2 is delivering price-performance improvement vs. A100, which we can expect with the H100, too.
At our booth you will also be able to experience the latest in Gaudi2 software advances, check out our text-to-image stable diffusion demo and multi-modal video retrieval demo.
Text-to-image Diffusion Demo
See Gaudi2 DL accelerators execute a text-to-image diffusion model that can spontaneously generate batches of 16 unique images in roughly 3.5 seconds. This blazing-fast deployment is executed on a single Gaudi2 DL accelerator. With this new and evolving text-to-image technology, we are making creativity more accessible to businesses and individuals, which is opening up an entirely new realm of creative possibilities. Come by and try it out yourself.
Multi-modal Video Retrieval Demo
See GaudiDL accelerators demonstrate mining knowledge from video at scale. The demo features a multi-model transformer developed by Intel Labs’ Cognitive AI team.
See You There
You can also see Habana Labs at the following events/sessions at SC’22:
- SC’22 Session: Habana’s Jiunn-yeu Chen will join a presentation with Argonne National Laboratory (ANL) on “Programming New AI Accelerators for Scientific Computing.”
- When and Where: Monday, November 14, from 1:30 – 5:00 p.m. CT., Room #D161
In this tutorial, Habana will present an overview of Habana’s Gaudi2, including both hardware and software features that have been optimized for deep learning. We present steps on how to program Gaudi systems and will provide attendees with an understanding of the key Gaudi and performance implications for scientific applications.
- When and Where: Monday, November 14, from 1:30 – 5:00 p.m. CT., Room #D161
- Supermicro Session: Don’t miss the Habana Gaudi2 presentation at Supermicro booth. Learn about the Supermicro Gaudi2 Server, which enables high-efficiency model training and inference, for more product details see here.
- When and Where: Monday, November 14, from 7:30 – 8:00 p.m. CT., Booth #1828
- Inspur Session: Join Habana’s Eitan Medina and Steve Thorne for a roundtable with Inspur to discuss the latest advances in training and inference. Learn about Inspur’s NF5688M7 Server, featuring Gaudi2, for more product details see here.
- When and Where: Tuesday, November 15, from 1:30 – 2:00 p.m. CT., Booth #2233
- Intel Session: Come learn from Eitan Medina about the latest performance advances by Habana Labs’ Gaudi2 DL accelerators along with how-to scale deep learning.
- When and Where: Tuesday, November 15, from 2:00 – 2:30 p.m. CT., Booth #2428
- DDN Session: Habana’s Steve Thorne to present at the DDN Booth. Learn about the Supermicro x12 Gaudi AI Servers, featuring Gaudi2 DL accelerators, for more product details see here.
- When and Where: Tuesday, November 15, from 2:15 – 2:45 p.m. CT., Booth #2828
- IEEE Computer Society Lunch: If you’re a student studying supercomputing or HPC, come meet up with Habana’s Jiunn-yeu Chen at the SC2022 Student Mentorship Lunch with IEEE Computer Society.
- When and Where: Wednesday, November 16, from 12:00 – 1:30 p.m. CT., Room #A306
- DDN Session: Habana’s Steve Thorne to present at the DDN Booth. Learn about the Supermicro x12 Gaudi AI Servers, featuring Gaudi2 DL accelerators, for more product details see here.
- When and Where: Wednesday, November 16, from 1:45 – 2:30 p.m. CT., Booth #2828
Want to Connect?
The Habana team would love to meet you and learn about your deep learning compute requirements. Feel free to come by our booth or drop us your info here and we’ll respond asap!
Also, since we’re here in Texas for SC’22, we invite you to join us in a round of Texas Hold ’em with us and our partners.
Come see us for more details.
Habana Labs SC’22 Schedule:
| Monday, November 14 | |
| 1:30 – 5:00pm | Programming New AI Accelerators for Scientific Computing, room #D161 |
| 7:30pm – 8:00pm | Supermicro booth #1828 |
| Tuesday, November 15 | |
| 1:00 – 2:00pm | Meet Habana Gaudi2 at Intel booth #2428 – media tour |
| 1:30 – 2:00pm | Inspur Roundtable at Inspur booth #2233 |
| 2:00 – 2:30pm | Intel booth #2428 |
| 2:15 – 2:45pm | DDN booth #2828 |
| Wednesday, November 16 | |
| 12:00 – 1:30pm | SC2022 Student Mentorship Lunch with IEEE Computer Society, room #A306 |
| 1:45 – 2:30pm | DDN booth #2828 |
