Pat Gelsinger’s keynote made it clear where Intel plans to innovate: AI. No surprise to anyone with a pulse, to be sure, but his enthusiasm was infectious. Every Intel employee I spoke with genuinely seemed excited to be a part of what they foresee will be the biggest turnaround in history of the technology industry. While the company still has much to accomplish, progress has been made on several key fronts.
Habana Labs is finally getting respect, and marketing help, from within Intel
OK, I could share scores of slides (yes, I have them and I’m not afraid to use them 😉 about the AIPC, Siliconomy, five process nodes in four years, the 288-core dual-die Siera Forrest, etc., etc. But I will focus instead on the core (pun untended) of Intel’s AI story. First, let’s start with the star of the show, Intel Gaudi. Yes, instead of calling it a chip from “An Intel Company”, the “Habana Labs” name was nowhere to be seen.
Instead, Gaudi, the company’s flagship AI platform, was now graduated to be an Intel product. This isn’t just a branding change. With the previously announced merger of Intel Ponte Vecchio GPU and the Habana architectures, I believe this change is meant to communicate which accelerator architectural team is in charge. It is Habana. In other words, if you use Gaudi2 or plan to adopt Gaudi3, Intel is subtly communicating that the company is not going to leave you in a lurch.
Some media reports say that Falcon Shores, the merged product expected in 2025, will be a two-die solution; one Gaudi and one Ponte Vecchio chip. I rather hope not. The market wouldn’t respond well to a repackaging; it needs to be Gaudi4 that also runs HPC (64-bit float) codes.
As for inference processing, the Habana Labs engineered Greco follow-on to the largely ignored Goya chip, appears to be dead, as I predicted 18 months ago.
Note that Intel Gaudi2 isn’t done with benchmarking. By adding support for FP8, we expect a major boost in performance when the next version of MLPerf training benchmarks are published in December. Stay tuned. We also note that Gaudi2 landed its first major generative AI Customer, Stability.AI with 4,000 nodes. Wowza! That’s a very big and important deal for Gaudi!
Intel went on to promote the Hugging Face Bridge Tower model, which combines vision and language into a single model. A live demo showed that, in this AI model, eight Gaudi2’s out-performed eight NVIDIA H100’s by over 30%. Multi-modal AI models are the next big thing, so this could be significant; or it could be an outlier.
One of the most lasting impressions of the event was the large array of demo booths for software providers. the Liftoff program touted that they are helping over 100 startups port their AI software solutions to Xeon or Gaudi. I hope AMD is taking notes and plans to invest similarly; software is AMD’s Achilles heel.
Finally, also in the software vein, Intel announced that the Intel Developer Cloud is up and running, making current and soon-to-be-release hardware platforms available to developers to experiment and develop their code. And the company announced that OpenVINO, its inference software runtime, will now support Nvidia, AMD, Arm, and RISC-V application targets.
Gaudi is for real, and now Intel is finally applying its marketing and technology muscles to optimize and fully support it. The Habana team in Israel is still in charge, and in fact could be assigned to build the GPU-ASIC bridge Intel needs to consolidate its high-end AI Training, Inference, and HPC data center offerings (speculation on my part).
No matter what, I look forward to attending this event in the future. Intel Dev Con is back!