Every year, I promise not to attend this tech-heavy confab of fast chips. And every year, I break that promise. Just too much to ignore!
I can deal with it: some technology is just beyond my pay grade. My masters in CS is decades old, and the deep technology underneath blindingly fast AI chips is often beyond my comprehension. But when I saw the lineup of companies presenting at this years Hot Chips conference, I relented and planned to absorb as much as my addled brain could handle. I am so glad I took the plunge. In this wrap-up, I will share my impressions, and links to a half-dozen other blogs I have penned for more in-depth analysis.
In the interest of brevity, I will suspend with my normal prose and images, and just provide you with a bullet list of items you might find of interest, with links to over a dozen relayed posts if you want more details. As usual, I will focus more on the “So What?” than the “What”.
- Intel provided presentations of Saphire Rapids, the next generation of Xeon CPU’s, Alder Lake for client platforms, and most interestingly the upcoming Ponte Vecchio GPU, planned for the ANL Supercomputer. My Take: Ponte Vecchio will give NVIDIA a run for the money for HPC, and looks to have decent AI performance. But NVIDIA’s AI position remains unthreatened given their massive lead in the AI ecosystem. Sapphire Rapids will continue to position Xeon as the leader in data center inference processing in terms of volume.
- AMD presented details about the Zen 3 core. My Take: the Zen 3 core and faster EPYC chip are going to take an additional 10 points of share from Intel Xeon.
- IBM presented the next-gen Z “Telum” CPU, with an on-die dedicated AI accelerator and innovative new cache and fabric. My take: Z customers have one less reason to leave this reliable enterprise platform and even more reasons to move new work onto the ‘Frame.
- Simple Machines provided more details on the Mozart AI chip. My take: these guys have promise in inference processing, but still a long way to go, especially in software.
- Synopsys Co-CEO Aart de Geus provided the first Keynote Address, explaining how the company’s journey into using AI to improved chip design is just beginning. My Take: Synopsys has lapped the competition and is probably 2 years ahead of anyone else.
- NVIDIA once again took the stage to explain why they see the need for massive processing on a NIC with the NVIDIA DPU. My take: Jensen Huang’s vision far outstrips the competition, and our ability to predict the future.
- Intel followed NVIDIA DPU session with their IPU. My take: Yawn. NVIDIA DPU will be far more interesting for HPC and AI clients.
- Esperanto, the startup that has been developing an AI chip using RISC-V for over 5 years, started by Dave Ditzel, and now led by CEO Art Swift, finally has real silicon. With an innovative architecture using over 1000 RISC-V cores, the company claims superior performance per watt over NVIDIA. My Take: This is very interesting. Let’s wait for production silicon, MLPerf, and more details on software. The ability to write kernels in C for a RISC core could be a winner.
- Qualcomm Technologies shared a lot more details about the Cloud AI100, with 14 TOPs/Watt. My Take: If anyone can take on NVIDIA for cloud inference processing, it is this company, team and architecture.
- Still with me??? ….. just checking 😉
- Graphcore Colossus Mk2 IPU looks like a cool and fast AI accelerator for a composable AI infrastructure. My Take: It is taking a long time to get traction, but the architecture looks solid and the company’s software is also ahead of the startup pack. I suspect there are large customers using this who don’t want to talk about it. Patience is a virtue.
- Cerebras, the inventor of the Wafer Scale AI Engine, announced new hardware and software that enable training of a 120-trillion parameter model on their second generation platform. This would be a first. To do it, they invented a new memory and fabric technology as well new approach to sparsity and weight streaming. My Take: If you want to beat NVIDIA, you must do something completely different. Cerebras fits the bill with room to spare. See my analysis of this announcement here.
- Samba Nova, the other Silicon Valley Unicorn, presented more info on their AI platform. My Take: Nothing new here. Move along.
- Xilinx presented their next generation ACAP for edge processing, the Versal Edge acceleration Swiss Army knife. My Take: Versal is still coming. Again. Patience is a virtue.
Ok, I’m exhausted. Hope I didn’t exhaust you, too! Several of these technologies merit in-depth research reports which can be found on my website, as well as related blogs found in the above embedded links.