Hot Chips 2020 to Highlight Machine Learning

Article By : Kevin Krewell

2020 Hot Chips starting next Sunday won't be held at Stanford Univ. Although it goes virtual, there’s no let up in content. Here's a skinny on what you might expect from many of the leading chip companies...

Hot Chips is held every August, and last year it opened with a record crowd in Memorial Auditorium at Stanford University. Approximately 1,200 people attended the event, which stressed the capacity of the Stanford venue. 2020 is going to be very different in format, but there’s no let up in content. For three days (Aug. 16-18), many of the leading chip companies will be streaming the latest processor innovations.

Large-scale data center machine learning

The optional Sunday (August 16) tutorials cover large scale data center machine learning deployments with presentations from Baidu, Cerebras, Google, and Nvidia. The afternoon session explores quantum computers with presentations from researchers at Facebook, Google, IBM, Intel, Microsoft, and UCSB. Both should be of interested to a wide group.

Two keynotes

The two keynotes this year include Intel’s senior vice president, chief architect, and general manager of Architecture, Graphics, and Software, Raja M. Koduri, with the title “No Transistor Left behind.”

Continuing the conference emphasis on machine learning processing, Dan Belov, distinguished engineer from DeepMind will give the second day keynote.

Server processors
The Monday session starts off with the server processors, which used to be saved for the last session of the last day for the conference finale. Of the server processor session, probably the most interesting will be the IBM POWER 10. This chip is a major redesign of the POWER architecture as it leaves behind 12nm with embedded DRAM for the more advanced Samsung Foundries 7nm EUV process. IBM will also provide an update to the mainframe series with the z15. Rounding up the server session, Intel will reveal more details of the 10nm+ Icelake-SP Xeon Scalable Server Processor and Marvell will offer more details of its forthcoming ThunderX3 Arm-Based Server Processor. All the mainstream server processors are loading up with 10’s of high-performance CPU cores with massive amounts of memory bandwidth and I/Os.

Mobile processors
The mobile processor session will pit perennial x86 competitors AMD’s 7nm Ryzen 4000 APU up against Intel’s Tiger Lake Mobile Client CPU. This has become a real horse race with AMD’s much improved mobile processors.

Monday afternoon has presentation from Chinese company Alibaba on using RISC-V for cloud and edge computing. This is one of two presentations on RISC-V designs at the conference, showing that the ecosystem is progressing. Not to left behind, Arm will talk about its high-performance Cortex-M55 microcontroller core and the Ethos ML accelerator that can be paired with it.

The day wraps up with a highly anticipated GPU session where Intel will finally reveal details of its discrete Xe GPU. It will have to take on Nvidia’s Ampere A100 supercomputer GPU and Microsoft’s Xbox Series X SoC built with AMD Radeon technology.

FPGAs
The Tuesday sessions starts off with the latest of FPGAs and reconfigurable logic with veterans Intel (Agilex) and Xilinx (Versal Premier) and newcomer Tenstorrent. The latter company is approaching machine learning applications with a chip and software that configures small processing elements for ML data flows.

SoCs, Fungible DPU
There’s a more open session on networking and distributed systems that includes an Intel/Barefoot high-performance Ethernet Switch for data centers. Alibaba has its second of three presentation on a bare-metal cloud memory expansion SoC. And then there’s the Fungible DPU – any company that names itself Fungible is going to offer something different.

Will Google drop a surprise?
The big data center machine learning training session lack Nvidia (which presented A100 in the GPU session). Google is back to talk about the TPUv2 vs. TPUv3, it seems like this is backwards looking as the company has already previewed some MLPerf numbers for the TPUv4! Maybe Google will drop a surprise. ETH Zurich has the second RISC-V presentation with a chiplet design with 4096 cores for power-efficient floating-point processing.

The most audacious design from last year’s Hot Chips was the Cerebras Wafer Scale Engine. The mega chip was a full wafer of logic and memory, squared off at 46,225mm2, with 400,000 cores and 18GB of memory. The company will be back again this year with a preview of its next generation design – expect at least a process shrink from the 16nm in the first product to 7nm.

The conference wraps up with four presentations on ML inference – three coming from Chinese companies. Including Alibaba, Baidu, and SenseTime. The most radical ML inference design is Lightmatter, which is using silicon photonics for ML acceleration.

Hot Chips has always been an event where the informal conversations around the event have been almost as important as the presentations. It is an opportunity to catch up with friends and find out the latest industry gossip. But now that the conference is virtual, most of that informal connections may be lost. But on the other hand, with three days of packed content in a mix of live video with recorded playback, it is now more accessible than ever, attendees can time shift any session. Without the cost of the event venue and food, the conference is also more affordable to engineer, professors, and students around the world. Luckily, sponsors for the conference have maintained support for the event, also helping to keep it affordable.

All three days of high-quality content is available for prices ranging from $40 to $160. While it is not free, it is still very affordable.

Leave a comment