Hot Chips 2019 Has Never Been Hotter, or Bigger
Two contrasting approaches to AI chips emerged at Hot Chips Conference this week. While Cerebras is building the largest chip ever, other vendors are looking to see how they can disaggregate the larger dies.
There were two keynotes at this year's Hot Chips conference, one given by AMD’s CEO, Lisa Su, and the other by TSMC’s Research VP, Philip Wong. Both touched on Moore’s Law, but while Su talked about how applications need to go beyond Moore’s Law, TSMC’s Wong committed that Moore’s Law is not dead (for transistor density, at least). Su contended that leading-edge silicon costs continue to increase, which is why AMD transitioned to its chiplet strategy. The TSMC presenter didn’t talk cost, but did offer many forms of multi-die packaging solutions, seemingly re-enforcing Su’s assertions.Hot Chips is held every August, and this 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. It is difficult to determine what drove the larger audience: the merit of the materials, better recognition of the event, better promotion, or the bubble in money flowing into machine learning chips these days. Machine learning was certainly a major theme of this year’s Hot Chips event; half of the presentations were related to machine learning and neural network processing.Cerebras: Wafer-scale die
The presentation that garnered the most attention was the introduction of a wafer-scale machine learning processor from Cerebras. This was the first detailed disclosure by the company since it was founded and the presentation just scratched the surface of the technologies that are required to build a wafer scale size die. The session chair’s introduction was quite appropriate, this was the hottest chip ever introduced at Hot Chips (although not disclosed in the presentation, it has been reported that the chip dissipates 15 kW of power). The scale of the chip is hard to imagine when the die sides are about 8.5 inches. The company certainly has expanded our concept of what it means to build a “big” chip....
....MUCH MORE
Last week:
Artificial Intelligence: The World's Largest Chip Is As Big As Your Head and Contains 1.2 Trillion Transistors
