Hot Chips’ conference included two updates from Arm, giving more insight into their newest V2 cores for HPC, ML, and next-generation cloud computing and Compute Subsystem (CSS) N2 cores for accelerated deployment of custom silicon. Therefore, they have the Veyron V1 aiming to address a data center, automotive, communications, and client computing all with the same 5nm (TSMC) RISC-V chiplet. But,one of the more interesting implementations we see is from Ventana, who is really applying the potential for flexibility that chiplets can bring. Of course, we have Zen 2/3/4 from AMD and the coming Meteor Lake products from Intel as high-visibility examples. ![]() At Yole Group, we define chiplets as a design philosophy that combines two or more discrete dies in a way that disaggregates or duplicates some or all the functions of the SoC. ![]() Of course, we have the growing momentum behind chiplet-based designs. Senior Technology and Market Analyst for Computing and Software, Semiconductor, Memory and Computing Division, Yole Intelligence (part of the Yole Group) “While the power consumption of data centers receives much more attention, the implications for increased efficiencies in personal computing can be quite broad, as Yole Group expects 182 million units of CPUs for laptop and desktop to ship in 2023, totaling US$29 billion of revenue for CPU vendors.” This method, using dynamic voltage and frequency scaling, revealed the possibility to reduce power consumption in client CPUs by 10-20%. This presentation from Efraim Rotem, Intel Fellow and Lead Power and Performance architect in client architecture systems group, showed a method for using an AI algorithm for detecting usage and consumption patterns and responding by throttling core voltage and frequency for different workload types. Next, we turn to a totally different application of AI: Intel tuning the real-time processor performance for increased optimization of CPU power consumption. This is all prescient for our current moment, where the race for generative AI hardware has driven revenues to new heights, where Yole Group expects 115% revenue growth in 2023 for server GPUs alone. The most striking was the relative performance gains achieved by advancements in number representation (~16x), complex instructions (~12.5x), process node (~2.5x), and sparsity (~2x), resulting in a 1000x performance gain in single-chip inference performance over the last 10 years. How the improvements in hardware and software have coincided to chase this growing computational load. What stood out was the parsing between computing improvements from process evolution vs from architecture and software. This presentation gave us a great tracking of the evolution of computing demand for AI acceleration. Our first highlight from Hot Chips is the keynote from Bill Dally, Chief Scientist and SVP of Research at Nvidia, titled “ Hardware for Deep Learning”. This article, written by John Lorenz, Senior Technology & Market Analyst at Yole Intelligence, part of Yole Group, is based on numerous analyses such as Status of the Processors Industry – High-end performance Packaging – Memory-Processor interface: CXL – Status of the Memory industry – Processor Quarterly Market Monitor… Hot Chips – A Yole Group’s outlook At the same time, Yole Group’s analysts have a major IPO launching, creating new opportunities for investors in the processor ecosystem. Today, Yole Group details a few of notable take-aways from the conference and share the relevant perceptions of these markets. This year’s Hot Chips conference provided once again a mixture of sharp tutorials, exciting disclosures, and timely keynotes. It is reprinted here with the permission of the Yole Group. This market research report was originally published at the Yole Group’s website.
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |