Size Matters: The secret to better hardware acceleration for AI algorithms seems to be related to the size of actual number-crunching microchips. Broadcom's latest custom "XPU" could be the largest of these AI accelerators shown so far.
Broadcom recently hosted an "AI investor day" to showcase its latest market solutions. The networking giant introduced a series of innovations specifically designed to accelerate and manage artificial intelligence algorithms. One of the innovations was a large "XPU," which, according to Broadcom, was built on world-class foundational technologies such as SerDes and DSP.
Intel defines an XPU as any compute architecture that best fits the needs of a company's application. Whether it is a CPU, GPU, FPGA, or ASIC with custom silicon design, an XPU can theoretically provide a seamless experience across PCs, cars, robotic arms, or any other "intelligent" piece of hardware. Tech analyst Patrick Moorhead attended the Broadcom event and was able to obtain some interesting visual documentation about the company's latest products.
In one of the photos shared on Twitter, Moorhead was holding a very large piece of silicon with an "XPU" label on it. According to Frank Ostojic, head of Broadcom's custom silicon group, the XPU in question has been specifically designed for a large "consumer AI company." The XPU is equipped with two compute units at the center, while high-bandwidth memory (HBM) populates the left and right sides of the chip.
Some fun pics from the @Broadcom AI investor day today. This is obviously a wafer in its protective case. This "XPU" (AI accelerator) chip uses a 3D package. The top die is a 700mm2 accelerator on top of an 800m2 SRAM. Yes I'm surprised as:
– Patrick Moorhead (@PatrickMoorhead) March 21, 2024
A/ I don't know $AVGO had 3D chops. (It… pic.twitter.com/SjoVY90yiq
Moorhead speculates that Broadcom's XPU packs "lots and lots" of computing power, very high-speed intra-chip connectivity, and the "highest performance" for external networking communications. Nvidia's recently introduced Blackwell AI accelerator also employs two compute units embedded in the same chip, but Broadcom's XPU is as large or even larger than that.
So far, no information has been provided about the third-party "AI company" for which Broadcom designed its supposedly powerful XPU. The corporation is mostly known for its networking gear and services, but it also provides custom chips for prominent consumer organizations like Google. Like many other fabless chip designers, Broadcom will use the foundry services of Taiwanese giant TSMC to manufacture the XPU.
Custom AI accelerators are part of Broadcom's strategy to sell data center solutions that are performant, efficient, and scalable. Consumer AI applications are driving the need for innovative custom AI silicon, the company said, while a proper "AI infrastructure" would require an entire portfolio of network-related products that can move all that data as fast as possible. And Broadcom is, of course, more than willing to sell its AI portfolio to AI-focused tech companies in the future.
https://www.techspot.com/news/102382-broadcom-showcases-huge-xpu-ai-accelerator-mystery-client.html
