Measurement Issues: The key to higher {hardware} acceleration for AI algorithms appears to be associated to the scale of precise number-crunching microchips. Broadcom’s newest customized “XPU” may very well be the most important of those AI accelerators proven thus far.
Broadcom lately hosted an “AI investor day” to showcase its newest market options. The networking large launched a sequence of improvements particularly designed to speed up and handle synthetic intelligence algorithms. One of many improvements was a big “XPU,” which, in keeping with Broadcom, was constructed on world-class foundational applied sciences similar to SerDes and DSP.
Intel defines an XPU as any compute structure that most closely fits the wants of an organization’s utility. Whether or not it’s a CPU, GPU, FPGA, or ASIC with customized silicon design, an XPU can theoretically present a seamless expertise throughout PCs, vehicles, robotic arms, or some other “clever” piece of {hardware}. Tech analyst Patrick Moorhead attended the Broadcom occasion and was capable of receive some attention-grabbing visible documentation in regards to the firm’s newest merchandise.
In one of many photos shared on Twitter, Moorhead was holding a really giant piece of silicon with an “XPU” label on it. In keeping with Frank Ostojic, head of Broadcom’s customized silicon group, the XPU in query has been particularly designed for a big “shopper AI firm.” The XPU is provided with two compute models on the heart, whereas high-bandwidth reminiscence (HBM) populates the left and proper sides of the chip.
Some enjoyable pics from the @Broadcom AI investor day right now. That is clearly a wafer in its protecting case. This “XPU” (AI accelerator) chip makes use of a 3D package deal. The highest die is a 700mm2 accelerator on prime of an 800m2 SRAM. Sure I am stunned as:
A/ I do not know $AVGO had 3D chops. (It… pic.twitter.com/SjoVY90yiq– Patrick Moorhead (@PatrickMoorhead) March 21, 2024
Moorhead speculates that Broadcom’s XPU packs “heaps and much” of computing energy, very high-speed intra-chip connectivity, and the “highest efficiency” for exterior networking communications. Nvidia’s lately launched Blackwell AI accelerator additionally employs two compute models embedded in the identical chip, however Broadcom’s XPU is as giant and even bigger than that.
Up to now, no info has been offered in regards to the third-party “AI firm” for which Broadcom designed its supposedly highly effective XPU. The company is usually identified for its networking gear and providers, however it additionally provides customized chips for outstanding shopper organizations like Google. Like many different fabless chip designers, Broadcom will use the foundry providers of Taiwanese large TSMC to fabricate the XPU.
Customized AI accelerators are a part of Broadcom’s technique to promote knowledge heart options which can be performant, environment friendly, and scalable. Shopper AI purposes are driving the necessity for modern customized AI silicon, the corporate said, whereas a correct “AI infrastructure” would require a complete portfolio of network-related merchandise that may transfer all that knowledge as quick as potential. And Broadcom is, in fact, greater than keen to promote its AI portfolio to AI-focused tech firms sooner or later.
