Broadcom has demonstrated that it is probably the world’s largest processor. But for what application? When we visited TSMC’s events, we were always shown a deck of multi-chiplet processors using the company’s chip-on-wafer-on-substrate (CoWoS) packaging technology and reticle range (858mm^2, 26mm by 33) use around mm) Compute Chaplet. We can’t take pictures of the deck, but there are definitely processors that get attention. One of those devices comes from Broadcom, and has been on display at the company’s recent investor events.
To most observers, Broadcom is a networking and telecommunications giant, but the company also has a significant custom chip design business. For those unfamiliar with this unit of Broadcom, Google Contract is one of the company’s most prominent clients in terms of chip design.
However, like TSMC, Broadcom does not disclose its clients. For those looking to resurrect their short-lived innovations, Broadcom has a list of them in their recent press release. What it does to impress is to demonstrate its vast achievements to its investors. These are very broad indeed, as our friend and colleague Patrick Morehead of the market analysis company Moore Strategies observes.
“Here’s another fun one,” Patrick Moorhead wrote in an X-post. “The guy who’s Frank Ostojic is smiling. [who] Operated by Broadcom’s Custom Silicon Group. He must be smiling because he announced that he has a third XPU design from a major ‘consumer AI company’.
Here’s another fun one. The Guy Who’s Smiling (Frank Ostojic) runs Broadcom’s Custom Silicon Group. He must be smiling because he announced that he has a third XPU design from a major “consumer AI company”. On the right is a close-up of the XPU. You can see the 2 compute units at… pic.twitter.com/sseCi02B4KMarch 21, 2024
Broadcom officially brands these chips as XPUs so as not to disclose their applications. Meanwhile, high-bandwidth memory usage reflects its target usage, which could be artificial intelligence or cutting-edge AI-infused network switching.
“On the right is XPU’s closest,” Morehead added. “You can see the two compute units in the center and all the HBM to the left and right. A full-up custom SoC with lots of compute, HBM, very fast intra-chip connectivity and, as you’d expect, the most Performance External Networking.”
Producing a chiplet of this scale (ie close to reticle size) is already an achievement. Getting it to a decent level is another dimension of success, and it seems that Broadcom’s foundry partner, TSMC, has probably accomplished that as well. Now, it’s time for software to catch up and harness that processor power.