If Jim Keller makes a move, it’s best to watch closely. Even though the chip design legend’s tenure at Intel was cut short in 2020 due to family matters, Keller still made his mark at the semiconductor giant in a smaller but similar way he did for AMD, Apple and Tesla, helping design the predecessor to the efficiency core that will inhabit Intel’s high-density “Sierra Forest” Xeon chips in 2024.
Keller is now nearly a year-and-a-half into his tenure as CTO and president at AI chip startup Tenstorrent, which is led by fellow AMD veteran Ljubisa Bajic, and the chip designer is starting to get more serious about convincing hyperscalers and other datacenter operators to adopt its line of RISC-V-based AI processors it says have been optimized for neural network training and inference.
This was signaled with the Monday announcement that Tenstorrent has hired yet another fellow AMD veteran, David Bennett, as its chief customer officer so that the startup can expand from being a primarily R&D-focused organization to one that can make money from such efforts.
A money-making endeavor like this would benefit from having someone at the sales helm who is well-versed in the components world and the systems world as well as how the two interact, which is why Bennett is very likely a strong fit for such a job. On the systems side, Bennett gained this experience more recently with his roles as president of Lenovo Japan and CEO of NEC Personal Computers, which is a joint venture between Hong Kong’s Lenovo and Japan’s NEC Corporation.
With his new job at Tenstorrent, Bennett will now have the chance to combine that experience with what he learned in his 10-year tenure at AMD, where he oversaw the chip designer’s relationships with major OEMs in the Asia Pacific and Japan before leaving for Lenovo in 2018.
It was in that final role at AMD, as general manager of OEM accounts, that had Bennett involved with system makers from the early design stages of new platforms to the sale of such platforms. We wouldn’t be surprised if Bennett has a similar spectrum of responsibilities at Tenstorrent.
“What excites me most about Tenstorrent is that its technology scales from a single chip to a thousand chips, and from low power all the way up to megawatt data centers. $1,000 cards to $1 million high-density racks are powered by a single software stack that supports inference and training and a wide range of models, which is a game changer,” he says.
If Bennett’s hire seems a little too soon for Tenstorrent, there’s a good argument to be made that it isn’t. The startup had a test chip, Jawbridge, in 2019, which was followed with its first production chip, Grayskull in 2020, and it recently laid out a roadmap with three new chips in the following years: Wormhole, Black Hole and Grendel. As reported by Dylan Patel of SemiAnalysis earlier this year, Tenstorrent expects to have 1,000 PCIe cards with dual Grayskull chips in datacenters by the end of this year in addition to a machine with 1,000 Wormhole chips in another.
These are very small deployments compared to the massive AI compute footprint held by Nvidia, but with Bennett now in place to grease the wheels of system discussions, there must be hope that soon we’ll see more datacenter operators buying into Tenstorrent’s unique approach of “conditional computation,” which we discussed in detail with Bajic in 2020.