It’s excessive reward when somebody like Jim Keller says that your organization ‘has made spectacular progress, and has essentially the most promising structure on the market’. That reward means twice as a lot if Keller really joins the corporate. At the moment Tenstorrent is saying that Jim Keller, compute architect extraordinaire, has joined the corporate as its Chief Know-how Officer, President, and joins the corporate board.
To our common viewers, Jim Keller is a identified knowledgeable in all issues pc structure. His historical past begins at DEC, designing Alpha processors, earlier than shifting to a primary stint at AMD for 2 years working to launch K7 and K8. Keller spent 4 years as Chief Architect at SiByte/Broadcom designing MIPS for community interfaces, 4 years at P.A. Semi, 4 years at Apple (A4+A5), then again to AMD for 3 years two years as Company VP and Chief Cores Architect in control of the brand new era of CPU architectures, K12 and Zen. This was then adopted with two years at Tesla as VP of Autopilot {Hardware} Engineering creating the Full Self-Driving chip, then two years as Intel’s Senior VP of the Silicon Engineering Group, earlier than leaving in June 2020. Since his departure from Intel, plenty of key business analysts (and ourselves) have been guessing the place Jim would land. He briefly appeared in the audience of Elon Musk’s Neuralink presentation in August 2020, alongside Lex Fridman.
Jim Keller: Work Expertise | ||||
AnandTech | Firm | Title | Vital Product |
|
Eighties | 1998 | DEC | Architect | Alpha |
1998 | 1999 | AMD | Lead Architect | K7, K8 |
1999 | 2000 | SiByte | Chief Architect | MIPS Networking |
2000 | 2004 | Broadcom | Chief Architect | MIPS Networking |
2004 | 2008 | P.A. Semi | VP Engineering | Low Energy Cell |
2008 | 2012 | Apple | VP Engineering | A4 / A5 Cell |
8/2012 | 9/2015 | AMD | Corp VP and Chief Cores Architect |
Skybridge / K12 (+ Zen) |
1/2016 | 4/2018 | Tesla | VP Autopilot {Hardware} Engineering |
Totally Self-Driving (FSD) Chip |
4/2018 | 6/2020 | Intel | Senior VP Silicon Engineering |
? |
2021 | Tenstorrent | President and CTO | TBD |
At the moment Tenstorrent reached out to tell us that Jim Keller has taken the place of President and Chief Know-how Officer of the corporate, in addition to being a member of its Board of Administrators. Jim’s position, primarily based on his earlier experience, would look like within the design of future merchandise for the corporate in addition to constructing on the workforce at Tenstorrent to reach that aim.
CEO Ljubisa Bajic confirmed Jim’s appointment as President and CTO of the corporate, stating that:
Tenstorrent was based on the idea that the continuing shift in the direction of ML-centric software program necessitates a corresponding transformation in computational capabilities. There may be no one extra able to executing this imaginative and prescient than Jim Keller, a pacesetter who’s equally nice at designing computer systems, cultures, and organizations. I’m thrilled to be working with Jim and past excited in regards to the potentialities our partnership unlocks.
Tenstorrent is a pure-play fab-less AI chip design and software program firm, which implies that they create and design silicon for machine studying, then use a foundry to make the {hardware}, then work with companions to create options (as in, chips + system + software program + optimizations for that buyer). For people who know this house, this makes the corporate sound like every of the opposite 50 corporations out available in the market that appear to be doing the identical factor. The standard break up with pure-play fabless AI chip design corporations is whether or not they’re targeted on coaching or inference: Tenstorrent does each, and is already within the technique of finalizing its third era processor.
Based in 2016, Tenstorrent has round 70 staff between Toronto and Austin. The crucial members of the corporate all have backgrounds in silicon design: the CEO led energy and efficiency structure at AMD in addition to system structure for Tegra at NVIDIA, the pinnacle of system software program spent 16 years throughout AMD and Altera, and there’s experience from neural community accelerator design from Intel, GPU techniques engineering at AMD, Arm CPU verification leads, IO virtualization experience at AMD, Intel’s former neural community compiler workforce lead, in addition to AMD’s former safety and community improvement lead. It feels like Jim will match proper in, in addition to have just a few former colleagues working alongside him.
Tenstorrent’s present era product is Grayskull, a ~620mm2 processor constructed on GF’s 12nm that was initially designed as an inference accelerator and host. It incorporates 120 customized cores in a 2D bidirectional mesh, and provides 368 TeraOPs of 8-bit compute for under 65 W. Every of the 120 customized cores has a packet administration engine for information management, a packet compute engine that incorporates Tenstorrent’s customized TENSIX cores, and 5 RISC-V cores for non-standard operations, similar to conditionals. The chip focuses on sparse tensor operations by optimizing matrix operations into compressed packets, enabling pipeline parallelization of the compute steps each by the graph compiler and the packet supervisor. This additionally permits dynamic graph execution, and in comparison with another AI chip fashions, permits each compute and information switch asynchronously, slightly than particular compute/switch time domains.
Grayskull is at present transport to Tenstorrent’s clients, all of that are nonetheless undisclosed.
The following era chip, generally known as Wormhole, is extra targeted on coaching than acceleration, and in addition bundles in a 16x100G Ethernet port swap. The transfer from coaching to acceleration necessitates a quicker reminiscence interface, and so there are six channels of GDDR6, slightly than 8 channels of LPDDR4. This may appear low in comparison with different AI chips discussing HBM integration, nonetheless Tenstorrent’s plan right here appears to be extra aligned for extra mid-range price construction, but additionally providing machine studying compute at a greater fee of effectivity than these chips pushing the bleeding fringe of frequency and course of node (a part of this will likely be in yields as effectively).
So the place precisely does Keller slot in if the present era is already promoting, and the following era is nearly able to go? In talking to the CEO, I confirmed that Keller ‘will likely be constructing new and attention-grabbing stuff with us’. This appears to counsel that the imaginative and prescient with Keller’s involvement goes to be on 2022/2023 {hardware} in thoughts, following Tenstorrent’s overriding Software program 2.0 technique that the {hardware}, compiler, and run-time supply a full-stack strategy to sparse (and dense) AI matrix calculations. In Jim’s personal phrases:
Software program 2.0 is the most important alternative for computing innovation in a very long time. Victory requires a complete re-thinking of compute and low stage software program. Tenstorrent has made spectacular progress, and with essentially the most promising structure on the market, we’re poised to change into a subsequent gen computing big.
Jim Keller formally began final Wednesday, and the official wire announcement is about for 1/6, however we have been allowed to share prematurely. Our request for an interview with Jim has been famous and filed, doubtlessly for just a few months down the road as the corporate has some extra particulars on its platform and roadmap (I’ve additionally requested for an up-to-date headshot of Jim!). For these , I interviewed Jim again in July 2018, simply after he began at Intel – you possibly can learn that interview right here.