
Tenstorrent’s COO Keith Witek on strategic breadth
Tenstorrent’s chief operating officr Keith Witek sat down with eeNews Europe to discuss how Hyundai Motor Group, Samsung Catalyst Fund and other investors will help the company take its AI architecture to a broader set of applications.
Tenstorrent (Toronto, Canada) has raised US$100 million, principally from two strategic investors from South Korea, Hyundai Motor Group and Samsung.
At one level the development is straightforward. Like many of the large group of AI startups formed in recent years and more established chip companies, Tenstorrent started by making sure it had an offering for the data center. Now Hyundai is going to be deploying Tenstorrent technology across vehicles in the Hyundai, Kia and Genesis brands.
The backing won from the Samsung Catalyst Fund does not guarantee any involvement from Samsung Electronics but it makes sense that if the venture capital wing of the firm finds investment worthwhile then Tenstorrent is in a good position to catch the attention of the parent.
The US$100 million announced in latest round could be followed up additional closings and comes after the company has raised about US$234 million up to now, Witek said. “The plan was to raise about US$60 million but then strategic investors started expressing interest, said Witek.
Tenstorrent founded in 2016 is more than just an AI accelerator company. It has also moved into designing high-performance RISC-V microprocessor cores as well as licensing in RISC-V cores, depending on requirements. It also has a strong emphasis on software, conditional execution and compilation. It is a ‘full-stack’ hardware-software company that is broad in terms of its technology offerings: from scalable neural network support to scalable high-performance RISC-V processing.
Business model
The company is also exceptionally broad in terms of its business model. This includes licensing intellectual property, selling chiplets, chiplet-based components, accelerator cards and even fully configured servers.
It may all seem to run against the conventional wisdom for startups of focus on one thing in one application sector to achieve differentiation and penetration. The wisdom of a previous generation of startups was never to compete with you customers. But the market is much larger, more compex and dynamic with multiple types of customer who want different forms of support. And Tenstorrent has always had a different view of AI processing from the point when Jim Keller, then with Tesla but now CEO of Tenstorrent, provide seed investment.
The broadening of Tenstorrent’s field of engagement with Samsung’s and Hyundai’s involvement is something that the company had already started. In May the company announced a collaboration with LG Electronics Inc. to build a new generation of RISC-V, AI and video codec chiplets to potentially power LG’s premium TV and automotive products of the future as well as Tenstorrent’s data center products.
And now Hyundai is taking a long-term position to deploy Tenstorrent technology across its automobiles. “If you look at the market, Samsung is going hard at the datacenter. And how many mobile phones do they make? 260 million, 280 million a year? They are looking for an alternative to ARM. With Google porting Android to RISC-V the future is clear,” said Witek.
Heterogeneous computing
It is no coincidence that Jim Keller, Tenstorrent CEO, spent a number of years at AMD helping that company pioneer heterogeneous computing. At the time of CISC versus RISC processor architecture had evolved into x86 versus ARM. The idea of multiprocessor heterogeneous SoCs was starting to come in, but was yet to be fully supported by standards and infrastructure.
Witek observed that a decade on chiplet infrastructure is starting to percolate down to a broader set application sectors.
“Chiplets is something we feel strongly about. The FPGAs and big processors ran out of lithography field of exposure,” he said. He observes that multi-die components were partially pioneered by Intel and AMD on high-performance x86 processors. The high-value of the component could justify what was then customized packaging around processors and cache memory. This evolved into something more recognizable as chiplet-style manufacturing with big processors such as AMD’s Fiji GPU which was assembled in Korea by SK Hynix, the provider of the high bandwidth memory (HBM).
“As Moore’s Law becomes exponentially more costly it doesn’t make sense to put every bit of logic on a die and pay $30,000 to $40,000 per wafer at the 2nm node,” said Witek. Not every bit of logic or functionality is on the critical path. “It makes sense to use trailing nodes where you can.”
Chiplets
The use of chiplets – if your architecture is optimized to make use of them and is scalable with them provides for the flexibility to go across a broad set of applications, said Witek.
In short Tenstorrent sees AI as more than just an addition to conventional processing. It is effectively the future of computing that will be implemented on a broad set of scalable software-programmable resources. These would include CPUs, GPUs DSPs, neural networks. It is almost the very definition of heterogeneous computing.
Witek said that being able to license out, to collaborate – or to provide chiplets, boards and boxes with a scalable architecture – means Tenstorrent can go from the datacenter down to milliwatts. “And the software will keep getting better and better.”
Witek concluded that with the latest investment Tenstorrent is “scaling up to take technology in different additional directions.”
Related links and articles:
News articles:
Garage chipmaker forms startup with processor veteran
Intel hires former Apple chip architect to lead silicon engineering
Conditional computation key to Tenstorrent AI processor
