MENU

CEO interview: Walter Goodwin, Fractile

CEO interview: Walter Goodwin, Fractile

Business news |
By Nick Flaherty

AI


UK chip designer Fractile is on the way to producing AI chips for datacentres. The company is developing an in-memory compute (IMC) architecture for AI inference in the datacentre with RISC-V technology and SRAM memory and is taping out test chips from its teams in London and Bristol.

The company is also developing reference designs all the way up to rack systems in cloud datacentres.

“Internally things are moving forward, we have multiple test chips planned for this year and a tapeout next month,” CEO Walter Goodwin tells eeNews Europe. “There’s also been a lot of external movement that corroborates our strategy.”

The lessons from Chinese AI technology developer DeepSeek has also helped, he says.

“Deepseek is a real clear signal for the impact of a mixture of expert models where small fraction of the weights are active at any one time and this has implications for the optimum way to put hardware together for efficiency.”

“We also see the impact and value of having a team that thinks about the full stack for efficient training and relatively efficient inferences as well. Part of the story of DeepSeek is a reminder that with extremely proficient models you can do more with less,” he said.

Pat Gelsinger, former CEO of Intel, came on board as an angel investor in January in a significant boost to the company as it looks to raise cash for producing datacentre chips. It has already raised £15m with backing from NATO.

Former Intel CEO Pat Gelsinger invests in UK AI startup Fractile

“He’s now able to make a huge number of amazing introductions so we are benefitting a lot from that,” said Goodwin. “I was introduced by a mutual friend, a US semiconductor VC turned angel investor just after he left Intel just before Christmas.”

The focus is on chips for the datacentre with a focus on low power consumption and built on advanced but not leading edge process technology.

Process technology

“We are using relatively advanced process technology but not bleeding edge FInFet. One of the factors for the test chip and products we are building is that a good portion of the chip is SRAM cells and the cost per sq mm of silicon has gone up so much and SRAM is not scaling aggressively,” he said.

“Fractile will be bringing in the money needed for a very competitive product but at the same time one of the things that is very challenging for companies that are directly pursuing Nvidia with GPU clones with HBM is they will be six to 12 months behind in the queue for the process technology.”

“One of the things we are really interested in is that IMC allows for low power density dies with a low W/mm2 but you need dense memory. The reason we don’t cram more GPUs into a unit is the power consumption.”

He points to Nvidia which has two versions of its NVL72, a 120kW single rack with custom cooling and a dual rack version for data centres that cannot meet that power envelope. Nvidia is donating the rack design with its Blackwell GPUs to the Open Compute Project.

“One of the things for the future is that we will be doing a huge amount of inference, and the minimum systems to serve these models are very expensive at a certain performance point so it makes more sense to amortise the costs across cloud based systems so that’s the reason we see that as the appealing market, with low latency and incredibly high throughput,” he said.

“So you are crunching through millions of tokens per second, and today there’s not the demand for that at the edge. The architectures have to change enormously to make that work, which makes it difficult to move form the edge to the datacentre”

“At the moment the chip and the software is the core product and we are building reference designs up to the rack but the cloud service providers (CSPs) we want to partner will have their own architectures so we are trying not to have a lock in up to the rack level.”

 www.fractile.ai

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News

Share:

Linked Articles
10s