@goodwin_ml joined @HarryStebbings on @twentyminutevc to talk about Fractile's technology, building a frontier AI company in the chip layer, and why he believes the AI buildout will only continue to scale.
https://t.co/Q8HQzIofmj
https://t.co/Jj1sec03IJ
"You have this huge increase in bandwidth, which translates to a huge increase in speed."
@goodwin_ml joined @tbpn to talk about how our chips achieve far higher memory bandwidth with stacked DRAM, and how that translates into a huge performance advantage across workloads.
Walter Goodwin believes the world's most capable AI models are held back from delivering on their potential by the chips of today. Yesterday his company, Fractile, announced a $720M Series C to fix that.
Walter, an Oxford PhD, founded Fractile in summer 2022, on a bet that there would be a huge dividend to speed in running AI models. The round is led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, and comes six months after Fractile's $220M Series B.
We sat down in London to discuss the future of semiconductors alongside Sir Robin Saxby, the founding CEO of ARM, who started one of the largest companies ever to come out of the UK with 12 engineers in a pub.
Congratulations to @goodwin_ml and the @fractile_ai team. Enjoy this one!
TIMESTAMPS
0:00 Intro
1:24 Starting Fractile
6:01 Inference Is the Profit Center
9:39 Chips Over My Dead Body: The Founding of ARM
18:45 The Art of the Opinionated Chip
25:09 Breaking the Memory Wall
32:15 How ARM Picked Winners
38:23 Robin Calls a Correction
42:51 The Stickiness of the Transformer
46:49 Is There a Limit to AI Demand?
56:11 How TSMC Became the Standard
57:44 AI-Designed Chips
1:03:52 Brutal Honesty
BREAKING: FRACTILE RAISES A $720M SERIES C CO-LED BY LIGHTSPEED
I got an inside look at the technical breakthroughs @fractile_ai has been quietly building for years, with Founder and CEO @goodwin_ml, alongside @guruchahal and @jonahcader
🧠 Why memory, not compute, is AI’s real bottleneck - and why today’s chips force a choice between big but slow (HBM) or fast but small (SRAM)
🧱 Stacking DRAM directly on logic gives ~20x the bandwidth per chip of HBM and 10x+ the capacity of SRAM
🔁 Why they chose not to bake the transformer into silicon when models change faster than a chip’s 3-5 year life in the data center
🌍 How a team spanning the UK, US and Taiwan pulled off a “secret skunkworks” in advanced packaging
🧑💻 Hiring across silicon, systems and software in London, Bristol and Palo Alto
(My favorite part was hearing how the team flew to London, drove three hours to Bristol for one day with the Fractile team, and even got to catch an eclipse! 🌘)
FULL SEGMENT BELOW
@lightspeedvp
If you are interested in joining a chip-first frontier AI company and taking part in this mission, we want to hear from you. Apply here: https://t.co/vKqblEE2n3
We’re delighted to announce that Fractile has raised $720M in a Series C led by @thrivecapital@foundersfund@lightspeedvp and @redpoint, with further participation by brilliant new and existing investors.
Our first product-scale chip, which we will be manufacturing next year, will stack DRAM directly on top of the processor, combining radically better bandwidth and speed with the capacity needed to run the world’s largest models at long context.