Hot Chips 2021 Keynote Live Blog: Department of Energy on AI Chipsby Dr. Ian Cutress on August 24, 2021 1:00 PM EST
12:55PM EDT - Welcome to Hot Chips! This is the annual conference all about the latest, greatest, and upcoming big silicon that gets us all excited. Stay tuned during Monday and Tuesday for our regular AnandTech Live Blogs.
12:55PM EDT - Starts at five past the hour
01:03PM EDT - Third keynote from the US government
01:05PM EDT - 'I don't claim to speak for the US government'
01:05PM EDT - 'These are my views'
01:06PM EDT - This is what I worry about
01:06PM EDT - Science and technology evolve differently but are intertwined
01:07PM EDT - Technology leads science and science leads technology
01:09PM EDT - Department of Energy is really the Energy of Everything
01:09PM EDT - Have to report to the president every year in March
01:11PM EDT - he's talking a lot about DoE holistically
01:11PM EDT - DoE 'solves problems no-one else can'
01:12PM EDT - 17 national labs, workforce of 100k
01:13PM EDT - it's not about being faster, it's about being smarter
01:13PM EDT - provide actionable decisions
01:14PM EDT - 'Can't talk about everything we do'
01:14PM EDT - Not necessarily specialists but know how to tackle complex problems.
01:15PM EDT - questions asked are often ill-defined
01:16PM EDT - moving towards exascale
01:18PM EDT - DoE was hammered by congress due to Fugaku
01:18PM EDT - Did a top-to-bottom review
01:18PM EDT - DoE does things that are viewed to be nuts but they prove fruitful
01:19PM EDT - 3 years ahead of Darpa programme to get to petascale with PS3
01:19PM EDT - Defining things for a purpose - to solve problems, and these are the instruments
01:22PM EDT - Learning how to solve complex problems
01:23PM EDT - Learning how to model
01:24PM EDT - quoting von Neumann
01:26PM EDT - Our largest computer simulations stress our largest systems - architectures have historically been developed to solve equations and not to deal with data at tomorrow's scale
01:29PM EDT - Working on Fukushima, had to answer in an hour about whether to evac US citizens from Tokyo
01:30PM EDT - Non trivial problem if you decide to do that
01:32PM EDT - no metrics for how close you might be with some of these decisions at the high level
01:34PM EDT - This is what the DoE does
01:35PM EDT - More about 'just simulate faster won't help'
01:37PM EDT - DoE is a big customer of SambaNova
01:38PM EDT - There's Cerebras, Groq, Graphcore
01:38PM EDT - Think beyond just the immediate market on chips
01:39PM EDT - Need to do much more of this
01:42PM EDT - Every slide is getting a long descriptive treatment, but there's little to bite
01:44PM EDT - Creating petabytes of data with petascale systems
01:45PM EDT - No longer can one person be the top-to-bottom solution for epiphanies
01:47PM EDT - health worries about rare outcomes
01:50PM EDT - Still waiting for the chip talk
01:51PM EDT - waiting to see if reconciliation passes in the house for money for the CHIPS Act
01:55PM EDT - Q&A Time
01:57PM EDT - Q: What do future chips look like? A: PIM is good, need to revisit the architectures. Old systems were locked into the world we've grown into. This used to be the only ways to do things. The 40s was the last time we did clean sheet top-to-bottom how we do things changed
02:02PM EDT - Q: Should the DoE develop its own chips, or buy commercial silicon? A: Have done it in the past. But have to have an ecosystem. Over time, perspective has been tweaking company roadmaps, doing the RNE, aligning it with DoE requirements. Has to be sustainable. Having a design others build is an idea, but working with the private sector is important. DoE understands roadmaps and gets involved, are there points of intersection with DoE and these private companies. Not always a point, but conversations start early
02:04PM EDT - Q: Startup investment A: Investing in startups unlike VCs - our investments involve thigns that have worked involve data modalities - well organized scientific data and sparse data. They invest with data. Have a place to exercise technologies that work for problems they face. Not organized enough to have a big enough aperture into the private market to do this - we need more test beds to give real data to people offering to solve real problems
02:05PM EDT - ROI for DoE is different to VC
02:06PM EDT - That's a wrap