Eric Seppanen

Checkout our latest report on how just eight SN40L chips run 1 Trillion parameter model at blazing fast speed https://lnkd.in/grwdFVAi. It was great to collaborate with some of the most talented folks in this area. Here is the link to the demo operating on similar principles https://fast.snova.ai/

28

Listening to some talks earlier today about the future of big foundation models reminded me of some CPU history: Pentium 4. Pentium 4 was Intel's final generation of single core processors and they optimized for a very specific performance element: high frequency. They shipped at up to 3.8GHz and overclockers would easily push them above 4GHz, increasing voltage and running high performance cooling setups to do it. They ran hot, were super power hungry, sounded badass (great for marketing against AMD's Athlon processors), and were the end of the frequency-chasing days before parallelism became the new thing to chase after. By comparison—16 years later—a modern Apple M3 processor also runs at around 3.5Ghz but achieves 128x (!!) the performance on some benchmarks. Hearing the panelists discussing the amount of hardware, electricity, and money being spent right now chasing higher parameter counts sounds a bit like Pentium 4. Can we chase the next level of sophistication with a 100T parameter model that costs $100M to train? Maybe we can. But maybe we're approaching the point where newer architectures shift the power-to-price boundary curve enough that model size isn't worth chasing anymore. I'm really not an expert here so maybe someone who knows more than me can comment—just wanted to share this connection my mind made to some 2000's era tech.

27

2 Comments

Do you know what it takes to design a differential pair that works past 30 GHz? If you are planning to take on PCIe 4, 5, or 6, USB 3.2, or any modern SERDES interfaces, I hope so! When we were building the JITX autorouter we wanted to be sure to address the issues plaguing modern designers. We found guidance in this DesignCon paper from Donald Telian, Kevin Rowett, and Ilya Teplitsky: https://hubs.ly/Q02wf7l70 Summarizing the key points for diff pairs here: 1. Above 8 GHz: have fear. Above 30GHz, everything is a stub. 2. Route as short as possible to minimize loss 3. Match P/N statically to within ~0.15ps 4. Match P/N dynamically to never accumulate 0.5ps of skew within a 100ps window 5. Size skew matching bumps below the relevant feature size so they become 'invisible' 6. Space out skew matching bumps irregularly to avoid impedance discontinuities. (the paper has a lot of other great lessons, so check it out) We make sure that JITX helps people design better boards, so when we asked ourselves "How should our autorouter draw differential pairs?" our answer is of course "Just do the Best Possible Thing." Put the computer to work making the detailed shapes so humans can focus on the more important stuff. Here's a video of the router matching skew to 1 femtosecond, enforcing dynamic phase matching, minimizing distance, minimizing impedance discontinuities, and avoiding obstacles. Feels good to offload all that to the computer! You can learn more on our website: https://hubs.ly/Q02wfdCL0 What do you think about this paper Robert Feranec, Benjamin Dannan, John Burkhert? Agree or disagree with the approach? #electricalengineering #signalintegrity #DesignCon

206

30 Comments

"Something that I was told early on in open source, that I sometimes forget," said Porter Lead Maintainer Sarah🦎 Christoff in her interview with Open at Intel's Katherine Druckman, "is that you just need to show up. So as much as I was striving at coding, worrying about my project's health around features, and responding to the community, I was doing that pretty much solo. And because of that, that created that isolation factor. No one knew about it because I was spending a hundred percent of my energy on the project and not building the community. Once I started going to TAGs, to SIGs…when you go to Kubernetes Days, and even your local Gopher meetup, or whatever your local meetup is…once I started getting out there, I started making those connections. That's when I could ask, can you help me? What do you think about this?" In this episode of the Open at Intel podcast, Defense Unicorns' Sarah Christoff, who is also lead maintainer of the open source project Porter, talks about the challenges of being an open source maintainer, the importance of community, Porter, Zarf, rescuing animals, and how as an open source community we are more powerful together. #opensource #Porter #Zarf #maintainers #CNCF #Kubernetes Read the full transcript: https://intel.ly/4dnBeTx

4

1 Comment