What's an AI supercomputer?
Our first topic is whether a supercomputer can or cannot be a “AI Supercomputer.” This is based on France (along with HPE) unveiling a new AI system which will double the capacity of French supercomputing. So what are the differences between a traditional super and a AI super. According to Dan, it mostly comes down to how many GPUs the system is configured with, while Shahin and Henry think it has something to do with the datasets. Send us a note or a tweet if you have an opinion on this.Software SMP hits 10k
The guys also discuss ScaleMP and how their announcement of record results, with close to 10,000 customers as of the close of 2018. This led to talk about SMP vs. MPP from a performance standpoint. Henry asserted that a clustered approach will always be superior to a big SMP approach, all things being equal. Dan doesn’t agree and Shahin confesses his love of ‘fat node’ clustering. Dan agrees with Shahin, but wonders why no one is doing it.We also note that Mellanox got a nice design win with the Finns, as they’ll be installing 200 Gb/s HDR InfiniBand interconnect in a new Finnish supercomputer to be deployed in 2019 and 2020. The interconnect will be used in a Dragonfly topology.
Catch of the Week
- Shahin’s catch of the week is a mathematical puzzle titled “The most unexpected answer to a counting puzzle.” Here’s a link to the video.
- Dan likes a good comeback story and in light of that, his catch of the week is AMD nabbing a design win at Nikhef.
- Henry HAS NO CATCH OF THE WEEK. This makes him the “RF-HPC Villain of the Week” :)
Sign up for our insideHPC Newsletter
No comments:
Post a Comment