MLperf Wars & AMD’s Gawdy Numbers

  • by

Summer inventory clearance days for Radio Free HPC! We’re blasting through our processing backlog and getting the episodes out as fast as possible (but no faster). In this episode, we talk about the spate of MLperf benchmarks and how AMD hit it out of the park on their most recent earnings.

We start the episode with introductions, paying particular attention to Henry’s Las Cruces rammed earth survivalist bunker. A quick segue to how Google and Nvidia are battling for MLperf dominance in the latest round of benchmarks (here and here).  Both companies have something good to talk about, which, as Shahin states “is the sign of a well-designed benchmark.”

Nvidia’s brand new A100 processor is showing great speed ups in the range of 1.5x to 2.5x vs. their former flagship V100 GPU. Meanwhile, Google’s new TPU v4 dominates their older v3 chips by 2.2x to 3.7x. In the call, the guys discuss the folks who didn’t submit MLperf results and the general question of whether vendors should submit public benchmark numbers. Shahin and Henry defend vendors who don’t submit benchmarks while Dan’s position can be summarized by “I needs me a benchmark.”

In perhaps the best segue of Dan’s podcasting life, he moves from record-breaking MLperf results to AMD’s record-breaking financial results. Revenue up 26% year to year, gross margin up 3%, earnings per share by 4x. What’s not to like? Great job AMD! This leads to a tangent about the difficulty of getting to smaller processes, TSMC failure to fab Intel’s 7nm parts and, somehow, to Henry making YouTube videos about melting stuff in his backyard wood-fired pizza oven. “Can Henry Melt It” would be the title and each episode would feature a variety of viewer suggested things for Henry to melt. I know that I’d watch it voraciously.

We also discuss the not yet revealed prize package for out latest contest and how it’s now much better because HPE/Cray sent us a few Cray keepsakes that will make their way into this bundle. Stay tuned for more details.

Henry Newman’s Reasons Why No One Should Ever be Online. Ever:  Big bug in Cisco’s data center manager that allows anyone to spoof an administrator role by using a rest protocol (bypassing the web gui).

Catch of the Week: 

Shahin: MIT makes a bigger atom to produce the largest quantum chip of its type. MIT blows up an atom to the size of a small melon in order to make a better quantum device. Ok, not really, but they did make a much bigger atom.

Henry: Netgear decides NOT to patch more than 40 models of home routers. While you can’t expect vendors to patch everything forever, some of these products were sold as late as 2017, which isn’t all that long ago. This has already cost the company one sale, as Dan, a loyal Netgear customer for more than a decade, said that his next purchase, a 16 port GB router, will not be a Netgear due to this issue. Henry also shared his rules for upgrading network gear:  always upgrade after 3-4 years. Network gear always seems to fail when you need it most so upgrading according to age is a good policy. Plus, even if you’re upgrading to the same product specifications (i.e. a new 1Gb switch to replace an old 1Gb switch), you’ll almost always see a performance increase.

Dan:  Intel’s Chief Government Affairs Office, Jeff Rittener, wrote an op-ed article discussing how it is vital for the US government at all levels needs to support US semiconductor companies in order to keep technology and jobs onshore. He made a very good case and it’s well worth a read.


Download the MP3

* Sign up for the insideHPC Newsletter
* Follow us on Twitter
Subscribe on Spotify 
Subscribe on Google Play 
Subscribe on iTunes 
RSS Feed
* eMail us

Leave a Reply

Your email address will not be published.