On today’s show we talk about MLPerf benchmarks and what we can infer from training results; we talk AI accelerator integration in HPC systems and software stacks; we take a look at a new cluster competition and talk “Moneyball Medicine” to close the program.
The latest iteration of the MLPerf benchmarks for machine learning training are out, and we get a look at how various accelerators, including Nvidia’s new A100 GPU and Google’s TPUv4 devices stack up. We talk to Paresh Kharya, senior director of product management and marketing for the datacenter GPU line at Nvidia, about the results and, more importantly, who the tests might be improved to help buyers actually decide which architectures to invest in for machine learning.
Today’s program takes a different look at AI hardware, this time in the context of HPC with two separate interviews, one focused more on systems integration of AI accelerators in supercomputing environments, the other more on the software and frameworks side with two researchers from Lawrence Livermore National Lab. Both are working on two ends of the AI accelerator puzzle, we look at why not just use GPUs for training and CPUs for inference, especially since they have those in abundance and also examine some of the lessons learned.
Also on today’s show, we talk to Dan Olds, principal at Gabriel Consulting, about a new Student Cluster Competition that he has started up to augment the similar competitions normally hosted at the ISC, SC, and ASC events each year. This one is a bit different but will have results that are compatible with the other competitions allowing for historical comparisons.
We close the show with something a bit different. The author of “Moneyball Medicine: Thriving in the New Data-Driven Healthcare Market” and strategic VC, Harry Glorikian, talks about where tech investments will be in healthcare and where some of the challenges are in bringing the industry up to date, technically speaking.
Thanks as always for tuning in. Timestamps below for your reference.
1:33 – Mlperf results analysis with Nvidia perspective
16:51 – Dr. Ian Karlin (Livermore National Lab) on AI hardware integration into HPC systems, workflows.
27:57 – Continuing from topic above, this time with software integration of AI accelerators in HPC with Dr. Brian Van Essen (LLNL)
41:36 – First look at a new HPC cluster competition for rising HPC systems engineers.
47:11 – “Moneyball Medicine” with VC, author Harry Glorikian.