2024_07_15.md

Oxide and Friends: July 15th, 2024

Heterogeneous Computing with Raja Koduri

We've been hosting a live show weekly on Mondays at 5p for about an hour, and recording them all; here is the recording.

In addition to Bryan Cantrill and Adam Leventhal, our special guest was Raja Koduri.

Some of the topics we hit on, in the order that we hit them:

• [0:12] Audience complaints about audio quality. Call backs to old Twitter space episodes
• [2:09] Audience requests for new episode topics invited
• [3:12] Forbes article about Vector Databases
• [4:49] Raja's tweet over the weekend
• [5:30] Raja's start of career
• [6:00] S3 Graphics
• [7:12] Tandem Computers
• [10:34] 3dfx Oral History Panel with Ross Smith, Scott Sellers, Gary Tarolli, and Gordon Campbell
  • 3dfx
• [15:50] OpenGL
• [16:23] Glide
• [17:20] Direct3D
• [19:12] "... GPU people think in terms of API, CPU people think in terms of ISA ..."
• [21:18] Exodus Communications

  • "The company created and enabled many "firsts" for the "modern Internet", including the first "storage as a service" cloud service branded as DataVault in partnership with Cisco systems and Sun Microsystems."

• [22:45] VIA
• [33:05] "In the history of parallel computation, there was no architecture that was successful whose programming model was different than its execution model"
• [34:19] CUDA
• [39:31] Transportation Analogy to understand CPU, GPU, FPGA and ASICs
• [53:59] Dennard scaling
• [54:40] VLIW
• [56:56] "...You become what you are measured by..."
• [1:01:28] John H. Crawford
• [1:02:12] SPEC benchmark
• [1:03:10] John L. Hennessy
• [1:09:37] GPGPU
• [1:09:50] Brook GPGPU programming language
  • https://en.wikipedia.org/wiki/BrookGPU
  • https://graphics.stanford.edu/projects/brookgpu/lang.html
• [1:15:47] "...If a customer wants a burger, give it to him"
• AMD APU
• William (Bill) Dally's talk
  • PPT Energy Efficiency and AI Hardware
  • Video
• [1:20:00] Windows being a GUI first OS will cause issues for GPGPU usecases, if GPU is held up for long time
  • Windows TDR for GPU timeout
  • https://learn.microsoft.com/en-us/windows-hardware/drivers/display/timeout-detection-and-recovery
  • https://www.intel.com/content/www/us/en/docs/oneapi/installation-guide-windows/2024-1/gpu-adjust-timeout-detection-and-recovery-setting.html
• [1:24:30] Raja's Performance Equations on Twitter
• [1:25:50] "The new UMA is NUMA... NUMA is here to stay"
• [1:26:20] Chip Packaging
• [1:31:49] AMD Fiji GPU - First GPU to use HBM
  • https://en.m.wikipedia.org/wiki/High_Bandwidth_Memory#Development
  • https://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/3
• [1:36:10] Cost of setting up a 1MW Data Center with NVIDIA hardware costs $35 million.
• [1:36:34] GW Solar Farms in India co-located with data centers. Transmitting power is more expensive than transmitting bits.
• [1:38:11] Raja sets a goal: Compute CapEx should be 1 dollar per Watt. Bryan tries to wrap his head around Oxide rack's pricecuts.
• [1:40:06] Rail Road Tycoons in US, being compared to huge data centres
• [1:46:15] Advice for people in their early career about boom and bust cycles
• [1:47:02] "Performance in a given constraint is a common theme"
• [1:50:38] Marc Andreessen - Why Software Is Eating the World
• [1:51:29] Hardware-Software Codesign is still rich. The next dirsuption might come from "the marriage of Python and memory"
• PRs needed!

If we got something wrong or missed something, please file a PR! Our next show will likely be on Monday at 5p Pacific Time on our Discord server; stay tuned to our Mastodon feeds for details, or subscribe to this calendar. We'd love to have you join us, as we always love to hear from new speakers!