Computex 2023 - Accelerated Computing and Generative AI (NVIDIA)

NVIDIA's Jensen Huang presented NVIDIA's software stack including AI enhanced gaming. their Accelerated Computing and Generative AI stack. Last but certainly not least NVIDIA presented their hardware stack for accelerated computing and generative AI.

And I bet you thought the RTX-4090i was expensive....ha!. How about:

Supermicro 8125GS-TNHR / AS-8125GS-TNHR has High-Performance Computing and AI/Deep Learning Training. It has 24 DIMM Slots; Up to 6TB DRAM; 4800 ECC DDR5 LRDIMM; RDIMM;8 PCI-E Gen 5.0 X16 LP, and 2 PCI-E Gen 5.0 X16 FHFL Slots. Its features include Flexible networking options, 1 M.2 NVMe for boot drive only, 12x 2.5″ Hot-swap NVMe drive bays, 2x 2.5″ Hot-swap SATA drive bays, 10 heavy-duty fans with optimal fan speed control and also has 8x 3000W redundant Titanium level power supplies.

For a discounted base price of $296,800 USD

NVIDIA software stack is Magnum IO. Magnum IO is based on OpenUCX, which is the Unified Communication X the open source framework for data centric and highly performant applications. If you are wondering if Linux is in there somewhere, yes it is.

Full video you can find it here: https://www.youtube.com/watch?v=i-wpzS9ZsCs&list=PLZHnYvH1qtOZ45xW29emPlPwlFCj9v7B0

However, be aware its 2 hours long.

In My Thoughts chapter, I spent some time talking about the Quest for Exaflops for 2 reasons, one is the NVIDIA's presentation seemed to be saying the Exaflop barrier was only recently exceeded and two because Nvidia is targeting the enterprise company with their AI solution, which is a market which has been declining in their use of supercomputers.

One final thing, on power comparison I don't think I explained very well, IF your datacenter is already hitting its power cap, then to add the NVIDIA HGX H100 or DGX GH200 you may need to replace some of the legacy general computers. That means your choice here might be: attempt to move classical computing workloads to the AI Supercomputer OR move the classical workloads into the cloud. The Grace Hopper Superchip is a "general" purpose computer, but I feel it might be under utilized running classical workload which rarely require the use of the accelerators or GPU.

Oh, I almost forgot: "Will it run Crysis?"

Once again I generated the thumbnail with AI, some of the variations looked like old SuperComputer designs, and I could not convince the AI to put in the network cable channels overhead, it insisted on putting them in the middle of the machines, which I thought was kind of odd, but funny. The last oddly hilarious thing is AI thinks assembly of a SuperComputer is a LAN party....

The martian lander video at the end is from NASA using AI to figure out the heating when entering the martian atmosphere

Chapters 00:00 - Intro 00:44 - Transitions 00:53 - Accelerated Competing & Generative AI 01:12 - Software Stack 01:21 - Magnum IO 02:26 - Generative AI in Gaming - ACE 06:48 - Hardware 07:05 - HGX H100 09:30 - Grace Hopper 09:42 - GH at Scale 10:02 - DGX GH200 10:08 - Grace Hopper 13:33 - DGX GH200 14:28 - Extend AI to the World's Heavy Industries 15:01 - Generative AI 16:04 - My Thoughts

Support me on Patreon: https://www.patreon.com/DJWare Follow me: Twitter @djware55 Facebook:https://www.facebook.com/don.ware.7758

Gitlab: https://gitlab.com/djware27

#nvidia #AI #Hopper ... https://www.youtube.com/watch?v=6Z2bImqOcPY