HPC GuruOct 3, 2023

Other than HPE, who wants to build OLCF-6 and NERSC-10?

Can #HPC #Cloud providers save DOE from a single vendor procurement?

RFP technical requirements are out for two of the DOE’s next #exascale-class #supercomputers

https://nextplatform.com/2023/10/02/the-first-peeks-at-the-doe-post-exascale-supercomputers/…

#HPC #AI

Show threadHPC GuruOct 3, 2023

👏👏Notable quote by TPM:

Intel has had quite enough of being a prime contractor for #supercomputers, and CEO Pat Gelsinger has come to his senses & stopped talking about reaching #Zettascale by 2027.

As far as we can tell, that mike drop broke the mike.

And the Raj.

#HPC

Show threadHPC GuruOct 3, 2023

TPM's view on other on-prem #HPC vendors for these #Exascale #supercomputers:

o Nvidia is making too much $$ on #AI to care

o Dell: Michael Dell doesn’t like to lose money - TACC
seems to be enough for its #HPC

o No way that Atos or Fujitsu can sell into US government labs

Show threadJanne MorenOct 3, 2023
@HPC_Guru
Idly wondering what would happen if HPE/Cray decided that with total lifetime demand of 2-4 machines or so, the development cost for post exascale wasn't worth it at any price, and also decided to bow out.
Show threadJanne MorenOct 3, 2023
@HPC_Guru
It's not just that the machines nowadays are eye-wateringly expensive. It's not even that there are very few workloads that can take full advantage. I suspect that ncreasingly the tech development needed for such massive machines isn't all that relevant for the much smaller machines they can actually make money on.
Show threadHPC Guru

@jannem The same argument was made when we had the first petascale systems.

Re-cycling one of my past quotes here:

The most exciting thing about Exascale is affordable Petascale.

Oct 3, 2023 at 5:21amWeb
Show threadJanne MorenOct 3, 2023

@HPC_Guru
I hope you're right but I do wonder. TOFU is a good example: it is an excellent interconnect - above 15k nodes. Below that other tech is better.

But nodes aren't really getting cheaper or less power hungry. The market for 15k+ node machines will remain very small. The cost of TOFU development will never really trickle down to benefit "normal" size machines to a significant degree.