US DoE Announces Summit with IBM Power and NVIDIA Tesla

[Patrick Kennedy]

Built for a number of high-performance computing and AI use cases, the US DoE ORNL Summit supercomputer uses IBM Power9, NVIDIA Telsa V100, and Mellanox EDR Infiniband to achieve over 200 PFLOPS of performance

The post US DoE Announces Summit with IBM Power and NVIDIA Tesla appeared first on ServeTheHome.

Continue reading...

 

[i386]

Spoiler: typo

and Mellanox FDR Infiniband to achieve over 200 PFLOPS of performance

EDR*

 

[Patrick]

Thanks @i386 - I actually made that oral typo this week as well. I must have FDR on my mind. 26K miles, 3 international round trips in 13 days :-/

Had it right int he article, messed it up in the snippet.

 

[gigatexal]

In everything I have read power9 is power hungry and perf per watt compared to intel x86 is not there. Am I wrong? I wonder if they could have saved money and power going EPYC for example.

 

[Evan]

Power hungry maybe but depends how you run them and what for, they are a pretty special cpu and for the right workloads will leave x86 way behind in performance (cost not)

Ignoring everything else, these big systems shall we say subsidised by government allow IBM to even make power cpu design and manufacture (3rd party foundry anyway) possible, that’s important for competition and innovation.

Not related the Power cpu usage in HPC applications but can anybody tell me there is a hypervisor that comes even close to what IBM offers ?
I may just be showing my ignorance of KVM, Hyper-V, ESX but do any of them easily offer NPIV SAN virtulisation and also simple boot off SAN for virtual systems ?