Inspur AGX-5 and Our SC18 Discussion with the Company

[Patrick Kennedy]

We got to see the Inspur AGX-5, the company's high-volume derivative of the NVIDIA HGX-2 at SC18 supporting 16x NVIDIA Tesla V100 GPUs today with NVSwitch and Volta-Next SXM3 modules in the future. We also take some time at SC18 to talk to the company about its AI plans

The post Inspur AGX-5 and Our SC18 Discussion with the Company appeared first on ServeTheHome.

Continue reading...

 

[LukeP]

why are all these servers 2u? why cant they just make them louder and halve them to 1u? there is nothing but fans and heatsinks in the top half.
(please pay me 100k if you steal this idea)

 

[Patrick]

Cooling more than 8x V100's + 2 CPUs, 24 DIMMs, PCIe switches, and storage in a 2U chassis is not easy. Most with these configurations are 4U. Doing what the AGX-2 does in 2U is impressive. Cooling all of that in 1U is nearly impossible with air.

 

[LukeP]

you can with smart heatsinks design. see recent cray designs what they do in 1u. it requires a bit of thought though. most of these 2u designs just throw in a row of fast fans and call it a day.

depends how much you care about density.

 

[Patrick]

@LukeP sure https://www.servethehome.com/cray-shasta-cooling-by-coolit/

8x Rome CPUs which have lower TDP than V100's. That is liquid cooled not air cooled.

A complete system also will have 100Gb NICs CPUs RAM and NVMe

4kW per U is not easy to cool. It is more efficient to do 4x1U in a 4U box than 1U servers because fans start having to spin so fast.


I cannot talk about the specifics. One of the large server OEMs had a NDA booth at SC18 where they were trying to do 8x 350-400W parts in 2U. They had massive copper heatsinks in these machines.

 

[LukeP]

i know what youre saying. i just don't assume the past is necessarily the best achievable. every layout has been the same for ages. i just think they could do a re-think. it would require the vendor making bespoke all parts of the case. which on the downside would introduce much higher costs.