AMD EPYC processors added to Cray CS500 product line

[HEXUS]

Supercomputer maker’s CS500 cluster system packs four dual-socket nodes in a 2U chassis.

Read more.

[shaithis]

Good density for a 2U

[aidanjt]

That's a huge boon for AMD. Well done chaps.

[lumireleon]

intel reports show Epyc processors perform EXTREMELY POORLY on Ethernet because only one CCM not all has the direct access to Ethernet connection I-O and the rest of the CCMs have to forward all their request through infinity fabric hampering connection speeds MISERABLY. Because of such a report plus lack of AVX512 would you recommend AMD Epyc processors to any one?

[afiretruck]

intel reports show Epyc processors perform EXTREMELY POORLY on Ethernet because only one CCM not all has the direct access to Ethernet connection I-O and the rest of the CCMs have to forward all their request through infinity fabric hampering connection speeds MISERABLY. Because of such a report plus lack of AVX512 would you recommend AMD Epyc processors to any one?

Would I believe a damning report coming directly from their flailing competitor? Definitely not, given their track record of smear campaigns against AMD.

Also, source please.

[lumireleon]

https://www.tomshardware.com/news/amd-intel-epyc-xeon-benchmarks,35993.html

[Lanky123]

https://www.tomshardware.com/news/amd-intel-epyc-xeon-benchmarks,35993.html

"Intel has released its internal benchmarking results" basically invalidates everything that is concluded afterwards. Intel are the very last people on Earth you would go to for independent performance analysis of AMD processors. And likewise AMD for Intel processors in the reverse scenario. The only party in this story with both independence and expertise are Cray...and funnily enough they've decided EPYC is worth using.

[Bagpuss]

Will they use one to render the sequel to 'The Last Starfighter' ?

)

[3dcandy]

https://www.tomshardware.com/news/amd-intel-epyc-xeon-benchmarks,35993.html

"Intel has released its internal benchmarking results" basically invalidates everything that is concluded afterwards. Intel are the very last people on Earth you would go to for independent performance analysis of AMD processors. And likewise AMD for Intel processors in the reverse scenario. The only party in this story with both independence and expertise are Cray...and funnily enough they've decided EPYC is worth using.

Also if you check the article above it does say they use a 2u compute node and a 2u service node. Any ethernet or other problems would be mitigated by the service node

[Tabbykatze]

Isn't it general knowledge that any publications made by a competitor are going to be done in the worst way possible to cast as much shade as possible? It's like if AMD released the details about Intel Spectre and Meltdown, it would absolutely have to be taken with salt until it's proven.

But you do like to cast AMD in a bad light, Lumireleon, it's not even Devils advocacy at this point.

On topic, this is great news for EPYC to get into more and more places will add credit when trickling down into commercial purchases.

[3dcandy]

intel reports show Epyc processors perform EXTREMELY POORLY on Ethernet because only one CCM not all has the direct access to Ethernet connection I-O and the rest of the CCMs have to forward all their request through infinity fabric hampering connection speeds MISERABLY. Because of such a report plus lack of AVX512 would you recommend AMD Epyc processors to any one?

I don't see AVX512 missing as a problem. Most software doesn't use it and even Intel takes a huge latency hit when running it. Plus it also downclocks cores when running AVX512. So all in all I'd say AMD has a more balanced outlook this time round. Factor in the rather substantial cost savings...and Intel looks worried

[Corky34]

intel reports show Epyc processors perform EXTREMELY POORLY on Ethernet because only one CCM not all has the direct access to Ethernet connection I-O and the rest of the CCMs have to forward all their request through infinity fabric hampering connection speeds MISERABLY. Because of such a report plus lack of AVX512 would you recommend AMD Epyc processors to any one?

I can't find that exact statement from the link you provided later but i suspect Intel are playing with words or going out of their way to cause an issue that wouldn't normally occur, each CCX (group of four cores) have their own CCM's (Cache-Coherent Master) that transmit data onto the Scalable Data Fabric (SDF) plane, if that data needs to go off package (SATA, PCIe, Infinity Fabric to another Epyc processor, Ethernet, USB, etc, etc.) then the Coherent AMD socket Extender (CAKE) module encodes local SDF requests into individual packets and ships them over any of the SerDes interfaces (Linky)

It's true that only one CAKE on a package has direct access to the Infinity Fabric InterSocket controller (IFIS) that can be MUX'ed with an Ethernet port but it seems for that to be a problem you'd have to actual go out of your way to make it so.