So we have been production benchmarking two AMD EPYC servers today. A Quanta 1U2Node chassis with a 7551P in one slot and a 7401P in the other. 512GB of memory, 25 Gbit/s Mellanox Card.
A production benchmarking installs a production puppet class, sets up nginx queue monitoring and then gradually increase load balancer weights until the nginx grows a queue of request backlog. At which point we call it a day and return to default weights for the box.
Systems under test have been four Intels, a E5-2620v4 and a Silver 4110, a E5-2690v4 and a Gold 6132. The Intels are of course all 2P.
With our memory intense production workload, the Xeon E5 and Xeon XP actually perform equally, so the 2620 and 4110 drop at 9000 workunits, the 2690 and 6132 at 15000, each.
The 7401P did 13000, and the 7551P terminated at 15000, but it ran into a worker limit at a load at 50 or so, so we know this is not the absolute ceiling.
So you have a $2000 7551P wiping the floor with two 6132 ($4200 list), and a $1000 7401P doing 44% better than $1000 of lower bin Intel. Notably, latency with AMD is 20% higher due to lower clock. But unlike lower bin Intel, AMD performance under load (all cores busy) is almost stable, where Intel clocks down 25%. So while slower, the CPU is harder, performance under load is as stable as a high bin part. This is good for VM work and K8s.
All in all, we need to go into vendors now and see what kind of 1P AMD is on the menu.
Also, AMD, ship me a 1P Rome board. Any form factor or vendor will do at this stage.