Article [VIdeocardz] Milan final specs

[jpiniero]

AMD 3rd Gen EPYC "Milan" full specifications confirmed - VideoCardz.com

AMD EPYC naming explained If a product needs a slide explaining its name then something must have clearly gone wrong. Not only do monitors have weird and indistinguishable names. Both Intel and AMD server processor naming is just as complicated to understand. For AMD EPYC server processors that...

videocardz.com

Base clocks are the same or a tad lower than Rome but even the ones lower should be okay. Now there is only two 64 core models and one 48 core model. Might be trying to push the 32 core models more (ironically enough) to maximize profits?

 

[Bam360]

I mean yeah, base frequency seems a tad lower on surface, but when we are in the 2-2.5GHz range, 200MHz less is a much bigger deal than 200MHz less in the 4GHz range.
For comparison with same TDP:
225W 7742 2.25GHz vs 7713 2GHz, this is 12% higher frequency for 7742.
280W 7H12 2.6GHz vs 7763 2.45GHz, this is 6% higher frequency for 7H12.

More proof that Zen3 is better the higher you are in the V/F curve, and the improvement from Zen2 diminishes the lower you are in the curve. Lower power limit means you are forced to go lower in the V/F curve.
Of course, base clock doesn't mean it will go that low in all MT workloads, but it is a clear indicator that efficiency won't improve much in some MT workloads, it depends on what the workload is.

 

[DrMrLordX]

It would be nice to see some Rome vs. Milan head-to-head benchmarks. Too bad AMD has been shipping Milan to various hyperscalars for months without anyone having a sample to bench publicly.

 

[itsmydamnation]

I mean yeah, base frequency seems a tad lower on surface, but when we are in the 2-2.5GHz range, 200MHz less is a much bigger deal than 200MHz less in the 4GHz range.
For comparison with same TDP:
225W 7742 2.25GHz vs 7713 2GHz, this is 12% higher frequency for 7742.
280W 7H12 2.6GHz vs 7763 2.45GHz, this is 6% higher frequency for 7H12.

More proof that Zen3 is better the higher you are in the V/F curve, and the improvement from Zen2 diminishes the lower you are in the curve. Lower power limit means you are forced to go lower in the V/F curve.
Of course, base clock doesn't mean it will go that low in all MT workloads, but it is a clear indicator that efficiency won't improve much in some MT workloads, it depends on what the workload is.

you have to remember that the base will be for heavy AVX only and depending on the exact workload Zen3 can see some very big perf per clock increases vs Zen2.

What we really need to see is the "general" code clock/ performance comparison.

 

[moinmoin]

It would be nice to see some Rome vs. Milan head-to-head benchmarks. Too bad AMD has been shipping Milan to various hyperscalars for months without anyone having a sample to bench publicly.

STH does have Milan already, but they wait for AMD to officially launch it.

 

[LightningZ71]

I suspect that the performance uplift for the full cache 32 core products might be significant. That was a rough case for ROME, having only two cores per CCX to achieve that. Now, that's half the CCX domains for the same cache, and more work done per clock while living comfortably in the same maximum power envelope. The 64 core parts might not improve a whole lot though. Maximum clocks seem to be harder on thermals, and I'm not confident that the needed reduction in clocks will be overcome well with higher IPC. Chatty workloads will like the fewer CCXs though.