AMD has kept details about its upcoming Epyc product family remarkably close to its chest. A recent leak (now deleted) at the publicly available Open Benchmarking database shows a tough competition between AMD’s upcoming 7nm Epyc CPUs and Intel’s equivalent Xeon products. Intel CEO Bob Swan has referred to AMD as offering increased competition in the back half of 2019, particularly in data center, so these figures aren’t automatically surprising — unless, of course, you remember the era just a few years ago when AMD’s market share in servers was basically zero.
According to the text of the now-deleted leak (picked up by THG before it went down), the AMD Epyc 7742 is a 64-core CPU with 128 threads, 256MB of L3 cache, a TDP of 225W, and a base / boost clock of 2.25GHz and 3.4GHz, respectively. The already-launched Epyc 7601 is a 32C/64T, 180W TDP CPU, with 64MB of L3 and a nearly-identical 2.2GHz base / 3.4GHz boost clock. The Xeon Platinum 8280 is 28C/56T, 2.7GHz base, 4GHz boost, and a 205W TDP, while the Xeon Gold 6138 (included for reference as well) is 20C/40T, 2GHz / 3.7GHz, and a 125W TDP.
If these rumors are accurate, AMD has managed to double core count and very slightly increase clock within a 1.25x larger TDP envelope. I am not sure what the “RDY1001C” refers to at the bottom of the results, though this configuration is always the fastest of the listed. Googling the term turned up no results.
There are more tests at THG than we’ve reproduced here; check their article for full results. And, as always, treat all results with a big ol’ bucket of caution. These are leaked results. Even if accurate, they may reflect engineering samples that are not representative of final performance.
SVT is a video encoder that’s heavily optimized for Intel CPUs, but optimizations for Intel chips often work well for AMD CPUs as well, and we certainly see that here. None of the encodes seem to scale particularly well when adding more cores, so we’re not going to try to make sense of the dualie figures. A single 7742 is significantly faster than the Xeon Platinum 8280 and the 7742 is more than twice as fast as the 7601.
In HEVC, the performance figures change. Here, Intel and AMD are at parity overall, but the 7742 is a huge uplift over and above the Epyc 7601.
POV-Ray 3.7 does scale with increased thread counts, but the gain from 1x CPU to 2x CPUs is much smaller from the 7742 as compared to the 7601. AMD only picks up about 24 percent more performance from adding another 64 cores, compared to 42 percent scaling for the Xeon Platinum 8280. This difference in scaling means that a pair of dual Xeon 8280’s nearly match a pair of Epyc 7742’s, even though one Epyc 7742 is significantly faster than one Xeon Platinum 8280.
Blender, and rendering more generally, are tests that AMD CPUs generally excel at. AMD decisively wins this test, though interestingly, we also see signs of significantly improved scaling for the Intel CPUs. This may simply reflect the fact that the Intel CPUs have far fewer cores. The Xeon Platinum 8280 is only a 28-core chip being compared to the performance of a 64-core chip. That’s a fairly massive advantage for AMD. Of course, there’s also the question of price and positioning — Intel has typically priced its Xeons far above AMD’s Epyc CPUs, and we tend to prioritize comparing on price above other factors.
Readers should, however, be aware that we may be seeing scaling issues on the AMD CPUs because of the sheer number of cores — 128C/256T, while the Xeon Platinum CPUs are only fielding 56 cores in a 2S configuration. The applications themselves may not scale well at these kinds of thread counts.
If these figures are accurate, they suggest AMD’s 7nm Epyc will be a significant challenge for Intel across a wider range of markets — which is pretty much exactly what we expected based on third-generation Ryzen and AMD’s previous statements about Epyc 2. Factor in Bob Swan’s acknowledgment of an increased competitive market, and we have a scenario teed up in which Intel will cut its Xeon prices, either by directly trimming them or when it launches Cooper Lake (currently expected in the first half of 2020). Intel’s CPU prices have historically run much higher than AMD’s, but it’s difficult to know exactly how much higher, because the company’s list prices (the best indicator we have to go on) don’t reflect what its volume customers actually pay.
If AMD’s Rome is as good as it looks, we should see increased OEM adoption of the part compared to first-generation Epyc, as well as some reaction from Intel. It can take server customers multiple product generations to move to new vendors, but they do eventually take notice.