Potentially Epyc performance
AMD is making some claims about the performance of various Epyc products' performance today, and the initial outlook is good. However, we do have to take issue with a couple of the choices AMD made on the way to its numbers. After compiling SPECint_rate_base2006 with the -O2 flag in GCC, AMD says observed a 43% delta between its internal numbers for the Xeon E5-2699A v4 and public SPEC numbers for similar systems produced using binaries generated by the Intel C++ compiler. In turn, AMD applied that 43% (or 0.575x) across the board to some publicly-available SPECint_rate_base scores for several two-socket Xeon systems.
It's certainly fair to say that SPEC results produced with Intel's compiler might deserve adjustment, but my conversations with other analysts present at the Epyc event suggests that a 43% reduction is optimistic. The -O2 flag for GCC isn't the most aggressive set of optimizations available from that compiler, and SPEC binaries generated accordingly may not be fully representative of binaries compiled in the real world.
Still, here are AMD's own numbers for two-socket systems. If we take the company's assumptions at face value, Epyc would appear to offer large (or even massive) boosts over Broadwell Xeons from top to bottom. The Epyc 7601 dual-socket system handily beats out the pair of 22-core, 44-thread Xeon E5-2699A v4 chips in these benchmarking conditions, and every Epyc chip duo enjoys wide margins of victory over the competition.
We wanted to see how the results might shake out with a different set of assumptions around the Intel SPEC results that AMD started with, though. Friend-of-TR David Kanter suggests that a 20% reduction to public SPEC numbers for Intel CPUs is fairer considering the impacts of Intel's compiler on the "libquantum" portion of the benchmark and the use of optimizing compilers in general. Accordingly, I rejiggered AMD's numbers with that in mind. (I've applied a 1/0.575x increase to AMD's own E5-2699A v4 results with a GCC binary to keep things consistent before applying David's suggested handicap.)
With Intel's compiler results adjusted this way, two-socket Broadwell systems generally come out slightly behind to moderately behind versus Epyc parts when running SPECint_base_rate2006, especially at the lower end of the market. AMD's systems still win out, but the differences in performance seem closer to our real-world examinations of Zen and Broadwell performance.
Since code compiled with GCC and ICC coexists in the wild, the truth of Epyc performance likely lies somewhere between these poles. SPECint_rate_base2006 is just one synthetic benchmark, as well, and we're eager to see real-world testing results with applications like databases, web servers, and so on.
It's also worth noting that Skylake server parts are coming (although we're not sure exactly when yet), and those parts will almost certainly compare more favorably to their Epyc counterparts when more details become available. Still, the fact that Epyc merits direct comparisons to Xeon performance is great news. We'll have more about AMD's server parts to share soon.