Benchmarks I
Performance evaluation
Noting the progress AMD has made with its newest Epyc processors, and with due knowledge that the Daytona reference system carries cutting-edge CPU hardware, the question of how to appropriately contextualise its performance becomes paramount.
Rival Intel's premium Xeons, being obvious competitors, use either the Cascade Lake SP or Cascade Lake AP blueprint. The resulting chips are productised as 2nd Generation Intel Xeon Scalable.
Cascade Lake AP, or Xeon Platinum 9200-series, top out at 56
cores and 112 threads, thus providing, on paper, a similar
topology to AMD's Epyc 7742. The problem with Cascade Lake AP is
that you cannot purchase the CPUs individually; they need to be
configured as part of a mostly-built system known as a Compute
Module. The top-line part, Xeon Platinum 9282, carries a
staggering 400W TDP, making cooling difficult in a 2U server rack.
There's also no concrete word on pricing, or any meaningful
performance numbers, though we expect it to be significantly more
expensive than the $10,000 charged for the Xeon Platinum 8280.
The Xeon Platinum 8280, a 28-core, 56-thread processor,
representing the fastest model in the Cascade Lake SP family, is a
better comparison fit for premium Epyc processors from a pricing
and implementation perspective. AMD's Epyc 7742 still has a lot
more cores and threads - 64C128T, remember - and is over 30 per
cent cheaper, at $6,950. Much like the desktop space, AMD provides
more cores and threads than Intel at similar, or lower, pricing
levels. The difference here, however, is more than 2x for a lesser
sum. That's not to say all cores are created equal, and it's
improper to draw tenuous conclusions purely from a core-and-thread
specification sheet.
But given that pricing and thermals remain key in the TCO
argument, it is therefore, on the face of it, reasonable to
compare the performance of the dual-socket Epyc 7742 Daytona
system against a dual-socket Xeon Platinum 8280. Configured with
otherwise identical hardware, the AMD system ought to be a fair
bit cheaper, too, due to the savings made by the processor cost.
Of course, these are retail prices, and many of Intel's
larger-scale customers are unlikely to pay the ticket price for a
significant server install. On the other hand, the Intel-based
server ought to draw a bit less power - the dual processors
consume 410W vs. AMD's 450W.
Software optimisations play a critical role in defining server performance, as well, and so the AMD Daytona server is benchmarked on Ubuntu 19.04 with the latest security mitigations in place. Our plan was to run a bunch of standard server-type tests on both the AMD and Intel platforms with a view of extrapolating conclusions of probable performance in the field. The rather large hiccup, however, has been the inability to source a dual-socket Intel Xeon Platinum 8280 server in time for this editorial: Intel UK is supplying one later this week, and we'll run the benchmarks in the labs and update the ensuing graphs as soon as possible.
As a short-term fix, we're taking our results obtained from the dual Epyc 7742-equipped AMD Daytona system and comparing them with the best published results from said 2P Xeon Platinum 8280 server from the Phoronix Test Suite 8.8.1, which runs exactly the same tests with respect to build versions. Going down this route ought to provide an adequate frame of reference until we can confirm results empirically.
The Stream memory benchmark is run via the default parameters that don't take each company's optimisations into account. AMD uses an eight-channel DDR4-3200 setup while Intel's Xeons run with a 6-channel DDR4-2933, as per spec.
There's not much in it, really, and AMD's lead is provided by a combination of faster memory and plain more channels. Running manufacturer-optimised versions of the same benchmark adds about 50 per cent to both AMD and Intel's scores.
