Sizing Up Servers: Intel's Skylake-SP Xeon versus AMD's EPYC 7000 - The Server CPU Battle of the Decade?
by Johan De Gelas & Ian Cutress on July 11, 2017 12:15 PM EST- Posted in
- CPUs
- AMD
- Intel
- Xeon
- Enterprise
- Skylake
- Zen
- Naples
- Skylake-SP
- EPYC
Memory Subsystem: Bandwidth
Measuring the full bandwidth potential with John McCalpin's Stream bandwidth benchmark is getting increasingly difficult on the latest CPUs, as core and memory channel counts have continued to grow. We compiled the stream 5.10 source code with the Intel compiler (icc) for linux version 17, or GCC 5.4, both 64-bit. The following compiler switches were used on icc:
icc -fast -qopenmp -parallel (-AVX) -DSTREAM_ARRAY_SIZE=800000000
Notice that we had to increase the array significantly, to a data size of around 6 GB. We compiled one version with AVX and one without.
The results are expressed in gigabytes per second.
Meanwhile the following compiler switches were used on gcc:
-Ofast -fopenmp -static -DSTREAM_ARRAY_SIZE=800000000
Notice that the DDR4 DRAM in the EPYC system ran at 2400 GT/s (8 channels), while the Intel system ran its DRAM at 2666 GT/s (6 channels). So the dual socket AMD system should theoretically get 307 GB per second (2.4 GT/s* 8 bytes per channel x 8 channels x 2 sockets). The Intel system has access to 256 GB per second (2.66 GT/s* 8 bytes per channel x 6 channels x 2 sockets).
AMD told me they do not fully trust the results from the binaries compiled with ICC (and who can blame them?). Their own fully customized stream binary achieved 250 GB/s. Intel claims 199 GB/s for an AVX-512 optimized binary (Xeon E5-2699 v4: 128 GB/s with DDR-2400). Those kind of bandwidth numbers are only available to specially tuned AVX HPC binaries.
Our numbers are much more realistic, and show that given enough threads, the 8 channels of DDR4 give the AMD EPYC server a 25% to 45% bandwidth advantage. This is less relevant in most server applications, but a nice bonus in many sparse matrix HPC applications.
Maximum bandwidth is one thing, but that bandwidth must be available as soon as possible. To better understand the memory subsystem, we pinned the stream threads to different cores with numactl.
| Pinned Memory Bandwidth (in MB/sec) | |||
| Mem Hierarchy |
AMD "Naples" EPYC 7601 DDR4-2400 |
Intel "Skylake-SP" Xeon 8176 DDR4-2666 |
Intel "Broadwell-EP" Xeon E5-2699v4 DDR4-2400 |
| 1 Thread | 27490 | 12224 | 18555 |
| 2 Threads, same core same socket |
27663 | 14313 | 19043 |
| 2 Threads, different cores same socket |
29836 | 24462 | 37279 |
| 2 Threads, different socket | 54997 | 24387 | 37333 |
| 4 threads on the first 4 cores same socket |
29201 | 47986 | 53983 |
| 8 threads on the first 8 cores same socket |
32703 | 77884 | 61450 |
| 8 threads on different dies (core 0,4,8,12...) same socket |
98747 | 77880 | 61504 |
The new Skylake-SP offers mediocre bandwidth to a single thread: only 12 GB/s is available despite the use of fast DDR-4 2666. The Broadwell-EP delivers 50% more bandwidth with slower DDR4-2400. It is clear that Skylake-SP needs more threads to get the most of its available memory bandwidth.
Meanwhile a single thread on a Naples core can get 27,5 GB/s if necessary. This is very promissing, as this means that a single-threaded phase in an HPC application will get abundant bandwidth and run as fast as possible. But the total bandwidth that one whole quad core CCX can command is only 30 GB/s.
Overall, memory bandwidth on Intel's Skylake-SP Xeon behaves more linearly than on AMD's EPYC. All off the Xeon's cores have access to all the memory channels, so bandwidth more directly increases with the number of threads.
Facebook
Twitter
RSS
219 Comments
View All Comments
oldlaptop - Thursday, July 13, 2017 - link
Why on earth is gcc -Ofast being used to mimic "real-world", non-"aggressively optimized"(!) conditions? This is in fact the *most* aggressive optimization setting available; it is very sensitive to the exact program being compiled at best, and generates bloated (low priority on code size) and/or buggy code at worst (possibly even harming performance if the generated code is so big as to harm cache coherency). Most real-world software will be built with -O2 or possibly -Os. I can't help but wonder why questions weren't asked when SPEC complained about this unwisely aggressive optimization setting...peevee - Thursday, July 13, 2017 - link
"added a second full-blown 512 bit AVX-512 unit. "Do you mean "added second 256 ALU, which in combination with the first one implements full 512-bit AVX-512 unit"?
peevee - Thursday, July 13, 2017 - link
"getting data from the right top node to the bottom left node – should demand around 13 cycles. And before you get too concerned with that number, keep in mind that it compares very favorably with any off die communication that has to happen between different dies in (AMD's) Multi Chip Module (MCM), with the Skylake-SP's latency being around one-tenth of EPYC's."1/10th? Asking data from L3 on the chip next to it will take 130 (or even 65 if they are talking about averages) cycles? Does not sound realistic, you can request data from RAM at similar latencies already.
AmericasCup - Friday, July 14, 2017 - link
'For enterprises with a small infrastructure crew and server hardware on premise, spending time on hardware tuning is not an option most of the time.'Conversely, our small crew shop has been tuning AMD (selected for scalar floating point operations performance) for years. The experience and familiarity makes switching less attractive.
Also, you did all this in one week for AMD and two weeks for Intel? Did you ever sleep? KUDOS!
JohanAnandtech - Friday, July 21, 2017 - link
Thanks for appreciating the effort. Luckily, I got some help from Ian on Tuesday. :-)AntonErtl - Friday, July 14, 2017 - link
According to http://www.anandtech.com/show/10158/the-intel-xeon... if you execute just one AVX256 instruction on one core, this slows down the clocks of all E5v4 cores on the same socket for at least 1ms. Somewhere I read that newer Xeons only slow down the core that executes the AVX256 instruction. I expect that it works the same way for AVX512, and yes, this means that if you don't have a load with a heavy proportion of SIMD instructions, you are better off with AVX128 or SSE. The AMD variant of having only 128-bit FPUs and no clock slowdown looks better balanced to me. It might not win Linpack benchmark competitions, but for that one uses GPUs anyway these days.wagoo - Sunday, July 16, 2017 - link
Typo on the CLOSING THOUGHTS page: "dual Silver Xeon solutions" (dual socket)Great read though, thanks! Can finally replace my dual socket shanghai opteron home server soon :)
Chaser - Sunday, July 16, 2017 - link
AMD's CPU future is looking very promising!bongey - Tuesday, July 18, 2017 - link
EPYC power consumption is just wrong. Somehow you are 50W over what everyone else is getting at idle. https://www.servethehome.com/amd-epyc-7601-dual-so...Nenad - Thursday, July 20, 2017 - link
Interesting SPECint2006 results:- Intel in their slide #9 claims that Intel 8160 is 2% faster than EPYC 7601
- Anandtech in article tests that EPYC 7601 is 42% faster than Intel 8176
Those two are quite different, even if we ignore that 8176 should be faster than 8160. In other words, those Intel test results look very suspicious.