The 64 Core Threadripper 3990X CPU Review: In The Midst Of Chaos, AMD Seeks Opportunity
by Dr. Ian Cutress & Gavin Bonshor on February 7, 2020 9:00 AM ESTAMD 3990X Against $20k Enterprise CPUs
For those looking at a server replacement CPU, AMD’s big discussion point here is that in order to get 64 cores on Intel hardware is relatively hard. The best way to get there is with a dual socket system, featuring two of its 28-core dies at a hefty $10k a piece. AMD’s argument is that users can consolidate down to a single socket, but also have better memory support, PCIe 4.0, and no cross-memory domain issues.
AMD 3990X Enterprise Competition | |||
AnandTech | AMD 3990X |
AMD 7702P |
Intel 2x8280 |
SEP | $3990 | $4450 | $20018 |
Cores/Threads | 64 / 128 | 64 / 128 | 56 / 112 |
Base Frequency | 2900 | 2000 | 2700 |
Turbo Frequency | 4300 | 3350 | 4000 |
PCIe | 4.0 x64 | 4.0 x128 | 3.0 x96 |
DDR4 Frequency | 4x 3200 | 8x 3200 | 12x 2933 |
Max DDR4 Capacity | 512 GB | 2 TB | 3 TB |
TDP | 280 W | 200 W | 410 W |
Unfortunately I was unable to get ahold of our Rome CPUs from Johan in time for this review, however I do have data from several dual Intel Xeon setups that I did a few months ago, including the $20k system.
This time with Corona the competition is hot on the heels of AMD's 64-core CPUs, but even $20k of hardware can't match it.
The non-AVX verson of 3DPM puts the Zen 2 hardware out front, with everything else waiting in the wings.
When we add in the AVX-512 hand tuned code, the situation flips: Intel's 56 cores get almost 2.5x the score of AMD, despite having fewer cores.
Blender doesn't seem to like the additional access latency from the 2P systems.
For AES encoding, as the benchmark takes places from memory, it appears that none of Intel's CPUs can match AMD here.
For the 7-zip combined test, there's little difference between AMD's 32-core and 64-core, but there are sizable jumps above Intel hardware.
Verdict
In our tests here (more in our benchmark database), AMD's 3990X would get the crown over Intel's dual socket offerings. The only thing really keeping me back from giving it is the same reason there was hesitation on the previous page: it doesn't do enough to differentiate itself from AMD's own 32-core CPU. Where AMD does win is in that 'money is less of an issue scenario', where using a single socket 64 core CPU can help consolidate systems, save power, and save money. Intel's CPUs have a TDP of 205W each (more if you decide to use the turbo, which we did here), which totals 410W, while AMD maxed out at 280W in our tests. Technically Intel's 2P has access to more PCIe lanes, but AMD's PCIe lanes are PCIe 4.0, not PCIe 3.0, and with the right switch can power many more than Intel (if you're saving 16k, then a switch is peanuts).
We acknowledge that our tests here aren't in any way a comprehensive test of server level workloads, but for the user base that AMD is aiming for, we'd take the 64 core (or even the 32 core) in most circumstances over two Intel 28 core CPUs, and spend the extra money on memory, storage, or a couple of big fat GPUs.
279 Comments
View All Comments
james4591 - Friday, February 7, 2020 - link
It's only useless to people who don't know how to use it.SanX - Friday, February 7, 2020 - link
Useless is probably Ian's own test of 3D particle movement which demonstrates just one single feature of AVX instruction set. Was there any real life or synthetic tests which use AVX512 or AVX256 to get a clue what improvement it really gives? Watching pure single instruction speed improvement having 12x speed boost is pathetic has no sense. Compile with and without AVX and show us if it gets any meaningful speedup at east from something useful like Gauss elimination Ax=B solvernsmeds - Saturday, February 8, 2020 - link
For AVX influence on Gaussian elimination take a look at HPL. It does have a huge impact as the matrix update in every reduction step fits extremely well to AVX512. See http://www.top500.org If you want to experiment there are implementations available from Intel (MKL) and AMD (BLIS, libFLAME and HPL-FLAME from github) see eghttps://www.google.com/url?sa=t&source=web&...
Several scientific workloads fit well to AVX512 usage but definitely not all. Adapting a code to effectively use AVX512 can be labour intensive though and for research purposes may make the code harder to adapt for researchers. It may be more important that researchers can implement new ideas easily than run at optimal efficiency. And only for large problem sizes the effort of putting data nicely in memory may be amortized by the the speedup from AVX512.
realbabilu - Saturday, February 8, 2020 - link
Just checking the Openblas, Intel MKL, and Blis with lapack DGETRI inverse matrix and Crout inverse of fortran polyhedron benchmark. The march=core-avx512 and march=core-avx2 help the calculation faster. From 30s to 17s in 9750H.nt300 - Friday, February 7, 2020 - link
Another phenomenal processor from AMD.Catering to a market that wants such processors.
Anybody claiming these are useless don't understand the Computer Industry.
AshlayW - Saturday, February 8, 2020 - link
Another useless comment from a tool.Zak90 - Saturday, February 8, 2020 - link
?timecop1818 "Another useless processor from AMD"-->Idiot!!!
levizx - Saturday, February 8, 2020 - link
Another waste of food water and air waste skin suit spottedevernessince - Sunday, February 9, 2020 - link
AnandTech would be better off disabling article comments if they aren't going to bother moderating blatant trolls.nt300 - Sunday, February 9, 2020 - link
Another Superior Processor from AMD.Bar None, AMD annihilates Intel in everything. AMDs price/performance is KING.