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 Prosumer CPUs
The first set of consumers that will be interested in this processor will be those looking to upgrade into the best consumer/prosumer HEDT package available on the market. The $3990 price is a high barrier to entry, but these users and individuals can likely amortize the cost of the processor over its lifetime. To that end, we’ve selected a number of standard HEDT processors that are near in terms of price/core count, as well as putting in the 8-core 5.0 GHz Core i9-9900KS and the 28-core unlocked Xeon W-3175X.
AMD 3990X Consumer Competition | ||||||
AnandTech | AMD 3990X |
AMD 3970X |
Intel 3175X |
Intel i9- 10980XE |
AMD 3950X |
Intel 9900KS |
SEP | $3990 | $1999 | $2999 | $979 | $749 | $513 |
Cores/T | 64/128 | 32/64 | 28/56 | 18/36 | 16/32 | 8/16 |
Base Freq | 2900 | 3700 | 3100 | 3000 | 3500 | 5000 |
Turbo Freq | 4300 | 4500 | 4300 | 4800 | 4700 | 5000 |
PCIe | 4.0 x64 | 4.0 x64 | 3.0 x48 | 3.0 x48 | 4.0 x24 | 3.0 x16 |
DDR | 4x 3200 | 4x 3200 | 6x 2666 | 4x 2933 | 2x 3200 | 2x 2666 |
Max DDR | 512 GB | 512 GB | 512 GB | 256 GB | 128 GB | 128 GB |
TDP | 280 W | 280 W | 255 W | 165 W | 105 W | 127 W |
The 3990X is beyond anything in price at this level, and even at the highest consumer cost systems, $1000 could be the difference between getting two or three GPUs in a system. There has to be big upsides here moving from the 32 core to the 64 core.
Corona is a classic 'more threads means more performance' benchmark, and while the 3990X doesn't quite get perfect scaling over the 32 core, it is almost there.
The 3990X scores new records in our Blender test, with sizeable speed-ups against the other TR3 hardware.
Photoscan is a variable threaded test, and the AMD CPUs still win here, although 24 core up to 64 core all perform within about a minute of each other in this 20 minute test. Intel's best consumer hardware is a few minutes behind.
y-cruncher is an AVX-512 accelerated test, and so Intel's 28-core with AVX-512 wins here. Interestingly the 128 cores of the 3990X get in the way here, likely the spawn time of so many threads is adding to the overall time.
GIMP is a single threaded test designed around opening the program, and Intel's 5.0 GHz chip is the best here. the 64 core hardware isn't that bad here, although the W10 Enterprise data has the better result.
Without any hand tuned code, between 32 core and 64 core workloads on 3DPM, there's actually a slight deficit on 64 core.
But when we crank in the hand tuned code, the AVX-512 CPUs storm ahead by a considerable margin.
We covered Digicortex on the last page, but it seems that the different thread groups on W10 Pro is holidng the 3990X back a lot. With SMT disabled, we score nearer 3x here.
Luxmark is an AVX2 accelerated program, and having more cores here helps. But we see little gain from 32C to 64C.
As we saw on the last page, POV-Ray preferred having SMT off for the 3990X, otherwise there's no benefit over the 32-core CPU.
AES gets a slight bump over the 32 core, however not as much as the 2x price difference would have you believe.
As we saw on the previous page, W10 Enterprise causes our Handbrake test to go way up, but on W10 Pro then the 3990X loses ground to the 3950X.
And how about a simple game test - we know 64 cores is overkill for games, so here's a CPU bount test. There's not a lot in it between the 3990X and the 3970X, but Intel's high frequency CPUs are the best here.
Verdict
There are a lot of situations where the jump from AMD's 32-core $1999 CPU, the 3970X, up to the 64-core $3990 CPU only gives the smallest tangible gain. That doesn't bode well. The benchmarks that do get the biggest gains however can get near perfect scaling, making the 3990X a fantastic upgrade. However those tests are few and far between. If these were the options, the smart money is on the 3970X, unless you can be absolutely clear that the software you run can benefit from the extra cores.
279 Comments
View All Comments
Pessimism - Monday, February 10, 2020 - link
Can it run Crysis?Mugur - Tuesday, February 11, 2020 - link
Yes, see LTT video. In software.XiZeL - Tuesday, February 11, 2020 - link
Doesnt the Ryzen 9 3950X have 64Mb L3 cache? the table states 32Mbchrkv - Tuesday, February 11, 2020 - link
What Windows version were using? I see claims that since version 18362.535 Windows 10 shows 1 socket for 3990X - look for "18362.535" here https://translate.google.com/translate?hl=&sl=...Betonmischer - Tuesday, February 11, 2020 - link
That's right. Here's a proof that it does:https://imgur.com/G2VqgoU
29a - Friday, February 14, 2020 - link
AT can't be bothered by the little stuff like OS patches when they're doing an AMD review. Haven't you seen any of their AMD launch reviews, they screw every one of those up.TokyoQuaSar - Wednesday, February 12, 2020 - link
Very interesting article, I hope you can update it with data from an Epyc 77xx (7702 or 7742). Would be nice to have a head to head comparison, if possible a test with equal frequencies and some tests on software that are very dependant on memory bandwidth, to see the influence of the 8 channels aside from the amount of memory.vivs26 - Wednesday, February 12, 2020 - link
Are there any linux distros for desktop that support more than 64 cores?TokyoQuaSar - Thursday, February 13, 2020 - link
Not sure exactly but this test was done on Ubuntu and they don't mention any problem coming from the OS but rather from the tested software:https://techgage.com/article/amd-ryzen-threadrippe...
They do say the number of cores scale better on Linux.
HikariWS - Thursday, February 13, 2020 - link
Very nice article! I've finally seen use cases where high core count counts!Indeed you should start adding some Lix benchs, I wonder how the kernel itself would handle that many cores. And of course M$ has to fix at least Pro Workstation.
I'd rly like to see a review comparing HT enabled and disabled, around 8C. Is it worth disabling or enabling HT on my 9900KS? Under full load, is there difference in performance and consumption?
How much performance the virtual cores have over physical ones? Do work load on one type affect the other? If we force affinity on one and leave its pair idle, and then put a full work load on it, how the tested core performs?