AMD’s new 32 – Core / 64 – Thread Threadripper 3970 x continues AMD’s 2019 trend of sweeping the field in desktop and server processors. In recent weeks, Ars has tested Threadripper head-to-head versus Intel’s top-of-the-line i9 – 10980 XE High End Desktop (HEDT) CPU, as well as its i9 – 9900 KS gaming CPU. To nobody’s surprise, the Threadripper is faster — alotFaster — than either, although with some caveats.
When comparing the rest of the Ryzen 3000 line to Intel’s 2019 desktop CPU lineup, one of the standout metrics is thermal design power (TDP). Non-threadripper Ryzen 3000 CPUs meet or beat the Intel desktop lineup on performanceandTDP, which means quieter, cooler systems that don’t cost as much to keep running. All that changes once you leave the “normal” desktop line and go Threadripper. With Threadripper, AMD is clearly far more concerned with raw power than niceties like running quiet or cool.
The above chart shows the whole system power draw as measured by a Kill-a-Watt power meter. Power draw was tested at minimum shown idling for one minute at the Windows 10 desktop and maximum during Passmark all-core CPU benchmarking.
Just for fun, we also did a little testing of the Threadripper with Precision Boost Overdrive (PBO) automatic overclocking enabled. The difference between default and PBO boosted clocks is more clear at the power meter than it is in the benchmarks themselves. Maximum power draw shoots up by about 20 – 25% with PBO enabled, but the actual performance hardly changed in most benchmarks.
|Specs at a glance: Threadripper 3970 x, as tested|
|(GB Corsair Dominator Platinum RGB DDR4) – $ 400 atAmazon|
|Motherboard||ASUS ROG Zenith II Extreme motherboard – $ (atAmazon|
|Praxis Wetbench test chassis – $ 200 atAmazon|
|Price as tested||≈ $ 4, 165|
In the above performance charts, we go head-to- head with some Intel parts. Some test results were sourced from public leaderboards and are marked by an asterisk next to the CPU name. Results for Threadripper 3970 x, i9 – 10980 XE, and i9 – 9900 KS were taken directly by Ars. The results show you just what you’re buying with that extravagant power draw — a system that kicks dirt in Intel’s HEDT parts and runs away laughing.
For general-purpose CPU performance evaluation, we like Passmark’s CPU test. Threadripper (x doesn’t) ********************************** (quite) double i9 – 10980 XE’s score there, but it comes pretty close . Moving on to Cinebench R 20 —Which is more appropriate for specific graphics-rendering and content creation evaluation — the Threadripper didn’t just beat Intel’s HEDT line, it even eked out a narrow victory over the top-of-the-line Xeon Platinum 8168 .
That “victory” is more of a draw within the margin of error, but the reason it matters is clear in the next chart, where we show the Cinebench R 20 performance-per-dollar by dividing each CPU’s score by its retail price. Normally, you’d expect to pay a pretty sharp penalty moving from HEDT-grade to high-end server-grade CPUs — but the 3970 x delivers equivalent performance to the 8168 at under a third of the cost. Meanwhile, although it’s roughly double the cost of the 10980 XE, it’s also roughly double the performance.
AI inference workloads
The one place that Threadripper 3970 x did (not) get a victory is in AI inference workloads. We tested Threadripper against the i9 – 10980 XE using the (OpenVINO) toolkit and (AIXPRT) ‘s reference configurations, and the advantage conferred by Intel’s Deep Learning Boost (DLB) x 86 extension was stark. Although it’s a much less performant processor for general-purpose workloads, the i9 – 10980 XE was able to deliver double to triple the image recognition inference throughput.
In testing with Tensorflow (results not shown), the 3970 x performed roughly on-par with The 10980 XE — which is still a serious upset, from a CPU that’s normally half the performance (and half the price) of the 3970 x.
Listing image by Jim Salter