AMD Threadripper 2950X
Few things make Jonni Bidwell as happy as 16-cores all maxed out working on chess problems.
Few things make Jonni Bidwell as happy as 16-cores all maxed out working on his latest chess problems.
Back in LXF229 we reviewed AMD’s Threadripper 1950X, a powerhouse of a machine that showed what the Zen architecture is capable of. Since then we’ve had two new lines of Ryzen chips: the 2000 series that launched in April, and then in August when this second-generation Threadripper series began shipping.
It would be nice if we could’ve gotten our grubby paws on the flagship of this new series, the 32-core 2990WX (which retails at $1,799, £1639), but this time we were only granted its smaller sibling, the 2950X (RRP $899, £849). Still, this makes for a good side-by-side comparison, so we thought we’d revisit our 1950X benchmarks and check out how much more performance AMD have eked out of 16 cores and 32 threads.
Base clocks are up from 3.4GHz to 3.5GHz, and boost frequencies have been, er, boosted from 4.0GHz to 4.4GHz, so it seems reasonable to expect things to be nippier. The shrink from a 14nm process to 12nm means that despite boosted clocks, the chip’s total power draw remains at 180W while cache latencies have been reduced. AMD’s propaganda touts the benefits of the revised XFR2 and Precision Boost 2 overclocking features, which can boost individual core clocks on demand. There’s also Precision Boost Overdrive which allows for yet more aggressive overclocking.
Running the numbers
When we first ran PhoronixTestSuite we saw some odd results, but a little investigating revealed that many of the tests had been updated, so comparing our old results to new made no sense. Recreating the old tests proved troublesome, so where this has happened, we’ve used new test data from openbenchmarking.org to provide 1950X readings. Our results are for the most part consistent with those on Phoronix (see http://bit.
ly/phoronix-2950x), which is reassuring… but benchmarks can be untrustworthy. We used a stock Ubuntu 18.04.1 install, freshly updated. Phoronix went all out and recompiled everything with GCC 8.2.0 to capitalise on the Zen enhancements there, which may explain why our x265 benchmark was slightly slower. It’s worth noting that although AMD chips are thought to be immune to Spectre v3 (aka Meltdown), new Spectre-like vulnerabilities are being discovered all the time, and mitigations to these may slow things down. On the other hand, improvements to already implemented mitigations may speed them up again. We tried running the benchmarks with the nospectre_v2 kernel cmdline, but this had no discernible effect.
Comparing results with Intel silicon is non-trivial. Costwise, the 2950X competes with Intel’s i9-7900X (RRP $999), but that’s a 10-core affair which is already over a year old. Clock speed and core count-wise, the i9-7960X is the natural competitor, but that retails for $1,399.
Intel still appears to have the upper hand for singlethreaded workloads, but AMD continue to rule the multicore school, and to do so with an impressive ratio of bangs to bucks.