by Rob Williams on November 29, 2019 in Processors
We’ve already put AMD’s newest Ryzen Threadripper chips through our Linux test gauntlet, so it’s now time to shift our focus back to Windows, where we have an even more daunting number of tests to cover. It won’t take long to notice a trend: AMD’s 24-core 3960X and 32-core 3970X are ridiculously fast processors.
On this page, we’re going to be tackling a few additional encoding-type projects. Since the beginning of its life, we’ve benchmarked with Adobe’s Lightroom, but dropped it for about a year or two because it wouldn’t reliably scale. Over time, things changed, and now the application seems pretty efficient on multi-core CPUs.
In addition to Lightroom, we’ve also added Blackmagic RAW Speed Test, which acts as a simple way to see how a CPU can handle playback of BRAW footage at different compression levels. In time, we’ll be adding a much fuller Resolve test to the suite, but this BRAW test fills in for now. Finally, we’re also testing with LameXP, an open-source music encoder that can take advantage of many-core CPUs, as well as the super-popular HandBrake transcoder.
Adobe Lightroom Classic
Months ago, we discovered just how much Adobe Lightroom loves Zen 2, because we saw the 12-core 3900X beat out every other chip in our collection – and by a seriously healthy margin. When the 3950X released, we encountered an oddity that had that chip fall behind the 3900X, something we can’t quite explain. Given that problem, we had worried that these bigger Threadrippers would suffer the same fate, but… not so.
That said, there are still some oddities to speak of. First, the 3970X didn’t manage to take the cake here, across multiple retestings to validate data. The result doesn’t surprise us, given what we saw with the 3950X, but it does highlight that it seems like further optimization in LR could be put to good use.
The main takeaway is that the 3960X becomes the fastest CPU we’ve ever tested in Lightroom. We should also mention that during our testing, we found that Lightroom can prove to be very sensitive to memory latencies and bandwidth. From testing we did eleven years ago, we had a suspicion of that anyway, but at least based on initial testing, we can say that density definitely matters, not just frequency. We don’t want to talk about it too much until we can devote more time to actually testing various configurations and generating some numbers. This is something to look forward to in the future.
What is worth pointing out is the huge difference in performance from one generation to the next with 2990WX to the 3970X, nearly 3.5x the performance for the same 32-cores. Those NUMA node issues of the 2nd gen chips are clearly highlighted here.
Blackmagic RAW Speed Test
The original many-core Threadripper chips (2970WX, 2990WX) didn’t scale too well with Blackmagic’s RAW, likely due to the same memory design that impacted many other video workloads. The new Threadripper has no problem rising to the absolute top, going toe-to-toe with each other. Since both perform so similarly, it could be that we’re running into a bottleneck of some sort – something future testing will include a look at.
HandBrake is an excellent transcode utility for many reasons, but a great one is that it makes for easy testing, and it’s also representative of more users than most of the other tools tested here. At the second-gen Threadripper launch, we encountered an issue with scaling in HandBrake, something that was fixed with the release of 1.2.2.
Fast-forward to today, and we are really excited to see that third-gen Threadripper is serious about scaling. These results are fantastic. Even the $749 3950X looks like a great value from where it sits.
Given the type of workload LameXP is, we weren’t sure if the new Threadripper could really shine, but behold: both chips still scale, but hit a plateau long before the others. Because of how many tracks are being encoded at once, we’re going to experiment with faster storage with our next round of testing and see if it impacts the scaling at all. In the past, I/O was never a bottleneck, but it could become one with such powerful processors at the top.