Threadripper 7000 Series – Tested for the studio

Following on from Intel's return to the HEDT enthusiast and workstation segment with their recent Xeon releases, AMD brings back its own high-end offering in the guise of the Threadripper range complete with all modern features. A server grade platform where the CPU's on offer in the regular (none Pro) Threadripper range cover 24, 32 and 64 core configurations with us taking a look at the 7960x 24 core and 7970X 32 core models here today.

Part of the appeal of the hardware found on HEDT platforms outside of their high core counts, is often their usage of ECC memory which tends to allow for greater quantities of RAM than consumer platforms. Even with a board like the test setup's ASUS Pro WS TRX 50 SAGE and 4 available slots, it's possible to max this out at 256GB. If you combine it with the platforms large PCIe lane count, allowing for sizable storage expansion both on-board and via add-in cards, on paper it sounds like it should have the potential to be an interesting setup for anyone working with large audio libraries.

Our delayed coverage in fact looks to be memory related and comes down to our initial round of testing, where troublesome low latency results occurred in the opening round of benchmarking whilst using the DAWbench VI test and the lowest ASIO 64 buffer settings proved completely unusable. Having seen this kind of behaviour in other generations and found that it would often improve with faster RAM it made sense to look into better performing kits as a possible solution before drawing any conclusion.

At launch the validated RAM kits were largely limited to 4800MHz but we've since seen faster kits slowly being qualified and becoming more widely available as times gone by. The result of this additional testing was both illuminating and a touch frustrating in equal fashion, as there's some astounding performance being held back by in low latency situations.

Whilst 6000MHz kits and faster are starting to become more commonly available, the fastest kit at time of writing which appeared on the QVL sheet and proceeded to run well for us were rated to 56000MHz. This should be enough to outline any inefficacies and establish if the 64 buffer results were going to be more universal than our initial 4800MHz based testing outlined. First up we’ll take a look at the differences we measured when stepping up the memory clocks from 4800MHz to 5600MHz.

Taking a look over the VI test first as this is where we have historically seen low latency performance suffer on some of the earlier generations, just as we do here. Most notably having seen a major impact on the older Ryzen Gen 1 and Gen 2 hardware, where it wasn't really until Ryzen Gen 3 along with improvements in DDR4 performance that the consumer offering really found its feet and started to achieve good low latency performance. It was a similar case with older Threadripper 3000 and earlier models where low latency performance was severely limited, although in later retesting the memory speeds and Windows support both improved significantly and to a point where it was reaching a state of usability just as AMD discontinued the range.

So, with this in mind when the VI testing started to show up a pattern of a none functional ASIO 64 buffer, it made sense to hold off until the RAM availability had improved. Ultimately it didn’t resolve the problematic 64 buffer setting result, but It did end up making a sizable difference not only in the memory heavy VI test, but it also had a smaller yet still notable effect to the DSP test performance. This is interesting as we don't normally see a marked difference when adjusting RAM speed on the consumer platform in the DSP test. It's clear that we're seeing performance scale here with memory and once more there looks to be good cause to try and match the internal clocking with your RAM selection. But the failing result on the 64 buffer showed no difference in how it presented itself even with the memory upgrade, with audio continuing to corrupt on the base test with no additionally load processing being added.

We've seen a history of troublesome low latency performance with server hardware where these chips often feature multiple processing clusters, in AMD's case here the CCD's (or “Core Chiplet Die's”) where they essentially collect together multiple smaller processing blocks to build the whole chip. A problem with constructing hardware in this fashion is inter-block latency can come into effect when data is being exchanged around the chip. In the consumer space the 7950X leads the range along with the 7900X with just two CCD's and the other models below those have one 1 block. These more modest arrangements minimise the amount of internal transfer latency and their results in prior testing speak for themselves when functioning under ASIO based software. By contrast the Theadripper chips we have here are based around 4 CCD arrangements which judging by the testing results look to be manifesting some degree of internal latency when dealing with the RAM intensive audio libraries.

Taking those results with the 5600MHz RAM kits and adding them into the overall chart, we've arranged the results via the 128 buffer to allow for the lowest set of viable results. Looking at the curve and then comparing it to the larger buffer settings, it helps to highlight that even the 128 buffer itself seems to be falling short of it's potential with both models falling somewhere between Intel’s older X299 based offerings or its newer consumer offerings along with AMD's own 7950X. It's only on the 256 buffer and above where it truly opens up and reveals the potential this platform really has.

In sharp contrast on the DAWBench DSP test both chips excel at even the lowest buffer setting and continue to scale well all the way up, with the 7970X notably running away from the pack. There is a lot of performance overhead overall that’s only really let down by the latency performance with memory sensitive applications and plugins. Whilst this might be problematic for recording situations there’s a possibility that for users running at a more relaxed 256 and above say for post-production or mastering scenarios where it could possible fit within the workflow.

Money no object and these limitations aside, these are very powerful chips but with perhaps an 80% increase over AMD’s own consumer flagship 7950X in DSP terms and 30 – 40% on the VI test, the cost for a 7970X setup likely to exceed three times the cost of a similar 7950X system for less than double the performance. In audio terms it’s an already niche offering that through its mixed performance levels, further limits the users it will appeal to. With the recent announcement from AMD that their Ryzen 9000 series is also just around the corner that performance gap is surely due to narrow even further as we see the value proposition further. Ultimately, this first DDR5 based Threadripper platform shows great potential and offers a huge amount of performance in the right scenarios, unfortunately low latency audio performance just isn’t one of them. We do however look forward to seeing what the second and third iterations bring in due course, as the performance continues to improve and as the platform further matures and it looks as AMD has a solid platform to build upon longer term.

SCAN 3XS CUSTOM AUDIO SYSTEMS >

SCAN FIXED SERIES AUDIO SYSTEMS >