AMD's third-generation Ryzen chips have reignited the CPU arms race with Intel — but which is best for a music-production PC?
A little over a decade ago, music-makers were reaping the benefits of the CPU arms race between AMD and Intel — the rate of increase in processor speeds seemed astonishing. Intel seemed to win that race, and for a long time their 'Core i' series has dominated this market sector, while AMD have focused their efforts elsewhere. We've seen worthwhile gains in other areas, such as core counts, more stable overclocking, and energy efficiency (lower heat means less noise from cooling systems), but in terms of sheer processing power, the lack of competition seemed to slow the rate of advance.
Now, with the third generation of AMD's Ryzen processors, that arms race looks set to resume — these new chips really do seem to be giving Intel a run for their money. I can say this with confidence because, along with colleagues at Scan UK, I recently ran benchmark tests on a number of AMD and Intel chips, to see how they compare for audio-production applications. As you'll see, both companies are now making very capable CPUs, but the answer to the question "which is best for a music PC?" isn't entirely straightforward
Before I dive into the results in detail, I'd better bring you up to date on the last few generations of AMD and Intel CPUs. Launched in 2017, AMD's Ryzen range offered more cores and improved handling of larger, parallel workloads than what had gone before. But while they were forward-thinking on paper, the chips fell short of the market-leading Intel CPUs when it came to audio production. Notably, Intel hardware still had the edge when dealing with low-latency ASIO-based workloads: there are various real-world scenarios in which the single-core performance is really important, and the Ryzens' per-core performance was lower. During tests, we also observed a tendency in the early Ryzens to overload early with ASIO buffers under 192 samples; they'd often distort audio with only around 80 percent of the CPU capacity being used.
The second-generation Ryzens offered better single-core performance, but they still often underperformed both in mainstream applications and in our more rigorous bench tests. Over the same period, Intel continued to release iterations of their own 'Core i' platform, and although they didn't make huge leaps forward, they succeeded in keeping well ahead when it came to handling audio workloads. It's only over the past 12 months that we've finally seen the Ryzen platform come into its own, in the form of the 3000 series.
So, what changed? The closing of the gap is partly due to Intel's development woes in recent years. A common technique for improving a processor's performance is to carry out a 'die shrink' between generations, downscaling the processor components so that more can be packed into the same space. Intel have missed a number of their own shrinking deadlines in recent years, whereas AMD have succeeded.
More important, though, has been the handling of memory. Historically, matching the memory to the rated memory controller speed has often proven best for audio handling, with optimal stability being a higher priority than the often marginal performance gains to had by overclocking the RAM. But the latest Ryzen platform changes this. AMD's own specification recommends 3200MHz RAM, but we've found the real sweet spot for audio handling to be higher.
AMD's chip design is based around 'Zen' microarchitecture, with the cores interconnected by an internal data bus known as the Infinity Fabric. We've found the optimum setting for the AMD chips to talk to the GMI (global memory interconnect) to be around 3733MHz. Sadly, for previous Ryzen generations this was impractical and the figure would often be tweaked downwards, to optimise interaction with slower memory.
But now, we're finally starting to see memory kits on the market that can easily match the internal bus speed. This means we can finally make use of all of the CPU's internal bandwidth and, in turn, look to solve the low-latency ASIO buffer issue experienced in earlier-generation Ryzens. Although 3733MHz RAM kits remain relatively rare (and expensive), it's not a strict requirement for the memory to meet that clock speed exactly; hidden timings under the hood can make up the difference. Many RAM firms offer Ryzen-optimised packs at the much more common speed of 3600MHz, and these can still unlock the full potential performance of the Ryzen CPUs for the audio user. What's more, the increasing popularity of faster memory kits amongst the enthusiast market means that prices have tumbled, making the extra performance more accessible. It's also worth mentioning that the older Ryzen chips were capable of handling around 3600MHz too, so early Ryzen adopters who are still running older, slower RAM kits could potentially achieve performance gains by upgrading their RAM.
It should be noted, though, that memory overclocking past the Infinity Fabric clock rate can lead an offset ratio being applied and this may make the performance worse, as more latency is added to the memory path. This brings us back to Intel and the question of RAM overclocking: it may previously have seemed of questionable value, but as those once marginal gains become more accessible, it's something we'll need to revisit in tests soon.
Also of interest to music makers is that, as CPU core counts have climbed to previously unseen levels in the consumer ranges, both Intel and AMD have increased the amount of memory that's supported. Currently, it's possible to fit up to 128GB of RAM on both Intel's Z490 and AMD's X570 platforms, and Intel's X299 range can handle up to a dizzying 256GB. So users working with large project templates that are laden with sample-based instruments now have various options to choose from.
With all that in mind, let's consider what headlines we draw from our latest round of CPU tests. We measured how well the various CPUs on test handled both plug‑in processors/effects (digital signal processing) and sample-based virtual instruments. To do this, we used an updated DAWBench VI test to take full advantage of the higher core-count of the flagship new chips.
See detailed larger views of the graphs in this Zip file: core-wars-graphs.zip
For the DSP test project, we retired the older multiband compressor plug‑in in favour of Shattered Glass Audio's more CPU-demanding SGA1566 mic preamp emulation. For the virtual instrument test, we used the latest version of Native Instruments' Kontakt 5 sampler in a small project using the older Kontakt 4 factory library. We hosted the plug‑ins in Reaper, as we know that DAW is great at multicore handling. Testing was carried using an RME Babyface USB audio interface, and involved loading instances of each plug‑in up to the point of CPU overload, or where we began to hear crackling during playback.
When stacking up instances of the tube mic-preamp emulation, we increased the load upon the CPU one plug‑in at a time. At the lowest buffer settings, the lower-end AMD chips — the 3700X and 3800X — performed slightly less well than we'd expected compared with the equivalent Intels. But with a 128-sample (or higher) buffer that situation changed. Those two AMD chips are particularly interesting, because they appear to be closely matched: the 3700X and its successor the 3800X are both commonly available and have the same number of cores, but the latter is clocked higher by default. In testing, the system pushed both those chips to a similar level under the turbo settings, which suggests the slightly cheaper 3700X may offer better value.
When it came to the high-end chips and the best of Intel's current mid-range offerings (the Intel 10900K), these ranked in the sequence we'd expected. As you'll see, the top-end Ryzen 12 and 16 core chips score well here, as would be expected from a test dealing with raw CPU performance (both chips perform strongly in more general tests like CPUbenchmark or Geekbench) and the R9 3950X is a clear winner.
The virtual instrument test places different demands on a processor, and the results are very different from those of the signal-processing test. To allow support for the higher core-count chips, the old test was reworked to use a single Kontakt instrument per track. In our previous tests, Kontakt ran with 16 instruments and used multiple output channels; as Kontakt can only address 16 cores per instance, we'd sometimes experience uneven load balancing when working with chips of 18 cores or more. The new approach avoids this. With so many instances running, it also places higher demands on system memory — to around 30GB once the template is loaded — so this test can also help us highlight differences in memory handling.
This time, AMD put in the strongest performance at the lowest buffer settings (which is useful when you play parts in, rather than programming them), but the Intels improved considerably with increased buffer sizes. The AMD chips overall exhibit a more muted performance with these workloads, but the 3950X still remains better value for money than other chips of similar performance.
In the process of this testing, we also evaluated AMD's 3970X (32 cores, 64 threads) and 3990X (64 cores, 128 threads) 'Threadripper' chips. When launched, these caused quite a stir, and hopes were initially high. However, we ran into similar low-latency performance issues as we'd observed with the earlier generations of Ryzens. In our sampled-instrument tests, the 3970X struggled at low buffer settings, and reminded us of the performance overhead we tend to associate with more traditional multi-CPU arrangements. It left us wondering if the design choices in facilitating so many cores may have led to fresh latency problems. These issues were most extreme with a buffer setting of 64: the CPU was constantly overloading, and noticeably underperformed until we got beyond a 256-sample buffer. On firing up the DSP tests, we saw additional symptoms, with the CPU approaching overload at around 65 percent of the supposed capacity when working with the smaller buffer settings. There's still a lot of raw performance there, despite the overhead, so don't write it off as a chip if your main application is mixing, where higher buffer settings are typically less problematic. But it proved difficult to make a viable comparison during tests focused on low-latency performance — hence it not featuring on the charts.
Having already seen the smaller chip flounder, we were unsurprised to see the beefier 3990X suffer similarly. While the software houses have done well to keep up with the current hardware explosion, the changes are not always within their control — in this case Reaper should, theoretically, have been able to utilise those chips fully.
Looking through the test results, it's clear that, despite the Threadripper limitations, AMD have found themselves back in a very strong position. So are there any other negatives worth noting? When it comes to music studio use, there are possibly two...
In the past, we've sometimes seen older USB hardware act a little fussily with non-Intel controllers, particularly on newer USB 3 setups. With the majority of PCs (including Macs) having had Intel chipsets for well over a decade, it's not surprising; its easy to see how they'd become the testing default. Also, the AMD X570 range has moved to PCIe 4.0, which affects some older cards and is causing some users teething problems. AMD have been working with various manufacturers to resolve such issues. Some BIOS updates have targeted audio hardware specifically, which is encouraging, and for other issues there are often documented workarounds. Again, while there are some risks at this stage, a little research upfront can result in a smoother ride.
It's clear that, despite the Threadripper limitations, AMD have found themselves back in a very strong position.
The second 'negative' relates to the AMD mainboards' cooling arrangements. There are two flavours: the more basic X470, and the premium X570 range. The latter offers the PCIe 4.0 support and additional bandwidth for more flexible USB and storage options. Crucially, the X570 range has seen mainboard manufacturers offering improved VRMs (voltage regulator modules), allowing better power delivery to the demanding high core-count chips, as well as continuing to improve memory support, making it the more attractive platform choice for running the power-hungry 3900X or 3950X chips.
Found on the vast majority of X570 motherboards and all of the more reasonably priced models is a small chipset cooling fan. This can end up running at audible levels in environments with a low ambient noise floor, such as studios. Thankfully, a lot of the boards we've looked at ship with fan-management software, which can bring the noise down to far more acceptable levels; a little tweaking here can go a long way.
So, that's the challenger, but what of the defending champ? Well, the maturity of the Intel platform is starting to show. It remains a well-tested platform for those who don't wish to tinker, but the performance headroom that enthusiasts have largely taken for granted in their recent CPU ranges is notably reduced. The current chips run well at stock settings with the turbo engaged, but efforts to overclock the higher core-count chips increase power requirements considerably, in turn raising the cooling requirements and possibly making squeezing these chips for extra performance a little impractical for users who wish to keep the system running quietly.
The AMD chips prove similarly challenging to overclock, though — our tests showed the performance gains to be equally achievable simply through the memory upgrade discussed above. Given the number of cores now being squeezed into each of these modern chips, and the pressure that competition between the two companies places to add more and more cores with each new model, the heat generated is going to become more challenging to mitigate while keeping noise low.
The bottom line is that the consumer is back in a strong position. AMD's aggressive pricing has swallowed up market share at pace, and solid competition looks set to drive advances in the computer world once again. Intel, of course, have been forced to react by price-cutting as they try to reposition their respective models against AMD's often more value-oriented range. With AMD's Ryzen 4000 series also on the horizon, we should see another round of generational improvements soon. And then, of course, there's Apple, who've recently announced that they're migrating away from the Intel X86 platform, and will focus instead on development of ARM-based hardware. While there's little point gazing into crystal balls at this stage, it seems fair to suggest that the introduction of another distinct platform in the desktop market could well shake things up further. All of this will inevitably serve to sharpen Intel's focus, and they'll undoubtedly already have something waiting in the wings. As we hurtle into the latter half of the year, the only thing that seems certain for now is that developments in the near future will be interesting!
The faster M.2.type of solid-state drive has risen in popularity in recent years — there's now support for two or three of this kind of high-speed storage drive on most mid- to high-end mainboards. Operating at up to 3500MB/s read speeds (around six times quicker than a standard SSD), these are obviously great OS drives, but with prices of larger-capacity M.2. drives continuing to fall, they're fast becoming a serious option for anyone looking to reduce large project and library loading times.
We're already seeing the emergence of the next generation, with AMD's X570 mainboards offering PCIe 4.0 support for an emerging range of drives capable of up to 5000MB/s. We really have come a long way from the 150MB/s we could expect of the average mechanical hard drive!
Of course, this means conventional SSDs have also dropped in price. As their capacities have also increased, the key differentiator now tends to be how much space you get for your money. Anyone still hosting their libraries on a humbler hard drive could greatly improve the quality of their daily experience by hosting all of their libraries on a newer, faster drive. After all, less time loading means more time making music!
Thunderbolt support remains a popular subject with my customers. The standard has long been a staple on Mac, but implementation on Windows PC hardware has been somewhat patchy. Thankfully, we're now seeing more mainboards with Thunderbolt support, both in BIOS and hardware form, and a few firms offer premium models with Thunderbolt already on board.
The lack of Thunderbolt support was, for some users, a weakness of the earlier Ryzens and a reason to choose Intel. But as a consequence of Intel opening up the standard in recent years, we're starting to see some AMD brands offering official Thunderbolt support. So both CPU platforms now allow you to take full advantage of the superb Thunderbolt hardware that's currently available.
The tests confirm that the raw power is more easily available than ever before, from both AMD and Intel. But can we actually make use of it?
The answer in most cases is yes. As core counts have increased, the majority of DAW software manufacturers have kept pace: the latest versions of most leading DAWs now offer full 32-core support, and some can handle more than that. Note, though, that each of these CPUs implements a version of multi-threading, which effectively doubles the core count from the DAW's point of view.
If you are planning to base a new system around a high-core-count chip in the future, it's always worth checking the supported recommendations for your software to help guide your choice. In particular, check if the current version of your DAW supports all those cores before spending silly money. Arguably the best way to do this is to consult the manufacturer's FAQs page and support forums, and try to establish what setups other users have already got working smoothly. The goalposts are shifting all the time, so a little research here can save you a lot of time and effort on any new build.