So, here we are, back with a look over the array of CPU’s that arrived around xmas 2019 and the start of this year. Rather late to the party this time I fully admit, but it’s been an interesting journey just to get here.
Around the start of last November Intel announced its next iteration of its X299 refresh as being imminent. The timing stood out as potentially taking the wind out of AMD’s own launch of the much anticipated 3950X, although what was notable at the time was a lack of chips from either firm being immediately available. AMD managed to start rolling out the 3950X’s properly a few weeks later and into our hands over the xmas period, whilst Intel also got a few chips out over January although since then however availablity has remained erratic due to at first Chinese new year and the subsequent rise in the Coronavirus impacting the supply chain of hardware coming out of the Asia region.
Since then AMD has gone on to release further new Threadripper solutions over the last few weeks, which are continuing to create a buzz from the market at large. Intel on the other hand continues to be dogged by supply chain issues with the 10900X only appearing whilst writing this up and the 10940X still sitting on the list of chips to acquire and test in the near future.
So, availability aside, why was there no testing done late 2019?
Well, we tried is the answer, but it wasn’t to be. Along with the stock scarcity, I did manage to test most of the chips I could find, only to watch it all fall over when the initial 10980XE sample arrived. This is the first time that I’ve put a 18 core chip through the Reaper based DAWBench VI test and it became clear in testing that once we passed 32 threads in this build, some internal dependencies caused by the way that Kontakt was being mapped were limiting its ability to balance across more than 16 cores/32 threads successfully.
This issue was escalated back to Vin at DAWBench who then spent much of his xmas break rebuilding the testing suite and making it more suitable for high core count CPU’s going forward. After a closed beta throughout a large part of January this year we’ve finally got a release candidate that can be run over the chips as we start again from the ground up.
This means the hybrid test’s that I ran with last time in order to solve the issue of running out of test overhead are retired in favour of two new builds developed by Vin himself. The SGA 1566 test is pretty close to the version I ran last time, with the SGA test remaining largely the same as the older public build but with the audio quality of the plugin being switched to “high performance” in order to cause more load on the CPU.
As noted the new DAWBench VI build has changed to ensure smooth balancing on extreme core counts and is now based around a larger number of single instances, rather than running a multiple instruments per Kontakt instant design, it now runs a one instrument per instance layout with multi-processing disabled within Kontakt to allow the sequencer to spread the load easier. The new test has some interesting restrictions such as pre-loading close to 30GB’s worth of data when all the instances are live, which make it a bit more unwieldy in some situations. However, having tested it on 64 thread successfully it does seem to allow Reaper to manage the load balance to the best of its ability, making me confident that this should be scalable for a quite a while longer going forward.
This isn’t going to be like a normal run down, in that I’m not going in depth with specific chips this time. This is more of a catch up or state of play post to allow us to see the market as it stands. All prior benchmarks are invalid as comparisons as these are brand-new tests to enable us to be able to validate the new hardware.
So, from the top I should mention what I haven’t covered and why as I’m sure a lot of people are asking the relevant questions in relation to those models already.
First off the big omission is the new Threadripper chips and model dependent I saw two different problems that tripped me up here.
In testing on the 64 thread 3970X I saw it refuse to run cleanly on the DB VI test at a 64 buffer, where it simply crackled constantly with little to no load applied. It did run better on a 128 buffer, but the score still placed it behind a number of far weaker chips, so it didn’t look to handle itself well. The 256 buffer and upwards seemed to slowly creep towards the sort of performance levels I would hope to see, pretty much repeating the sort of issue’s we saw in previous generations with the low latency performance hole.
So, we thought there might be a few usable scenarios to be had, but then we ran the DSP test and hit another snag. The projects in that SGA DSP test would always overload at around 60 %- 65% load and I didn’t see a way around this in the time I spent with it. I’ve tried different memory types and speeds, but no matter what it seemed to cap out there.
I also took a look at the mighty 3990X, although was beaten here as well for a slightly different reason. There have been reports in the wider press about software having trouble addressing all the cores currently due to underlying Windows interactions, and we saw pretty much that there. With Reaper, we could only address 64 threads out of the 128 and it behaved much like the 3970X in the testing above. I had wondered if the Windows Pro Workstation build with its support for improved addressing across more cores might help, but the feedback so far is that it’s currently ineffective and comments from Microsoft and AMD would appear to currently bear this out and it remains one to keep an eye out in the future.
I have managed to take a look at a number of the Intel 10 series enthusiast range chips, although what we see here is a platform that is starting to look like it’s approaching the end of its current life cycle. Having attempted to change process a few times now, the strain of extracting ever more performance out of the current platform is beginning to show. This is the first time in a long time, where I’ve had to drop the all core turbo clock lock setting from the Intel enthusiast range as the overhead simply doesn’t seem to be there at this time to allow for it.
I’ve also retested the common Intel mid-range selection, including the ever popular 9900K. These were amongst the last chips that I took a look a proper look at and throughout 2019 remained popular options, still offering a strong solution where the utmost compatibility is concerned.
However, where the excitement stands then is with AMD’s mid-range in the shape of Ryzen and most notably the previously mentioned 3950X launch. One of the last articles I published last year was looking into the differences that overclocking your RAM could make to the system, with improvements to performance becoming ever more obvious as you increased the RAM clock and tightened the timings.
If you look at the officially rated RAM recommendations that accompany the current Ryzen chips they still outline that 3200Mhz kits are recommended and whilst they work fine, they still result in the performance hole that we’ve seen in previous generations of hardware. Even with the 3950X it remains recommended and whilst I’m sure it’s fine for gaming or other less intensive time sensitive workloads, for high performance audio system you want to be cranking those RAM clocks for the best results it seems.
It’s been discussed quite a lot with this generation that the internal data bus is running around 3733MHz or there about. Having tried it with 3733MHz RAM kits, it does indeed work great, but packs of RAM at that speed are both fairly rare and pretty expensive in comparison with more common speeds.
The 3200MHz “AMD Optimized” RAM packs that are in circulation do help, reducing the performance hit from 15 – 20% we’ve seen previously, down to around 8 – 15%. By going up to the 3600MHz AMD optimized kits we can pretty much remove the performance hole that we saw with previous generations.
So the question comes up, is more, simply more? Apparently not, as going over 3733MHz in a lot of testing appears to be having the opposite effect than the one desired, as it switches the controller ratio to another divider setting that allows the performance hole to creep in again, meaning that just adding a kit of 4000MHz isn’t being advised as a solution.
If you don’t have pre-optimized RAM packs already, fear not, with a bit of RAM tweaking you can possibly clock your RAM up to be more efficent. There is an excellent AMD Ram Calculator tool that I was playing with early in testing and you want to be looking at getting better than 90NS for optimizing your RAM for audio. I was hitting 74NS through manual tweaking of a none AMD optimized kit and 80NS through using a kit with a predefined profile and simply selecting that in the BIOS. For the sake of a few hours work overclocking it and testing, I would suggest that anyone who wants the best from their system pays the little extra for the pre-approved packs and have the hassle taken out of it, as regular packs running at stock settings were offering in the region of 110NS and above which isn’t great for our requirements.
So, testing this time is carried out with that in mind and a few additional notes before I lay out the charts.
All tests done using the current Reaper 6.0.4 build.
All testing done with an RME Babyface.
X299 systems are running 2933MHz which is the optimum RAM for the current 10 series.
Z390 systems are running 2666MHz RAM, again the optimum for the platform.
The AMD systems are all testing with 3600MHz RAM packs with pre-optimized timings as discussed above.
Windows 10 running the 1909 build, which is latest at the time of writing.
Regarding all core boosting, as already noted the X299 platform is simply running too hot for this to be viable without some much heavier duty cooling, due in large to the 260w+ power draw that it requires when clocked up. Whilst my standard cooler can handle it, by the time you’ve ramped up the cooling to that level it’s not exactly ideal for an audio recording setup.
AMD has suggested not overclocking the chips themselves this generation, rather letting the system manage it for you. Indeed, so far in practice I’ve seen that any attempt to carry out a manual overclock will then tend to restrict the RAM clocking we can then carry out. Given that optimizing the RAM gives us far more benefit over adding a few hundred MHz to the CPU itself, we’ve chosen to go that route instead. Our Z390 setups by comparison remain active with an all core turbo setup on each chip and officially rated RAM as the more optimum setup.
With that said about the AMD overclocking, I’ve been particularly impressed by AMD’s own turbo handling in this round of testing, with the cores boosting fairly consistent to all core levels when dealing with busy multi-threaded workloads and then pretty much just sitting there, which in most cases with the Ryzen 3000 series appears to be about 4.2Ghz.
With that out of the way, what exactly are we looking at results wise this time?
DAWBench DSP SGA1566 (2020 Build).
DAWBench VI Kontakt Test (2020 Build).
So, as expected with the memory lag cleared up, we see the performance restored at the lowest latency settings and it makes a sizable difference.
The DSP test, which is essentially loading up each core to see where they top out was always the stronger test for AMD and the additional gains put them squarely ahead now at each price point.
Of the chips tested, the 9600K and 9700K sit at the bottom of the results and given the competition that’s not overly surprising. It’s interesting to note that the 3700X and 3800X are fairly similar results wise, with about 300MHz separating the chips on paper, we saw both of these turbo to roughly the same level meaning that there wasn’t a lot between them in real terms. That means that In this instance the 3700X looks to offer the better value out of these two models and indeed from pretty much the whole sub £500 segment.
Moving up to the more expensive “enthusiast” segment, this is where the AMD’s value lead is most apparent at this time. The Intel chips that we’ve seen look have made the standard 10% – 15% generational gains on the previous models, but the launch of the 3950X really pushed the results in AMD’s favour.
Of course, the performance hole we saw previously was always more apparent in the VI test and again the difference here is clear.
The 10980XE holds its crown in the VI test but not by a huge margin, whilst falling behind in the DSP test, meaning that if the chips were more equally priced then this would be far too tough to call. However, even with Intels price adjustments over recent months the 10980XE is sat around the £1250 price point at the time of writting and looking at the chart and specifically that 3950X in comparison which is sat at less than £700, it becomes clear that whilst the 10980XE is a perfectly fine chip in isolation it really should be £300 cheaper than it currently is in order to make sense in a value comparison with the rest of the market.
In fact, that’s the take home right now. Intels prices cuts should have made them more competitive and to be fair they briefly did, but AMD’s continued monitoring of the market and aggressive counter pricing has left Intel with very little to shout about at this time.
For those old enough to remember the days of the P4 chips and FX64, this was the last time that AMD secured a solid tech jump and dominant market share, right now we find ourselves in a period very much reminiscent of that. For Intel to come back now, it’s going to require that long awaited node change just to keep up with AMD and going forward I suspect that ideally they need a new platform in order to reposition themselves fully, kind of what we saw when the “i” series launched and until then I can’t see this highly competitive era easing up.
Looking forward is always interesting, although at this point I’m not sure anyone can predict how the year ahead will pan out. With much of the first and even second quarter of this year likely to see the entire component industry hit with delays in production and shipping, we might well be seeing shortages across the board for the more popular lines for some time to come. There has certainly been short availability of the 10980XE so far and this is unlikely to clear up anytime soon. The knock on effect has seen the 3950X go through waves of short supply too as AMD continue to try to meet the high demand for its flagship Ryzen chip, although it’s an envious problem to have, I’m sure.
It also means that at the time of writing this may prove to be quite possibly the best time to buy in quite a while, if you can source the parts. With possible supply issues in mind, it could be awhile before it all stabilises again and prices may start to drift if those available supplies struggle to meet demand in the short term.
Intel has 10nm mobile chips out there and more on the way, but it’s all a bit subdued on its desktop range. The already known about i9-10900K 10 core successor to the 9900K is the obvious model of interest but looking at the charts, the predicted up to 20% gains to be had via IPC, additional cores and increased cache are attractive, but I’m not sure it’ll elevate it much above the 3900X a chip that’s already out, established and in easy supply.
Also, notably AMD’s 4000 series were announced at this years CES and looks to be arriving later in the year if they remain on schedule. Crucially this is another overhaul for AMD and it seems like they are well-placed to keep making gains from their platform too, which undoubtedly will apply further pressure to Intel.
The earlier concerns about a limited number of incompatibilities have so far not grown and we’re aware that AMD have been working hard to help smooth these out. With the platform not so much “on the map” at this point, as stomping all over it in size 12 boots, we would hope that that developers are now sitting up, paying attention and baking in full compatibility from the ground upwards on future product releases.
We’ve already seen chip prices tumble since the start of this CPU performance war and no doubt it’ll continue for quite possibly a few more years yet.
As we’ve said before the consumer of course remains the big winner in all this. At this point we’re watching the software firms attempting to play optimization catch up as now that more consumers have cheaper and easier access to some absolutely stupendous core counts, it should hopefully lead to more developers taking advantage of this jump in available processing power.
Exciting times and it’s going to be insteresting to see how the rest of the year ahead pans out.