It’s taken as gospel that Ryzen CPU performance, in particular Zen 2, scales with ram speed. We wanted to investigate this further so that we can find out about the trade off with ram cost, and also ascertain if this is just a theoretical benefit or if it can bring a real world advantage.
For this test we also wanted to use more mainstream hardware. That’s why we stuck to the Ryzen 3600 and a mid range B450 motherboard. We don’t feel that you can accurately call faster more expensive RAM a benefit if you can only reap those benefits on more expensive hardware, increasing the overall system cost.
Why does RAM speed seem to matter with Ryzen?
Ryzen CPU architecture is unique in that it consists of a number of discreet but interconnected chips: The CPU is made up of ‘core complex dies’ or ‘CCD’s which in turn are made up of a pair of quad core ‘CCXs’ or Core complexes. AMD manufactures CPUs by selecting the chips with the right number of viable cores on a core complex, and then fusing off the failed or less worthwhile cores. On CPUs like the six core Ryzen 3600, 2 CCX’s each with 1 core fused off work in parallel on a single CCD. This results in a six core CPU. But these cores have to communicate both with each other and the outside world, and to achieve this there is a memory interconnect known as the ‘infinity fabric’. This is just a fancy name for a data bus, and it’s across this bus that data flows into and out of the CPU.
A diagram showing the layout of a 16 core AMD Ryzen CPU – 2 pairs of CCX’s on a pair of CCD’s communicate via the infinity fabric.
From this we can begin to understand the importance of the speed of the infinity fabric: It dictates the rate of information exchange from the CPU to the outside world and within the CPU itself. On most motherboards the infinity fabric clock is known as the ‘Fclck’ and its speed can be set in BIOS. Further, this infinity fabric runs at a rate of speed that may seem familiar: around 1600Mhz – that is the speed of DDR RAM. DDR means double data rate and DDR RAM can perform memory operations twice per clock cycle, hence the common RAM frequencies, 3000MHz, 3200MHZ being twice the basic RAM Frequency clock.
The key tuning improvement of faster ram is this: Ram speed can be tied to Infinity Fabric speed, and this means that you can move more data through the CPU if you run everything faster. However it’s not simply a case of higher frequency is better: Synchronisation also matters. If infinity fabric runs at a different speed to the ram, then there’s potential for offset between requests for data and the response of the RAM. If it ‘misses’ a clock then the CPU must wait for the next opportunity. This has the effect of increasing overall ram latency. This is why ‘1:1’ infinity fabric ratio is so important – and why ram that can run at higher frequencies allows higher synchronised infinity fabric speeds, and so faster CPU operation.
To test this we set up our rig with a Ryzen 3600 CPU with PBO on, the MSi B450m Mortar motherboard, and some very flexible RAM: The Patriot Viper 2x8GB DDR4 4400Mhz Kit. This memory kit uses Samsungs B-Die ram, now legendary for it’s overclocking ability. To be absolutely clear, no-one should expect this RAM to hit 4400Mhz on the Ryzen platform, at least not without a huge amount of overclocking effort and high end parts complimenting it. Our intention is to use the flexibility of this RAM kit to manually input RAM timings that emulate everything from slow 2133Mhz RAM (or someone who has neglected to set their RAM profile in BIOS) to as fast as we can get the infinity fabric and RAM to run.
We were able to find stable settings from 2133MHZ representing the default speed if no A-XMP profile is set, through to 4000MHz in increments representing the popular RAM speeds. Full timings are shown in the Annex. The infinity Fabric hits a speed limit before our RAM and isn’t stable past 3733MHz, so above that speed we break the 1:1 Fclk ratio to investigate what happens with faster RAM speeds.
Our test methodology is simple: Dial in our representative RAM speeds and use benchmarks in games and synthetic workloads to gauge the impact on performance. To ensure we were fully CPU bound in gaming benchmarks we used an RTX 2080ti or GTX 1080ti and 1080p resolution. To keep the CPU cool we slapped it under a Corsair H100x 240mm AIO.
Artificial workloads are a great starting point for this investigation. They offer fully repeatable tests that can distinguish subtle differences in performance. Some isolate pure CPU performance, others test a wider range of performance factors such as RAM bandwidth and access speeds.
Firestrike is a Direct X11 gaming benchmark tool but we’re only concerning ourselves with the ‘physics’ portion here which is purely dependent on the CPU performance.
In this benchmark we can see the results are fairly uniform, with almost all cases within run-to run variance across the board. The two prominent results are the base JDEC timings at 2133MHz clearly hurting performance, whilst the 3733MHz score is anomalously high and it is repeatable – I can’t offer an explanation for that other than as a quirk of the benchmark and some aspect of it’s demands on the CPU. There’s no clear trend in performance with RAM speed here, so it appears likely that the benchmark tests CPU performance largely independently of its RAM access speeds. In short, RAM overclocking does little to help CPU performance in this test.
Moving on to the more recent DX12 powered Timespy, we see more encouraging results:
Looking at the CPU score in isolation we can see clear evidence of RAM speed scaling. The use of the Direct X12 API and more a more complex test of CPU performance clearly brings the benefits of fast RAM and fast access to that RAM into play, along with the overall boost to CPU performance of the faster Fclk. Of note here is the lower score as we surpass the 1:1 Fclk ratio. Ram speeds at 3866MHz actually hurt performance because we’re no longer synchronising the infinity fabric with the RAM clock as we are able to do at 3733Mhz, and whilst 4000MHz makes up some of the deficit it’s not enough to overturn the damage of desynchronising RAM and Infinity Fabric.
Cinebench R20 is another popular CPU benchmark, using the processor to render a scene tile by tile. It heavily favours multi core CPUs as it’s a parallelisable workload.
We obtain similarly inconclusive results to the Firestrike benchmark in this test. 2133MHz clearly hurts performance, but from 2666MHz all the way through to 3600Mhz there’s no massive performance benefit of faster ram and no clear trend. Again this load on the CPU appears to care little for RAM access speeds although very slow speeds, or breaking the 1:1 infinity Fabric ratio does appear to be detrimental to performance.
Finally, another rendering benchmark: Blender is a popular 3D production tool and it provides a useful standalone Benchmarking tool. We ran the BMW27 and Classroom preset renders using the CPU alone to render them. The completion times in seconds are shown in the chart below.
The BMW27 render appears not to be greatly affected by Ram speed, there’s just a 6 second variance from fastest to slowest and I’m hesitant to call it a trend although each speed step is faster than the last. The Classroom Render is more involved and shows a downwards trend saving 30 seconds from slowest RAM speeds up until we break 1:1 infinity fabric ratio at 3866MHz. 4000Mhz shows a marginal improvement on 3866Mhz but again doesn’t overturn the deficit of the desynchronised FClk.
For the avoidance of doubt here I’ve included the render time when you let a CUDA enabled GPU handle the workload: A GTX 1080ti is four times as fast as the CPU. The $280 RTX 2060 KO handles these renders nearly twice as fast again thanks to the use of the OPTIX renderer: If you’re looking to build a rendering workstation for blender, don’t worry about your ram speed or even too much about the CPU: Invest in the best RTX enabled GPU you can afford.
These results aren’t as clear cut as we might like. The benchmarks most closely resembling gaming workloads appear to show some promise of scaling with RAM speed. However the synthetic or productivity workloads either don’t scale, or don’t offer scaling significant enough to make a persuasive case for high speed RAM on Ryzen Zen 2 CPUs. This tallies with the nature of these workloads: A CPU core isn’t constantly accessing RAM to process renders. It obtains work, processes it, and is then fed more. Whilst there is some benefit to faster RAM and lower latency it makes up such a small proportion of the workload that we can’t see the evidence of RAM speed scaling. However we are able to draw some important takeaways here:
- Very slow RAM or running RAM at the default speed is severely detrimental to performance in almost all cases.
- Breaking the 1:1 Fclk ratio is also harmful to performance, as seen by the performance dip after 3733Mhz.
Gaming is of course far more important than work so we set out to identify games across a spectrum to test the impact of RAM speeds. To obtain these results we had to run the games as close to 100% CPU limited as possible, necessitating the use of a powerful GPU at 1080p. The results are therefore somewhat artificial but we’re aiming to find out the limitations of RAM and CPU performance not GPU performance here. We settled on three AAA titles – Red Dead Redemption, Shadow of the Tomb Raider, and Microsoft Flight 2020. More importantly we also chose two first person shooters with consistent inbuilt benchmarking tools, as it’s competitive players who care the most about every last frame to gain an in-game advantage.
First Person Shooters:
The Division 2 is a competitive shooter that runs at a lower framerate than some better optimised and less graphically intensive games. Running 1080p at low settings on the GTX 1080ti allows us to isolate CPU performance.
The in built benchmark is very consistent but only returns average frame rates. Nevertheless a clear trend is visible with incremental performance increases right up to 3733Mhz, then as we saw in some of the synthetic benchmarks performance takes a hit past 1:1 Flck ratio. Let’s dive into some games with more comprehensive benchmarking.
Rainbow 6 Siege is a highly competitive shooter with high frame rates and well optimised performance.
The inbuilt benchmark allows us to examine Average, Min and Max FPS giving us more insight into performance. We see a distinct trend to higher FPS as we increase RAM speeds towards the maximum 1:1 Fclk speeds. As we exceed that speed all metrics take a hit and again stepping up to 4000MHz cannot recover the lost performance.
Shadow of the Tomb Raider
First we’ll look at Shadow of the Tomb Raider. This well optimised DirectX12 title scales well with hardware improvements but remains playable even on modest hardware and settings. We used our GTX 1080ti at 1080p medium to achieve the following results:
The inbuilt benchmarking is comprehensive and allows us to evaluate relative CPU and GPU performance. Even at these lower settings and with a powerful GPU we are on occasion GPU limited. However framerates are high and smooth and we see clear scaling with RAM speed increases up until the FClk 1:1 ratio is broken. The slowest ram speed is a serious detriment to performance costing nearly 40FPS average. The improvements from 132FPS average to 150+FPS as we go from 3200MHz to 3733MHZ are less marked but still appreciable.
Red Dead Redemption 2
This title has a reputation for punishing hardware requirements and low frame rates – it’s the only title on test that prompted us to break out the RTX 2080ti to ensure we could see through GPU limitations to understand the effect on CPU performance. Whilst we obviously don’t recommend running a top tier GPU at 1080p Balanced settings, doing so again yields the prevailing trend in this test: Ram speed matters. Whilst not as marked as some other titles we see gains with Ram speed, and the slowest RAM is detrimental to performance. The drop in minimum FPS to 69 produced by 2133MHz RAM is likely to have an appreciable impact on playability, and could be one cause of stutters and frame drops as it induces a CPU limitation.
Microsoft Flight 2020
Microsoft Flight 2020 is somewhat unique in that it is as CPU limited as any title we can think of in recent years. Obtaining acceptable and stable frame rates relies on extracting maximum performance from your CPU. In this test we used an A320 and took off from JFK airport, overflying Manhattan before turning left and overflying statue of liberty. This demanding situation is a ‘worst case’ for FS2020 but the Ryzen 3600 copes admirably.
Once again we see the same trend line, albeit muddied by FS2020’s performance challenges. In our in depth examination of Flight 2020 we found that a CPU limit induces stutter and lag when any single core hits maximum utilisation. That appears to be part of the issue here with the minimum FPS scores showing more variance. However we see the same trend of RAM speed scaling, and note that the slowest RAM can cost you a full 10FPS or over 20% performance hit Vs better optimised RAM at 3600MHz. The most prominent feature here is that whilst in many of the other games frame rates appear unrealistic to induce a CPU limitation, in Flight 2020 most users will absolutely see an appreciable benefit: Even a GTX 1660 Super is capable of this level of performance at 1080p high settings, whilst high end GPUs at higher resolutions cannot overcome the CPU limitation imposed by this simulators game engine. If you’re building a PC for Flight 2020 with a Ryzen CPU, pay close attention to RAM Speed to enhance your experience – there’s no cheaper way to buy 10 extra FPS.
Our results highlight the importance of playing to the Ryzen architectures strengths. It’s not really RAM speed that matters here. What is critical is the Infinity Fabric speed and using RAM that can match that speed, synchronising RAM clocks to CPU clocks and minimising RAM latency. There’s little value in pushing past the maximum stable Fclck value of the CPU and in our example that was 3733MHz. It appears that some examples 3600MHz is the ceiling of stability. The market has also had it’s say, and high speed RAM has come down in price as manufacturers cater to demand for faster RAM. Therefore we can say with confidence that 3600MHz CL16 Ram hits the sweet spot of performance, price and stability.
For a gaming build it makes excellent sense to buy 3600MHz RAM, with timings at CL16 or lower. Whilst you may not notice a benefit in all games, it lifts the performance floor of the CPU considerably and will prolong the competitive life of your processor by enhancing it’s performance. You should also ensure that you enable XMP to obtain that benefit – the most egregious performance hit in all our tests was simply failing to set RAM speed in BIOS. You may find that your RAM doesn’t work at the rated speeds right out of the box, but don’t give up or return it as you can still make it work. Using the flexibility of the AMD Motherboard platforms and the tuning utility ‘Ryzen DRAM calculator’ you will be able to generate and input settings that stabilise your RAM. If you want to avoid this headache, use RAM on the QVL list of your motherboard for guaranteed compatibility. The motherboard manufacturer will have ensured that the BIOS holds optimal settings for the RAM so it’s guaranteed to work at advertised speeds.
For workstations the situation is less clear cut and varies more widely on application. In many cases it will be quantity of RAM that’s critical rather than speed and high capacity RAM kits can get incredibly expensive. We’d urge you to consider the nature of the work, with video editing and 3D rendering unlikely to be critically impacted by less fast RAM. First ensure you have the adequate quantity, then see if there are kits available at optimal speeds and a reasonable price.
- Fast RAM matters on Ryzen because it synchronises the infinity fabric interconnect to RAM speeds and lowers latency.
- 3600MHz at CL16 is the current sweet spot of compatibility, affordability and speed on ZEN2 CPUs.
- Ensure you set RAM speed correctly.
- Unless you have RAM at 2400MHz or lower or run particularly sensitive games, a Ram speed upgrade is unlikely to bring appreciable performance benefits in most situations. It pauys to buy it right first time.
Recommended RAM for Ryzen Builds – Our Recommendations
RAM Kit used in this test: Patriot Viper 4400Mhz CL19 Samsung B-Die PVS416G440C9K ($129)
This RAM kit uses blazing fast Samsung B-Die IC’s for maximum speeds and overclocking potential. In all cases it was our CPU that became the limiting factor but we were still able to run 4000MHz comfortably on a Ryzen 3600 and mid range B450 Motherboard at stock voltages – and that’s impressive. The flexibility of this RAM can’t be overstated and using it allowed us to effectively emulate 8 different popular RAM specifications with just one kit. If you want to try your hand at overclocking particularly on Intel, or simply want to explore the extent of RAM overclocking yourself this is an excellent purchase. There’s rumour that ZEN 3 CPU’s may scale even further with infinity fabric speeds perhaps stable out past 2000MHz, in which case RAM like this will be in short supply as tuning enthusiasts snap it up to wring as much as possible out of those upcoming CPUs.
Best Gaming RAM Kit for Ryzen: Crucial Ballistix 2X8GB 3600Mhz CL16 BL2K8G36C16U4B ($79.99)
You don’t have to spend much to obtain premium RAM that is optimized for Ryzen. This RAM kit is just $80 and brings all the benefits of fast ram ensuring maximum gaming performance from your CPU. The tight timings. There’s also an RGB Version for just $10 more. (BL2K8G36C16U4BL)
Best Workstation RAM Kit for Ryzen: G.Skill Ripjaws V 64GB CL18 (2x32GB) F4-3600C18D-64GVK ($240)
This RAM Kit offers a fantastic balance of capacity, speed and affordability. A 2 Stick Kit allows 2 slots free for future expansion to 128GB if required, whilst G.Skill offer comprehensive reverse QVL’s on their website allowing you to easily check compatibility with your motherboard and CPU choices. Whilst speed isn’t as critical for many workstation and productivity tasks you can be sure your high end Ryzen CPU is running to full potential with this RAM kit.
The PremiumBuilds test rig warms up for another tough day in the lab…..
Appendix: Ram Settings. Sub timings also tweaked for stability, guidance via Ryzen DRAM calculator 1.7.3
|Speed||Timings CL-tRCD-tRP-tRAS||FClk||Voltage||CPU PBO|
|4000MHz||16-16-16-34||1833MHz||Auto (1.4 Max)||ON|
|3866MHz||16-16-16-34||1833Mhz||Auto (1.4 Max)||ON|
|3733MHz||15-15-15-30||1867Mhz||Auto (1.4 Max)||ON|
|3600MHz||16-16-16-34||1800MHz||Auto (1.4 Max)||ON|
|3200MHz||16-17-16-34||1600MHz||Auto (1.4 Max)||ON|
|3000MHz||16-16-17-34||1500MHz||Auto (1.4 Max)||ON|
|2666MHz||16-16-16-32||Auto (default)||Auto (1.4 Max)||ON|
|2133MHz||15-15-15-36||Auto (default)||Auto (1.4 Max)||ON|