DDR5 Memory Performance Scaling with Alder Lake Core i9-12900K 58

DDR5 Memory Performance Scaling with Alder Lake Core i9-12900K

(58 Comments) »

Conclusion

I spent almost a week benchmarking—there are over 4000 benchmark scores, but it was worth it. We now have a much better understanding of DDR5 performance scaling, how memory clocks affect applications and games, and what effect timings have.

Let's start with applications first; here, we see huge differences between our various workloads, which is as expected—it has always been like that; some applications are more memory dependent than others, some favor high clocks, some work best with tight memory timings. For example, tasks like rendering and media encoding see very little differences between our various test configurations. The reason is that they work with a relatively small working set in memory, but apply a lot of CPU math to that data. Modern CPUs have various caches that store a copy of highly accessed information inside the CPU, which means it can be accessed directly without a round-trip through the memory subsystem. Tests like Cinebench run almost completely in cache, which reduces the effect of memory speeds even further. As a result, the difference between DDR5-4800 and DDR5-6000 in Cinebench is a meager 137 points or 0.5%—it is clearly a better investment to save on the memory and buy a faster CPU. The opposite is true for applications like WinRAR compression and Adobe Premiere video editing as we see tons of data moving across the memory bus; these tests are highly dependent on memory bandwidth. For example, WinRAR performance drops by 13% when going from DDR5-6000 to DDR5-4800, at a constant CL36. Those 13% are not that far off from the 20% numerical difference between "6000" and "4800."

On average, across our 38-test strong application test suite, we see a relatively constant trend for memory clock. For example, going from DDR5-6000 to DDR5-5200 (800 MT/s lower) will cost you 2.6%. Going to DDR5-4800 instead, 1200 MT/s less, means 3.8% average performance loss. It's important to realize that this is an average depending on your workloads. The differences might be smaller or bigger as outlined in the previous paragraph. As rule of thumb, you can use use "1.5% for each 400 MT/s step."

I always wondered what happens when memory is run at excruciatingly slow speeds, but never found the time to test this for DDR4. That's why I also included performance results for DDR5-2400 to DDR5-4400 even though such speeds are unlikely to ever make it to market—the baseline configuration seems to be DDR5-4800. Especially DDR5-2400 shows shockingly bad performance as 30% performance is lost on average, with much higher losses in some applications that make it fall behind the 4c/4t Core i9-9100F; I had to lol. In games, at lower resolutions, performance is even cut in half (!), which at least means we finally have some entertaining frame-time charts.

I also looked at memory timings, specifically DDR5-4800 CL30 vs. CL36 vs. CL40. These are typical timings available on the market at this early state of DDR5 maturity. In applications, the average difference vs. CL36 is around 2% gained or lost depending on whether you go up to CL30 or down to CL40. I picked CL30 because that's the tightest setting my G.SKILL Trident Z5 DDR5-6000 CL36 kit can handle. Increasing DRAM voltage to 1.45 V made no difference, and adjusting System Agent voltage by +0.4 V didn't help, either. The tightest timings I got were DDR5-4800 30-30-30-70. I also tried running the memory at Gear 1 or 1T—impossible. Overclocking beyond 6000 MT/s was also a bust. 6400 MT/s was always slightly unstable no matter what I did. I suspect it's the memory controller in my Core i9-12900K that can't go that fast.

What's noteworthy is that overall, DDR5-4800 CL30, the tight set of timings, is our second-fastest memory configuration, right behind DDR5-6000 CL36—faster than DDR5-5600 CL36 and everything else. This confirms that just like on DDR4, brute "MHz" aren't the only thing that matters since memory timings are important, too. The same is true for games. CL30 beats all CL36 configurations except for DDR5-6000.

Games have been known to be very sensitive to memory speed and latency for a long time. This remains a constant, even with DDR5. We include gaming performance results at 720p in our reviews even though we know that nobody plays at that resolution. It is still very useful data because the reduced resolution ensures the bottleneck is almost exclusively at the CPU, and the GPU plays only a minor role. This data sets an upper limit on achievable FPS with a GPU even if infinitely fast—the CPU will always be the bottleneck. The difference is roughly 5% between DDR5-6000 and DDR5-4800—not that much. Specific games will see bigger swings; for example, Far Cry 5 loses 12%, while other games barely post a difference: DOOM Eternal sits at -0.3%.

At higher resolutions, the differences will become smaller as higher resolutions run at lower FPS due to the GPU becoming the bottleneck, which reduces the pressure on the memory subsystem. The amount of memory "activity" is roughly constant per frame and fairly independent of the GPU load—the number of pixels per frame only plays a minor role. If we stay with our DDR5-6000 vs. DDR5-4800 example, the differences shrink from 5.1% (720p) to 3.8% (1080p) and 3.2% (1440p), and, finally, 0.5% at 4K. The gap is small enough to be considered insignificant especially at 4K, but it's still there, so maybe don't spend $10K on a gaming PC and pair it with super-slow memory. On the other hand, everyone else who considers price/performance can definitely skip the super-expensive memory and go for something more affordable.

That's also why I included the DDR4-3600 CL16 "sweet spot" configuration in the test results. DDR4 is readily available and has come down in pricing a lot in recent months despite the crisis and shortages. In applications, we see DDR4-3600 CL16 roughly on par with DDR5-4800, which is the lowest speed grade DDR5 announced at this time, so almost nothing is lost despite the numeric difference of "3600" vs. "4800" and huge price increase tied to it. In gaming, which is more sensitive to memory timings than memory bandwidth, our DDR4-3600 configuration works even better, achieving FPS similar to DDR5-5600. The takeaway here is to focus on DDR4 unless you absolutely want to play with the new DDR5 tech. Motherboards are cheaper, too, and you can use the DDR4 memory you already have.
Discuss(58 Comments)
View as single page