FIXING MY MISTAKES…. Ryzen RAM Testing REVISITED
A few weeks ago I posted a video titled “I CAN’T EXPLAIN THIS…. Ryzen RAM Testing”, which was me testing out various RAM speeds using this Ryzen 5600X and RTX 3080 PC. I was using the same kit of memory to try and eliminate variables, just using the BIOS to change the frequency but leaving everything else including the timings set by the XMP Profile. The gaming results all made sense, generally speaking the faster memory you have, the more performance you got – except for some games that didn’t seem to care much. But the productivity tests.. I couldn’t make sense of the results. The slowest two speeds, 2133 and 2666 were often significantly faster than 3600 which was the fastest speed I was testing.
In the video I asked if anyone had any suggestions, and much to your credit I was inundated with excellent suggestions and theories. So in this video I want to go through your suggestions and test them out to work out what is going on, what to do to fix it, and incidentally also see how Precision Boost Overdrive or PBO for short handles on this 5600X.
The first suggestion was to check the timings, specifically the tertiary timings which as you expect change based on the training the motherboard does on startup. I found that even the primary timings slipped when at 3600, or specifically 1800MHz since it’s Double Data Rate memory. It’s not massive, but goes from 17, 19, 19, 39, 58, to 18, 19, 19, 39, 84. It’s not a massive change, but could have a slight impact in performance although wouldn’t explain such a big swing having the slower speeds destroy the faster ones.
So, onto your second suggestion, one from the good Doctor himself. Power consumption. Specifically, a unique quirk of Ryzen and it’s chiplet design. This 5600X CPU actually has two separate “chips” under this metal heat spreader, or IHS. One is the chiplet with the cores and cache, and the other is the I/O die which houses the memory controller amongst many others. The motherboard enforces power limits on the CPU, which for a 65W TDP CPU like this is 88W of total socket power. That 88W then needs to be split up between the cores and the I/O die, and the more power you can send to the cores the faster they can boost meaning more performance.
As you suspected, when running at 2133 the I/O was barely sipping power at just 5W, but at 3600… 17W. The cores? At 2133 they had up to 56W of juice flowing through them, but at 3600 that was down to just 45W. That’s a significant decline, and very much explains the performance delta I saw.
So, by default Ryzen’s power limits mean for certain chips you’ll struggle to see a benefit of higher clocked memory in CPU heavy workloads – although if you just game there is still an advantage to getting faster kits. A chip like the 5800X might fare a little better as that’s in the next class up of TDP meaning it has up to 142W of total socket power available and only 8 cores to power with that, rather than the 5900X and 5950X with 12 and 16 respectively.
But, if you own a 5600X and want to make use of your 3600 kit, what do you do? You head to your BIOS and enable PBO. Specifically, you set increased PPT, TDC and EDC limits. I set 125W PPT, 150A TDC and 200A EDC, which resulted in the total package power jumping from 76W to 105W. It also meant there was more than enough power available for both the cores and the I/O die.
Performance saw a marked improvement, in fact a full 10% in Cinebench R20 nT, going from just shy of 4000 points to 4410 points. The same goes for the Gooseberry scene where we drop a full 100 seconds from the render time. Now you might be thinking, Andrew you’ve just overclocked it obviously it’s going to be faster! Well, to fully isolate how much of that performance gain comes from the RAM speed I also re-ran the tests with the power limits increased but running at 2133. In Cinebench the difference is negligible between them. Again, that might come down to RAM utilization being rather minimal in Cinebench, but luckily that’s not a problem with the Gooseberry Scene, which shows the PBO enabled 2133 running 50 seconds slower than the same PBO settings at 3600.
So, in the right scenario, 3600 offers around 5% more performance than 2133. That’s not massive, but it does help extend the advantage you get by increasing your power limits. What about gaming?
Well, in Watchdogs Legion the advantage still lies with the faster memory, although you do gain 3FPS average by increasing the power limits too. In my testing Watchdogs is a pretty CPU limited game so any increase in horsepower translates well into performance.
In Cyberpunk it’s within margin of error between the runs, although enabling PBO at 2133 did see a slight improvement of 2FPS both on average and the 1% lows, plus there is still a 20 FPS gap between running 2133 and 3600.
In CSGO it’s back to whacky races style performance figures, all over the place. 2133 saw a decent 23FPS advantage by enabling PBO, whereas 3600 saw a 10FPS loss. I’d argue percentage wise it doesn’t matter much here so I won’t dwell on it.
In Fortnite, while the difference between 2133 and 3600 is still there, there isn’t any benefit running with PBO enabled here.
And in Microsoft Flight, surprisingly, there is no difference enabling PBO on either run. Like, at all. They may as well have been the same runs. There’s still a decent difference going from 2133 to 3600 though.
So, how do we wrap this all up.. Well, we found the culprit of the strange results, Ryzen’s power limits. For the average gamer who builds a system, it’s still a good idea to get the faster memory speeds. In gaming no matter what your setting is, at worst faster RAM does nothing, but at best gives you a significant increase in performance. If you want to do heavy CPU workloads like rendering, either increase the power limits, or don’t sweat getting slower speeds especially if you are getting larger capacities. I would quite like to test the 5800X and see if it’s position as the lowest end 125W TDP chip means it can run faster memory without needing to unlock power limits, so if you’d like to see that do let me know in the comments.
I also want to thank you guys for your suggestions, you were spot on. I’m glad we got to the bottom of it!
