- Aug 7, 2009
- 302
- 168
- 116
With the release of 3DMark's new "CPU Profile" benchmark, I decided to revisit an issue I've had trouble wrapping my head around and overcoming and would appreciate any input or experiences other Ryzen 2700X users have regarding overclocking while running fast memory. All key system specs are in my signature. TLDR at the end.
Essentially, when I am running my memory at 3533MT with all timings tuned (see Ryzen Timing Checker screenshot below) the CPU power consumption absolutely sky rockets (see OCCT Power draw in spreadsheet below). Some of you might ask whether most/all of the increased power draw is SoC or core and I confirmed, but forgot to write down, that 80-90% of the increase in power draw is from the CPU cores.
My core clocks speeds while gaming are 4000-4050Mhz and light single core loads will see clock speeds around 4100-4150Mhz. During OCCT stress tests (AVX, medium data set, 16 threads) the clock speeds are 3800-3850Mhz. While running my memory at these speeds and timings, I never see 4200-4300Mhz boost clocks.
All attempts to manually OC to 4150Mhz, for example, while running these fast timings results in instability during stress tests at reasonable daily voltages. At unreasonable, but not super unsafe voltages, it results in very high temps, then instability once it reaches 90*C+ (1-2 mins after the stress test begins). Even a fixed clock speed of 4100Mhz is unstable.
Having said all that, we all know that Ryzen loves fast memory. And I've gotten significant gains, as shown in the spreadsheet below, just from the memory OC.
Below you will see a spreadsheet of the benchmarks I ran. The column key is as follows:
3533 14fs PBO PE2 = 3533MT CL14 + tuned primary, secondary and tertiary timings. PBO enabled. Asus "Performance Enhancement Level 2". (this is my current config)
3533 14fs defaults = 3533MT CL14 + tuned primary, secondary and tertiary timings. CPU settings at defaults.
3533 14 auto PBO PE2 = 3533MT CL14 + tuned primary timings. Secondary and tertiary timings set to auto. PBO enabled. Asus "Performance Enhancement Level 2".
2400 auto PBO PE2 = 2400MT, all timings auto. PBO enabled. Asus "Performance Enhancement Level 2".
Optimized defaults 2400 = All settings default/auto (defaults to 2400MT).
I used Time Spy to simulate a very GPU limited situation. While the CPU scores changed quite a bit with the faster memory configuration, the GPU scores are within the run-to-run margin of error.
I used Night Raid to simulate a CPU limited situation. Both the CPU and the graphics scores changed significantly with the faster memory configurations.
I also ran the API benchmark. The DX11 single threaded benchmark showed a nice boost in performance with faster memory configurations. But the DX11 multi-threaded benchmark showed an even larger increase in performance. DX12, which is inherently multi-threaded, also showed a significant performance uplift. And Vulkan, also inherently multi-threaded, more than doubled its score. Vulkan also appears to be particularly sensitive to memory latency. Going from 3533MT CL14 + auto secondary/tertiary timings to 3533MT CL14 w/tuned timings showed quite an uplift in performance.
Rainbow Six Siege is the most demanding game I play right now. It is also unique in that it has both a DX11 client and a Vulkan client, allowing me to do as close of an apples-to-apples comparison between the two APIs as you can get. I used the in-game benchmark and was on the low preset at 1080p to ensure a CPU limited situation. Both clients showed ~25% gains, though the Vulkan client started and ended with a higher overall frame rate.
The OCCT power draw row speaks for itself. Even without PBO, running at 3533 with fast + tuned timings results in a ~50% increase in power draw over the rated 105W TDP.
****TLDR**** - So, in summary, while I am way ahead of "stock" performance just from the memory OC, I want to see if I can squeeze more with a fixed all core OC. Which the memory OC makes difficult. Even a slight PBO undervolt results in instability (-0.05v offset).
Essentially, when I am running my memory at 3533MT with all timings tuned (see Ryzen Timing Checker screenshot below) the CPU power consumption absolutely sky rockets (see OCCT Power draw in spreadsheet below). Some of you might ask whether most/all of the increased power draw is SoC or core and I confirmed, but forgot to write down, that 80-90% of the increase in power draw is from the CPU cores.
My core clocks speeds while gaming are 4000-4050Mhz and light single core loads will see clock speeds around 4100-4150Mhz. During OCCT stress tests (AVX, medium data set, 16 threads) the clock speeds are 3800-3850Mhz. While running my memory at these speeds and timings, I never see 4200-4300Mhz boost clocks.
All attempts to manually OC to 4150Mhz, for example, while running these fast timings results in instability during stress tests at reasonable daily voltages. At unreasonable, but not super unsafe voltages, it results in very high temps, then instability once it reaches 90*C+ (1-2 mins after the stress test begins). Even a fixed clock speed of 4100Mhz is unstable.
Having said all that, we all know that Ryzen loves fast memory. And I've gotten significant gains, as shown in the spreadsheet below, just from the memory OC.
Below you will see a spreadsheet of the benchmarks I ran. The column key is as follows:
3533 14fs PBO PE2 = 3533MT CL14 + tuned primary, secondary and tertiary timings. PBO enabled. Asus "Performance Enhancement Level 2". (this is my current config)
3533 14fs defaults = 3533MT CL14 + tuned primary, secondary and tertiary timings. CPU settings at defaults.
3533 14 auto PBO PE2 = 3533MT CL14 + tuned primary timings. Secondary and tertiary timings set to auto. PBO enabled. Asus "Performance Enhancement Level 2".
2400 auto PBO PE2 = 2400MT, all timings auto. PBO enabled. Asus "Performance Enhancement Level 2".
Optimized defaults 2400 = All settings default/auto (defaults to 2400MT).
I used Time Spy to simulate a very GPU limited situation. While the CPU scores changed quite a bit with the faster memory configuration, the GPU scores are within the run-to-run margin of error.
I used Night Raid to simulate a CPU limited situation. Both the CPU and the graphics scores changed significantly with the faster memory configurations.
I also ran the API benchmark. The DX11 single threaded benchmark showed a nice boost in performance with faster memory configurations. But the DX11 multi-threaded benchmark showed an even larger increase in performance. DX12, which is inherently multi-threaded, also showed a significant performance uplift. And Vulkan, also inherently multi-threaded, more than doubled its score. Vulkan also appears to be particularly sensitive to memory latency. Going from 3533MT CL14 + auto secondary/tertiary timings to 3533MT CL14 w/tuned timings showed quite an uplift in performance.
Rainbow Six Siege is the most demanding game I play right now. It is also unique in that it has both a DX11 client and a Vulkan client, allowing me to do as close of an apples-to-apples comparison between the two APIs as you can get. I used the in-game benchmark and was on the low preset at 1080p to ensure a CPU limited situation. Both clients showed ~25% gains, though the Vulkan client started and ended with a higher overall frame rate.
The OCCT power draw row speaks for itself. Even without PBO, running at 3533 with fast + tuned timings results in a ~50% increase in power draw over the rated 105W TDP.
****TLDR**** - So, in summary, while I am way ahead of "stock" performance just from the memory OC, I want to see if I can squeeze more with a fixed all core OC. Which the memory OC makes difficult. Even a slight PBO undervolt results in instability (-0.05v offset).