Welcome Stranger to OCC!Login | Register

A Look at RX Vega 64 Efficiency

   -   
» Discuss this article (5)

Procedure:

For these recordings I used three features of AMD's Adrenalin drivers; Performance Monitoring, Frame Rate Target Control (FRTC), and Wattman. Performance Monitoring allows the user to record various measurements of AMD GPU performance, including ASIC power use, core and memory speeds, fan speed, and temperature. It can make these measurements at most once a second. With FRTC the user can impose a frame rate limit on a game through the driver configuration utility, or through the Radeon Overlay that can be brought up. It does have its quirks, such as requiring the game run in (Exclusive) Fullscreen to function, or at least for the overlay to be available and control the FRTC target. (As The New Colossus uses the Vulkan graphics API, its 'Borderless' option still supports both the overlay and FRTC. My impression is that Vulkan's borderless option is actually fullscreen, but it might be more resilient to Alt+Tabbing than fullscreen is for some other APIs.)

After finding a location to make my runs in (Roswell – Underground for The New Colossus, a path in Nurnen for Shadow of War, and the Outpost in Killing Floor 2) I did a recording using the usual settings for my RX Vega 64, which is with an undervolt and the Power Limit slider raised to +50%, configured within the Wattman feature. Making a recording involves beginning the Performance Monitoring recording, starting OCAT, a utility for capturing frame time information, and a timer. While OCAT can be set to stop after a period of time (300 seconds or five minutes), Adrenalin Performance Monitoring cannot be, so I used this timer to let me know when to manually stop the recording. Using the OCAT data, I determined the average frame rate and then set FRTC to that value to make another recording. I then stepped down by 10 FPS for each run, finishing with a 60 FPS run.

To make the Undervolt (Power Limit at 0%) and Stock runs I took advantage of the Wattman feature to load saved profiles. Curiously the Performance Monitoring feature was not working for me when just using the Balanced configuration in Wattman (which is what I consider as stock), which is why I apply the Balanced settings to the Custom configuration option. You can get these same settings by applying the Balanced configuration and then returning to the Custom Configuration. Usually my custom profiles also have a change to the fan limits, 500 RPM to 3000 RPM instead of the stock 400 RPM to 2400 RPM, but I manually changed the fans to the stock setting, so no profile had an advantage over another due to better cooling.

Before going further, here are my system specs:

  • Processor: AMD Ryzen Threadripper 1950X @ 4.0 GHz (The New Colossus) or 3.7 GHz (Shadow of War, Killing Floor 2)
  • Cooling: Enermax Liqtech TR4 240 AIO
  • Motherboard: ASUS Zenith Extreme
  • GPU: AMD RX Vega 64 at Stock Clocks
  • Memory: G.Skill TridentZ 4x8 GB (32 GB) @ 3200 MHz 14-14-14-28
  • PSU: OCZ Fata1ty 750 W
  • OS: Windows 10 Pro 64-bit version 1709
  • Drivers: Radeon Adrenalin Software 18.3.4

I previously mentioned I missed a run with The New Colossus. This was the 130 FPS target run when Undervolted with the +50% Power Limit. It took some research and experimentation, but I eventually managed to get R, the statistics application I am using to process the data and produce graphs, to add an empty run at this target, so while the data is missing, the graphs between the three different sets in these games have identical formatting.

My usual Wattman Profile

 

I should mention, as it is very important, that the way the RX Vega 64 (and RX Vega 56) undervolt is unusual. Within Wattman there are three voltage settings you can change: GPU State 6; GPU State 7; and Memory State 3. These states refer to the P-state or power state (so GPU P6, GPU P7, and HBM P3), but the voltage setting for Memory State 3 is not directly the voltage setting for the memory. Instead this appears to be a floor voltage for the GPU though it does likely impact the memory voltage. What this means is that while I have the voltages for GPU P6 and GPU P7 set to 900 mV and 950 mV, respectively, the actual minimal voltage for these is the 965 mV setting in HBM P3. Why do I bother with the lower values for those two GPU P-states? Because I have never taken the time to change them, but I also think there is a chance this is still important, which I will get to later with the Killing Floor 2 data.

I do quickly want to explain what P-states are, at least based on my understanding. The power states for a processor, GPU and CPU, are pairs of clock speed and voltage settings, so at a specific P-state, the processor will run at that clock speed and use that voltage. While they are a very powerful feature of modern hardware, allowing the processors to reduce power when it is not needed and remain stable at high speeds, they are not as firm as they once were. The boost technology the Vega 10 GPU has allows it to run at clock speeds between the P-states, which we will see in the data. I am unsure if the voltage will take an intermediate value at these times. Also I am unsure if the different HBM P-states will result in different floor voltages, but I have reason to think it does, as we will see later.

I should also mention that the power usage of a processor (GPU or CPU) is dependent on both its voltage and its clock speed. If you fix the voltage of a processor you will see the power use increase as its clock speed increases. This is because it is the switching of the transistor within the processor that uses energy, and higher clock speeds means these transistors are switching faster, using more energy in a unit of time. Power is energy used in a period of time, with a Watt being defined as one Joule, the metric unit of energy, per second, so this is why power use is directly related to clock speed. Voltage is directly related to the energy (one Volt is one Joule divided by the charge difference between two points) so it too is directly related to the power use. My sense is that power is more strongly related to the voltage than the frequency, but both contribute.

I think that is enough here, so time to move to the The New Colossus data and graphs.




Related Products
Random Pic
© 2001-2018 Overclockers Club ® Privacy Policy
Elapsed: 0.1632530689   (xlweb1)