- Joined
- Jul 20, 2020
- Messages
- 851 (0.61/day)
| System Name | Gamey #1 / #2 |
|---|---|
| Processor | Ryzen 7 5800X3D / Core i7-9700F |
| Motherboard | Asrock B450M P4 / Asrock B360M P4 |
| Cooling | IDCool SE-226-XT / CM Hyper 212 |
| Memory | 32GB 3200 CL16 / 32GB 2666 CL14 |
| Video Card(s) | PC 6800 XT / Soyo RTX 2060 Super |
| Storage | 4TB Team MP34 / 512G Tosh RD400+2TB WD3Dblu |
| Display(s) | LG 32GK650F 1440p 144Hz VA |
| Case | Corsair 4000Air / CM N200 |
| Audio Device(s) | Dragonfly Black |
| Power Supply | EVGA 650 G3 / Corsair CX550M |
| Mouse | JSCO JNL-101k Noiseless |
| Keyboard | Steelseries Apex 3 TKL |
| Software | Win 10, Throttlestop |
This thread is about power-optimizing the Radeon RX 7700 XT but to start, please read W1zzard's bullet point notes about overclocking in the 7700 XT reviews. It's a great way to OC and is a good starting point, but when power limiting one of those settings (Max Frequency) will trip you up.
www.techpowerup.com
I'm not going to get into stock performance for my 7700 XT but it's the Sapphire Pulse version, same as in the link above. Instead I UV/OC everything I get because why throw away potential performance or power savings with conservative stock settings? I'm using Time Spy (DX12) and Fire Strike Ultra (DX11) Graphics Scores as my tests for easy repeatability and both are 1440p like my display. The FPS deltas will be a little different in other games.
Test system:
AsRock B450M Pro4 (PCIe 3.0)
Ryzen 5 5600 OC to 4650MHz
32GB 3200 MHz CL16 DDR4
4TB Team NVMe SSD
The tl;dr for my power-optimizing interests is that IMO the Power Limit slider in AMD's Adrenaline is too limiting but not the upper limit, instead at the lower limit. This card supports +15% and it seems the standard lower limit in Adrenaline is -10%. Maybe that happens to be a great efficiency cutoff but I very much doubt it, especially for DX11. In Time Spy and Fire Strike Extreme, after setting the card's undervoltage settings properly, the FPS difference between 208W (-10%) and 262W (+15%) was:
3.6% fewer FPS in Time Spy (cores at 2580-2630 MHz @-10%, ~120 MHz higher @+15%)
2.7% fewer FPS in Fire Strike Extreme (cores at 2670-2720 MHz @-10%, ~130 MHz higher @+15%)
with 20.6% less power used
Having the Power Limit slider scale down to -25% like Nvidia commonly allows in Afterburner in would be much more useful. Yes, you can instead set Max GPU Frequency to a lower limit like 2500 or 2300 which will work but if the point of Power Limiting is to literally limit the power and not the speed, then directly controlling power is better. I'm familiar with this from 75W slot-powered GPUs like the 1050 Ti where undervolting and setting a hard MHz limit wastes power under lighter loads, whereas simply using a smooth undervolt curve with the existing 75W cap realized more overall FPS as every bit of available power is being used at all times, clocking higher and lower as needed.
OK how to do this: First follow W1zzard's instructions above to find your card's max performance settings and definitely heed his +0.05v safety offset. Those last little slivers of performance aren't worth the potential instability. My 262W +15% OC settings are the result of those instructions. So now you can just swing that Power Limit slider down to -10% and get a cooler running and lower power GPU, right? Nope.
Key point from W1zzard: "... but rather acts as some sort of guidance for AMD's clocking algorithm." The 7000's clocking algorithm is different from the RX 6000 series. I have a few of those I'm very familiar with OC/UVing, and I needed to be retrained for this 7700 XT. What the algorithm seems to do if you set Max Core Frequency to 5000MHz and start using a lower power limit, is steal power from the Memory subsystem in the form of lower memory clocks and give it to the Core clocks. So when doing those tests the memory clock was 1500 MHz instead of 2400 MHz (OC from 2250 MHz stock). Yes, the algorithm underclocked the memory with the Core clocks set to 5000 MHz! With that setting, the GPU was about 20% slower than at max power.
The solution is pretty easy: just take that Max Core Frequency slider and bring it down to about 3400 MHz and set the Minimum Frequency slider to about 2700 MHz. I used to leave the minimum at the default 500 MHz but that seems to let the low-load voltage sag too far and I would get GPU instability from time to time. With this setting, the VRAM stays at 2400 MHz at all loads in GPU-Z and the Core frequencies dipped down about 100 MHz to compensate for the increased power the now fully-enabledbattlestation memory subsystem was drawing. Importantly, it was overall 19% faster this way because of course memory bandwidth is extremely important.
The point of this is to say that optimizing 7000 series GPUs for high performance but modest power use is not as simple as: lower the power slider. There's much more to be gained with some experimentation.
Another observation which is completely different from any GPU I've used before is it seems the power scaling when limiting MHz (2500, 2300, 2100, 2000 MHz, etc.) is far more linear than with the RX 6000, RTX 2000, GTX 1000 cards I've used before. I hope to do more testing on that but it was a bit disappointing as usually there are good efficiency gains to be had doing MHz limiting on non-power limited cards. Instead those gains may be in the 2500-2900MHz range on this card, where they were in the 2300-2700 MHz range on RX 6000 and the 1850-2050 MHz Range on the Pascal and Turing GPUs.
Sapphire Radeon RX 7700 XT Pulse Review
The Sapphire Radeon RX 7700 XT Pulse is priced at AMD's MSRP of $450, which makes it an attractive choice. You still get a great cooler that's paired with outstanding fan settings that result in a whisper-quiet gaming experience that's quieter than most GeForce cards.
I'm not going to get into stock performance for my 7700 XT but it's the Sapphire Pulse version, same as in the link above. Instead I UV/OC everything I get because why throw away potential performance or power savings with conservative stock settings? I'm using Time Spy (DX12) and Fire Strike Ultra (DX11) Graphics Scores as my tests for easy repeatability and both are 1440p like my display. The FPS deltas will be a little different in other games.
Test system:
AsRock B450M Pro4 (PCIe 3.0)
Ryzen 5 5600 OC to 4650MHz
32GB 3200 MHz CL16 DDR4
4TB Team NVMe SSD
The tl;dr for my power-optimizing interests is that IMO the Power Limit slider in AMD's Adrenaline is too limiting but not the upper limit, instead at the lower limit. This card supports +15% and it seems the standard lower limit in Adrenaline is -10%. Maybe that happens to be a great efficiency cutoff but I very much doubt it, especially for DX11. In Time Spy and Fire Strike Extreme, after setting the card's undervoltage settings properly, the FPS difference between 208W (-10%) and 262W (+15%) was:
3.6% fewer FPS in Time Spy (cores at 2580-2630 MHz @-10%, ~120 MHz higher @+15%)
2.7% fewer FPS in Fire Strike Extreme (cores at 2670-2720 MHz @-10%, ~130 MHz higher @+15%)
with 20.6% less power used
Having the Power Limit slider scale down to -25% like Nvidia commonly allows in Afterburner in would be much more useful. Yes, you can instead set Max GPU Frequency to a lower limit like 2500 or 2300 which will work but if the point of Power Limiting is to literally limit the power and not the speed, then directly controlling power is better. I'm familiar with this from 75W slot-powered GPUs like the 1050 Ti where undervolting and setting a hard MHz limit wastes power under lighter loads, whereas simply using a smooth undervolt curve with the existing 75W cap realized more overall FPS as every bit of available power is being used at all times, clocking higher and lower as needed.
OK how to do this: First follow W1zzard's instructions above to find your card's max performance settings and definitely heed his +0.05v safety offset. Those last little slivers of performance aren't worth the potential instability. My 262W +15% OC settings are the result of those instructions. So now you can just swing that Power Limit slider down to -10% and get a cooler running and lower power GPU, right? Nope.
Key point from W1zzard: "... but rather acts as some sort of guidance for AMD's clocking algorithm." The 7000's clocking algorithm is different from the RX 6000 series. I have a few of those I'm very familiar with OC/UVing, and I needed to be retrained for this 7700 XT. What the algorithm seems to do if you set Max Core Frequency to 5000MHz and start using a lower power limit, is steal power from the Memory subsystem in the form of lower memory clocks and give it to the Core clocks. So when doing those tests the memory clock was 1500 MHz instead of 2400 MHz (OC from 2250 MHz stock). Yes, the algorithm underclocked the memory with the Core clocks set to 5000 MHz! With that setting, the GPU was about 20% slower than at max power.
The solution is pretty easy: just take that Max Core Frequency slider and bring it down to about 3400 MHz and set the Minimum Frequency slider to about 2700 MHz. I used to leave the minimum at the default 500 MHz but that seems to let the low-load voltage sag too far and I would get GPU instability from time to time. With this setting, the VRAM stays at 2400 MHz at all loads in GPU-Z and the Core frequencies dipped down about 100 MHz to compensate for the increased power the now fully-enabled
The point of this is to say that optimizing 7000 series GPUs for high performance but modest power use is not as simple as: lower the power slider. There's much more to be gained with some experimentation.
Another observation which is completely different from any GPU I've used before is it seems the power scaling when limiting MHz (2500, 2300, 2100, 2000 MHz, etc.) is far more linear than with the RX 6000, RTX 2000, GTX 1000 cards I've used before. I hope to do more testing on that but it was a bit disappointing as usually there are good efficiency gains to be had doing MHz limiting on non-power limited cards. Instead those gains may be in the 2500-2900MHz range on this card, where they were in the 2300-2700 MHz range on RX 6000 and the 1850-2050 MHz Range on the Pascal and Turing GPUs.