Temperatures and Clock Rates
Remember that there are two representations of temperature in play here: edge, or GPU temperature, and junction temperature. In the measurements below, reported as GPU temperature, Vega 20 appears to operate coolly, even though we know the junction temperature AMD uses to control thermal throttling and fan control is quite a bit higher.
Temperatures do rise faster in the closed case and settle about 2°C higher than an open test bench. This isn't too painful at first. After all, AMD's throttling mechanism works differently from Nvidia's, where whole GPU Boost steps are lost as temperature increases.
As our test progresses and Vega 20 warms up, we see that Radeon VII starts switching back and forth between ~1500 MHz and 1740 MHz on the open bench. The jumps progressively become larger and more frequent as the temperature rises. In a closed case, the changes in frequency also seem to be dependent on the rate at which temperature changes. The card even sporadically drops to 1367 MHz!
Since fan speeds and temperatures are held within a very strict limit, clock rates are carefully balanced every step of the way. This behavior is very different from what we've seen from AMD's Radeon RX Vega and Nvidia's GeForce cards. In fact, we haven't seen such severe clock rate fluctuations from any other graphics card.
From what we can tell, Radeon VII doesn't really have a fixed boost frequency. So, you have to distinguish between peak, minimum, and average clock rates. Programs like WattMan and 3DMark reflect maximum values. But the practical relevance of those numbers is debatable. In order to make reliable observations about Radeon VII's boost clock, you need to record several seconds and calculate an average from them.
Let's again draw a comparison between Radeon VII and GeForce RTX 2080 Founders Edition using our benchmark results:
| Header Cell - Column 0 | Radeon VII Initial Value | Radeon VII End Value | GeForce RTX 2080 End Value |
|---|---|---|---|
| Open Test Bench | |||
| GPU Temperature | 34°C | 73°C | 75°C |
| GPU Clock Rate | 1,758 MHz | 1,498 to 1,741 MHz (Alternating) | 1,815 MHz |
| Ambient Air Temperature | 22°C | 22°C | 22°C |
| Closed Case | |||
| GPU Temperature | 36°C | 75°C | 75°C |
| GPU Clock Rate | 1,755 MHz | 1,367 to 1,747 MHz (Alternating) | 1,800 MHz |
| Ambient Air Temperature | 23°C | 46°C | 43°C |
Board Analysis: Infrared Images
We brushed the board with a special lacquer before collecting measurements, and calibrated the system with thermal sensors at four unique reference points. The homogeneous coating gives us a known emissivity of 0.975, and we take readings at an angle of 90 degrees.
The following infrared images, captured during our gaming loop and stress test in a closed case and on an open test bench, convey meaningful real-world information. Compared to WattMan, which made logging data problematic at best, our infrared measurements are far more reliable.
Differences in the gaming loop between our open bench table and closed case are visible, but within the expected range of two to three degrees.
Our stress test imposes up to 20W-higher power consumption, which is dissipated as waste heat. Interestingly, the temperature we read below the GPU package was higher than the value that WattMan reported as GPU temperature.
In a closed case, everything gets hotter. What's reported as GPU temperature nearly hits 80°C, and according to AMD, junction temperature should be even higher.
Unfortunately, we couldn't log the results of our stress test with WattMan because the program returned nonsensical data. In several measurements, the software simply stopped recording after only a few lines. At least we have values from our camera.
MORE: Best Graphics Cards
MORE: Desktop GPU Performance Hierarchy Table
MORE: All Graphics Content