Clock Rates, Temperatures & Noise
Temperature & Clock Frequency
A maximum temperature of 85°C is reached fairly quickly due to conservative fan control. At that point, the card loses approximately 6% of its performance compared to when it was still cool. This is almost entirely due to the automatic clock frequency throttling by ~9%. We calculated an average clock frequency across the entire run for every 5°C step. The results ranged from 1533 MHz when the card was cool all the way down to 1401 MHz when it was hot.
Temperature Curves & Power Consumption
This where things get interesting. At the warmed-up card’s 1401 MHz, we measure approximately 285W. However, at the cold card’s 1533 MHz, we measure approximately 310W. This means that a 9% power consumption increase nets us a 9% frequency increase, which, in turn, yields a 6% gaming performance increase. In other words, the efficiency curve is already starting to dip. There’s not much headroom left.
That also means losses due to leakage don’t really play much of a role anymore. The days when 40W could be saved by keeping the card cool at the same frequency are gone.
GPU vs. HBM2 Temperatures
Unless the sensors are lying to us, the GPU’s maximum temperature is 84°C (85°C peak), while the HBM2 gets up to 90°C (94°C peak during the stress test). That latter figure seems fairly high, but it does end up close to the ceiling for GDDR5X. We'll keep an eye on both temperatures during future tests; it's important to ensure the sensor data is 100% accurate.
During the stress test, the card heats up so quickly that the curves for the open and closed PC cases are practically on top of each other.
What jumps out to us is that the board just below the GPU reads ~5°C cooler than the inside of Vega 10! The obvious question is why. To answer, remember that AMD's Vega 10 GPU and its HBM2 sit on an interposer attached to a package substrate positioned on top of the PCB. The interposer doesn’t seem to make full contact with this substrate, causing a so-called underfill issue. Air between the layers acts almost like insulation.
During the stress test, temperatures are a little lower due to the fan spinning faster and Vega 10's lower (throttled) frequency.
Using the primary BIOS and the Balanced power profile, Radeon RX Vega 64 generates a maximum of 48.2 dB(A). Switching to Turbo mode causes the card to exceed 50 dB(A).
We praised AMD’s cooling solution in our Vega Frontier Edition review. However, a pleasant breeze turns into a raging tornado this time around due to power consumption that's way too high. Then again, an Nvidia GeForce GTX 1080 Ti Founders Edition with a 295W maximum power target is almost as loud.
The version of AMD’s cooling solution used on Radeon RX Vega 64 is significantly different from the one we found on Vega Frontier Edition. It’s too aggressive, too hot, and, of course, too loud. Designing a thermal solution to be "good enough" never works out well for thermals or noise.
MORE: Best Graphics Cards
MORE: All Graphics Content