For our power testing, we used AIDA64 to stress the AMD Radeon RX 6800, then HWiNFO to monitor power use and temperatures.
After the stress test has ramped up the AMD Radeon RX 6800, we see it tops out at 203 Watts under full load and 7 Watts at idle. The power draw for the RX 6800 is very low considering past generation AMD GPUs high power draw. It also puts the power consumption very competitive with the GeForce RTX 3070 / RTX 2070 series.
A key reason that we started this series was to answer the cooling question. Blower-style coolers have different capabilities than some of the large dual and triple fan gaming cards.
Temperatures for the AMD Radeon RX 6800 ran at 68C under full loads, which shows AMD’s cooling solution performs very well and makes very little noise. The idle temps we saw were 33C, which is also excellent for a GPU of this size.
Overall, the AMD Radeon RX 6800 performed well across the workloads that we could run. Power consumption and thermals were very reasonable compared to the NVIDIA GeForce RTX 3070.
It seems like AMD has an edge in double-precision FLOP performance. We have received feedback that there are students out there who view this as a need. At some point, a data center card is going to be faster, but that is an area where AMD seems to have an edge in the consumer space. By the same token, we did not get to show off Tensor core performance now that NVIDIA is building AI acceleration into its GeForce parts.
From a software perspective, this is AMD’s big challenge. ROCm is getting better, but it is an extra step. For example, for our deep learning containers, we have to re-build the containers with ROCm, troubleshoot why things are not working, and then ensure we are getting something comparable to what we had on the NVIDIA side. Realistically that last step is specific to us but the first steps are more universal. If you are coding net-new work from scratch, this is likely not an issue, but CUDA acceleration is present in other applications as well, and not all are open source.
The good news is that software tends to iterate faster than hardware and AMD has a competitive hardware platform. Having a competitive hardware platform means that software developers are more willing to work on optimization. We think that AMD is on the right side of a cycle here. Frankly, the 16GB of memory in a card of this class is a great feature. Moving data over the PCIe bus if you run out of on-card memory today is a relatively slow operation. That is why we want to see NVIDIA cards like the NVIDIA GeForce RTX 3060 12GB with more memory than the 2-generation old $650 GeForce GTX 1080 Ti. AMD has delivered on that while NVIDIA has not necessarily done so in this price bracket (yet.)
Overall, this is a very nice card. AMD clearly has a new GPU platform that performs well.