With NVIDIA launching Pascal tomorrow, and AMD set to host a special live stream for their Polaris architecture on Wednesday, both GPU manufacturers are basically clawing at each other’s throats, desperately trying to find their footholds to grab market share in the coming GPU bloodbath. While NVIDIA seems to be ready and waiting for AMD with the Geforce GTX 1080 and GTX 1070, AMD is also ready for a fight with upcoming Polaris 10 GPUs. Benchmark results for Polaris-based cards running the Ashes of the Singularity test have been popping up on the site’s leaderboard, and they are very interesting.
Because AMD is careful not to reveal model names just yet, the current crop of Polaris benchmarks on the internet use their hardware ID instead. In a recent patch to AMD’s AMDGPU Linux kernel driver, six other card names could be seen alongside their codenames Baffin and Ellesmere, which represent Polaris 10 and 11 respectively. Searching for these device ID strings can lead you to some benchmark results, but the model numbers themselves are quite generic and using Google often brings up thousands of unrelated hits.
The ones that are currently out there are 67DF:C7 and 67DF:C4, which have been found in other tests, both from the Polaris 10 family. The change in the numbers after the colon are models that derive from the fully unlocked GPU, so C7 is the unlocked GPU and C4 is the cut-down one that will be sold for a bit cheaper. The benchmark results seem to support this as well.
Here the benchmarks for the 67DF:C4 variant of Polaris 10 look very promising. At the 4K Crazy setting for the game, the average framerate is down at 20.5fps, but there’s very little variance between the average batch run and the individual batch results, which means that the card is performing right on its limit. Two users have submissions for the same card in different systems weeks apart, which suggests that this is either AMD engineers testing the cards, or reviewers getting early hands-on time in preparation for a launch. There’s one more submission running at UltraHD 5K resolution by yet another user, as well as two more at 1080p with low settings. One is much higher than the other, possibly as a result of major changes in the game’s code that improved performance.
The next set of results are for 67DF:C7, which seem to be benchmarked at lower settings, but this could be intentional on AMD’s part to hide the true performance of the cards. Note the username submitted, “ATI_MA_DEV_03”, which is either a troll trying to throw people off, or a real developer with several cards on hand to test. These results look somewhat more interesting, although the 4K performance still leaves a lot to be desired, and it’s only a tiny bit ahead of the 4K Crazy results for the 67DF:C4 variant of Polaris 10 as well.
There are more listings to be found as well. Tested at 1080p with the standard preset is an AMD GPU with the codename 67EF:C3, a model number that currently does not show up in any driver list anywhere on the internet. It would be a new low-power GPU AMD has never talked about, because its performance barely scrapes playable at those settings. Another test, this time at 1080p low using DirectX 11, was completed by the card 67FF:C8, which is a Baffin GPU. The performance is really good, but the MSAA has been dropped to 1x, and most options are turned off.
Going by the results, I think we can more or less figure out card placement, but most of the settings are different and there’s no way of comparing cards to each other because they’re also in different systems. We’ll have to wait and see what AMD says in their live stream event on wednesday, and then come back to compare settings and performance, and hopefully get a glimpse of what AMD is planning.