AMD: “Maxwell is not capable of asynchronously executing graphics and compute” | PCGamesN

AMD: “Maxwell is not capable of asynchronously executing graphics and compute”

Nvidia and AMD

Following claims from developers Oxide that NVIDIA’s Maxwell GPUs can’t handle asynchronous compute/shading in DirectX 12 applications, and a communtiy-created benchmark which appeared to disprove that, AMD’s senior graphics PR manager Antal Tungler has reached out to PCGamesN in response.

Oxide stated that NVIDIA’s Maxwell GPU family couldn’t process graphics/compute commands asynchronously in DirectX 12 game Ashes of the Singularity. To put that to the test, a Beyond3D forum user MDolenc built a benchmark that throws increasing numbers of graphics/compute tasks at a given GPU. The basic idea: if there’s a linear increase in time taken by that GPU to process greater and greater command queues, it isn’t running asynchronously.

The results showed that Maxwell cards processed up to 31 command queues without any latency increase, the time taken increasing further with every 31 additional queues. Many users over at Beyond3D and elsewhere – including this guy – took that as a strong indication that NVIDIA cards were, in fact, capable of executing asynchronous graphics/compute.

But AMD see the results as proof that they’re not. Here’s the full statement from Tungler:

“The results produced by the benchmark do, in fact, illustrate that Maxwell is not capable of asynchronously executing graphics and compute. If you look at the Maxwell async compute results, you will see that the bar heights are the result of adding graphics and compute together. This indicates that the workloads are being done serially, not asynchronously. Compare that to the AMD results, where the async compute results show graphics and compute being processed simultaneously with no noticeable rise in overall frame latency.

“If Maxwell supported asynchronous compute, their results would look like the GCN results. Remember that asynchronous compute isn’t whether or not a GPU can do compute and graphics across a long workload, it’s whether or not the GPU can perform these workloads simultaneously without affecting the frame latency. MDolenc’s benchmark clearly shows that only GCN can do this.”

We’ve reached out to Oxide and NVIDIA for comment, and will update you as soon as they respond.

Clicking on links in articles to retailers or publishers may mean we earn a small commission.