A simple test of the throughput of atomics on your gpu using webgpu.
While building a very custom GPU memory allocator for my game engine, I've been relying heavily on atomics. This was inspired by the old CUDA blog post on warp-aggregated atomics which demonstrated that compiler magic can counterintuitively make certain GPU atomics extremely fast (faster than a CPU). I've been very curious to know if the same holds true for modern APIs and GPUs in general. Results indicate that even on non-nvidia systems and high level APIs such as WebGPU, these optimizations are clearly available!
PRs adding results for your GPU are welcome!
Current configuration is 32 atomic adds per thread, launching a total of 15M threads, all writing to a single global memory address.
| GPU | Max Bandwidth | Ops/s | Bandwidth Utilization* |
|---|---|---|---|
| M1 Max | 400 GB/s | 20B | 40% |
| RTX 4090 | 1008 GB/s | 62B | 49% |
*This may not be actual global memory utilization, but the utilization that would be required if operations were not aggregated prior to global memory.
-
Copy the result:
Operations per second -
Calculate results using this formula:
operations_per_second = # your result here
gpu_max_bandwidth = # your gpu max bandwidth (look this up online)
# 1 read + 1 write for a 4 byte u32
bandwidth_utilized = ((operations_per_second * 4 * 2) / gpu_max_bandwidth) * 100