Posts tagged cuda

TorchFX 0.6.0: FP32 on the GPU, CUDA Graphs, and a Hardened Realtime Path

TorchFX 0.6.0 is a performance and realtime release. The headline is the GPU follow-up promised back in 0.5.4: the CUDA kernels now run natively in float32 instead of silently upcasting to float64, which is 3.0–3.6× faster on consumer GPUs and finally lets the GPU beat its own CPU on multichannel workloads. On top of that, a new CUDA Graph path collapses the per-chunk launch overhead for streaming — up to 4× lower latency on short chunks — and the realtime engine moved its DSP off the audio callback into a dedicated worker thread.

Read more ...


FP32 on the GPU: 3–3.6× and the End of the Consumer-GPU Penalty

This is the GPU half of the promise we made in 0.5.4: “retuning the CUDA SOS kernel for mixed precision so float32 gets the same fast path on GPU that it now has on CPU.” TorchFX 0.6.0 delivers it.

Read more ...


CUDA Graphs for Streaming: One Launch Instead of a Launch Storm

For offline batch processing, GPU kernel-launch overhead disappears into the noise. For realtime streaming, it is the cost. TorchFX 0.6.0 adds torchfx.realtime.CudaGraphRunner, which captures a fixed-shape filter forward into a CUDA Graph and replays it per chunk — up to 4× lower per-chunk latency.

Read more ...


TorchFX 0.5.0: Custom CUDA Kernels & Native C++ Extension

I’m excited to announce TorchFX 0.5.0, a performance-focused release that introduces custom CUDA kernels, a JIT-compiled C++ native extension, and major algorithmic improvements across the entire filter pipeline.

Read more ...