New blog: how FFmpeg uses Vulkan Compute shaders to accelerate professional codec workflows — FFv1, ProRes, ProRes RAW, APV and more — on any Vulkan-capable GPU, with no fixed-function hardware required.

Fully GPU-resident, no CPU hand-offs, and portable across the broad Vulkan device ecosystem.

Learn more: https://www.khronos.org/blog/video-encoding-and-decoding-with-vulkan-compute-shaders-in-ffmpeg?utm_medium=social&utm_source=mastodon&utm_campaign=FFmpeg_Video&utm_content=blog
#Vulkan #FFmpeg #VideoEngineering #GPUCompute #OpenStandards #ProRes #FFv1 #ProResRAW #APV #VC2 #JPEG

The Hard Maths Behind AI Token Costs: Why Simple Wins Are Disappearing

As AI inference scales, the real economics come down to latency-throughput trade-offs, software optimisation, and ruthless cost discipline—not just raw chip performance.

The Daily Perspective

How do ML teams actually share GPU clusters when demand outstrips supply?

Three dominant allocation models keep appearing:

1️⃣ Quota-based allocation (virtual clusters)
Give each team a guaranteed slice. Others can borrow your idle capacity, but you can reclaim it.

Microsoft's HiveD takes this further with "topology-aware" guarantees so distributed training jobs get GPUs that can actually talk to each other efficiently.

#MLOps #GPUCompute #MachineLearning

Vulkan question: when creating multiple pipelines with vkCreateComputePipelines which are all specializations of the same shader code with the same pipeline layout do drivers optimise this better than I could by just submitting each creation to a task on a separate thread?

I'm testing this right now on the hardware & platforms I have, but was wondering if others had wider experience with this.

#Vulkan #GPUCompute #Shaders

One thing I haven't been able to find out is if any optimization is happening when I pass in my wgsl code. Does it do constant folding, does it do load optimization, etc?

#WebGPU #GPUCompute

In my playing around w/ WebGPU, I think I have a bilinear debayer function working that runs on the GPU.

As usual, the hole is deep when looking beyond the surface (e.g algorithms, optimizations).

#WebGPU #GPUCompute

Well, I'm half impressed with wgpu. It managed to compile fine on my 9 month old version of rust, and a few of the demos worked, but not all of them.

Just over half worked, 15 of 29. (Note that I assume that an example worked if it popped up a window to show me something. If the example just prints something out, I might have considered it a fail. Also, this may be because the ray tracing support is experimental, and 6 of the tests are of this type.)

#wgpu #GPUCompute

As usual, getting something like GPU compute that's cross platform working is a message because everyone likes to do their own thing and reinvent the wheel.

I would like something that is [modern] macOS and FreeBSD compatible, but doesn't look like that's possible since Apple deprecated OpenCL.

(Also, could Apple have picked a less searchable term for their new GPU framework?)

It's again looking like the best way to be cross platform is to use JS+browser.

Or am I missing some library?

#OpenCL #GPUCompute #FreeBSD

Disabling Intel Graphics Security Mitigations Can Boost GPU Compute Performance By 20%

While not talked about as much as the Intel CPU security mitigations, Intel graphics security mitigations have added up over time that if disabling Intel graphics security mitigations for their GPU compute stack for OpenCL and Level Zero can yield a 20% performance boost

We’ve just released NAP 0.7.7— a minor update with a bit of a punch! Including a new integrated FFT system module, a cool audiovisual demo, support for newer versions of Visual Studio and many other improvements and bug fixes.

Check out the release notes here: https://github.com/napframework/nap/releases/tag/v0.7.7

pps: turn on audio for this video:

#nap #napframework #realtime #vulkan #gpucompute #audiovisual #cpp #opensource

Release NAP 0.7.7 · napframework/nap

Incremental release that adds support for new platforms, bumps various dependencies, introduces some new features and fixes some outstanding issues. This is the final incremental release before NAP...

GitHub