The problem with DX12/Vulkan isn’t just that “low-level control is hard”, it’s that a lot of performance-critical decisions are now exposed at a level where they’re extremely GPU- and generation-specific. The same synchronization strategy, command ordering, or memory usage can work great on one GPU and badly on another.

A GPU ISA wouldn’t fix that, it would push even more of those decisions onto the developer.

An ISA only really helps if the underlying execution and memory model is reasonably stable and uniform. That’s true for CPUs, which is why x86 works. GPUs are the opposite: different wave sizes, scheduling models, cache behavior, tiling, memory hierarchies, and those things change all the time. If a GPU ISA is abstract enough to survive that, it’s no longer a useful performance target. If it’s concrete enough to matter for performance, it becomes brittle and quickly outdated.

DX12 already moved the abstraction line downward. A GPU ISA would move it even further down. The issues being discussed here are largely a consequence of that shift, not something solved by continuing it.

What the blog post is really arguing for is the opposite direction: higher-level, more declarative APIs, where you describe what you want rendered and let the driver/hardware decide how to execute it efficiently on a given GPU. That’s exactly what drivers are good at, and it’s what made older APIs more robust across vendors in the first place.

So while a GPU ISA is an interesting idea in general, it doesn’t really address the problem being discussed here.

But the driver can't decide how to execute efficiently more than the application does, that's how we got the modern APIs. The declarative API would necessarily have to tackle very specific use cases, which again is what the older APIs did.

So I guess we're stuck with what exists today for a while.