Hacker News Reader

Comment by BearOso | Hacker News Reader

Polaris and Vega were ROP-limited, and performed worse than nvidia cards with similar tflop measurements.

The Xbox has a higher tflops rating because of more compute units, but it has the same number of raster units. Because the PS5 clocks higher, its ROP throughput will be greater. In a rendering workload like games this will make a difference. Whether or not this offsets the fewer compute units remains to be seen, but I’d wager the consoles are going to come out pretty much on par. There’s not going to be a big discrepancy like PS4 and Xbone.

View on HN

You're missing one of the key pieces: software. I don't think many would argue that Sony can produce better tooling and operating system efficiency than Microsoft.

Even with identical hardware, I'd place my money on a small Xbox edge.

The PS3 was on paper significantly more powerful than the Xbox 360, however it didn't result in significantly better performance or graphics. In some games in some cases yes, in others not really.

Doesn't Sony use clang? My understanding was LLVM does a better job at optimisation than MSVC.

Whether the texturing/antialiasing units (called ROP in this GPU) are a bottleneck depends on the program you run on the GPU, and how it's tuned to the underlying fixed console hardware, and what tradeoffs are made to do work on the ROPs vs other hardware. With consoles the benefit of the fixed hardware configuration is that you can tune the program so that it's using the hw in a balanced way.

I’m not too familiar with AMD/XSX architecture, but why wouldn’t ROP scale with number of GPU cores? Why would MS design something with lower ROP/core? Also, where did you get these details from?

ROPs are only useful for pushing pixel out, if you intended to use a large enough portion of the GPU for compute you don’t need as many ROPs.

Since compute shaders can write out directly into memory you don’t need to use the ROPs to write the result into a buffer / texture like in the early days of GPGPU.

If you assume 20-30% or even more of your frame GPU budget is spent on compute you can reduce the number of ROPs to match.

Unless you have other bottlenecks you really only need as many ROPs as the fill rate you’ll be targeting which will be dependent on the resolution and targeted frame rate.

10-15 years ago ROPs and TMUs scaled 1:1 with GPU “power” the more you moved the load into shaders and then compute the less meaningful those figures are.

Reply on news.ycombinator.com