[ksec]

It is important to Note, both the 2060 Super and RX5700 has similar transistor count, ( 10.8B vs 10.3B ). So RX5700 isn't winning because it has new node and stuffing in more transistor.

I am wondering on the state of GPU Compute on Mac, CUDA, OpenCL, ROCm. Apple has now abandoned OpenCL and working on their Metal Compute, which is ( I think ) only available on iOS. CUDA is not available on Mac, and ROCm is only available on Linux.

[rrss]

> RX5700 isn't winning because it has new node and stuffing in more transistor

Sure, I guess, but it is probably still winning because it is manufactured on a new node - transistor count isn't the only thing that matters.

Performance is frequently power-limited, so because 7nm is more efficient it can achieve higher performance.

If you took the same design from 16nm (or 28, ...) and simply shrunk it to 7nm, you could achieve higher performance with the 7nm version - even though the number of transistors hasn't changed.

[kitsunesoba]

Metal compute is available on macOS as well. As far as I can tell, it’s been available since at least 10.11.

[ksec]

Thanks I was not aware of that. So Intel and AMD has to write drivers specifically for it?

[ohyea]

You can use Intel PlaidML right now on OSX to utilize your metal supported GPU for machine learning. I have tried the python library with Keras and my GPU (Geforce 760) with no issues.

[tmd83]

That is good news. On the other hand compared to Zen2 the power usage is kind of disappointing. Navi is taking similar or even significantly higher power while using 7nm vs. nvidia 12nm.