M5 Ultra and M6 Chip Benchmark Results Reveal Graphics Performance

M6 isn't far off the M4 Pro (just thinking in the context of the Mac mini). Insane.

Some impressive hardware - especially when you consider the performance per watt.

I mean it's getting to the point where only AI workloads and certain intensive creative work can make use of this sort of compute. My M1Pro with 16GB RAM is still perfectly serviceable for pretty much anything that normal computer users need. I mean the hardest I push it is 4 x 4K multicam in FCP and it crunches through that quite happily. I'm not sure what I'd to with an M5 ultra!

If the gains are equally impressive two years from now with the M8, then I’ll switch up from my M4 mini as long as it doesn’t involve selling organs.

I realize the author wanted some score to compare to discreet graphic cards on the PC side but isn’t OpenCL pretty much deprecated on Mac for years now?

Yes. It was deprecated back in 10.14 in favour of Metal: https://developer.apple.com/opencl/

Sadly, there is no comparable metric because Windows and DirectX (coupled with games being built for Microslop's rendering APIs and then ported to macOS's Metal framework) are just not comparable. There are way too many moving parts. Power draw is another huge factor given ARM runs at a sliver of what X86-64 does. 5090s are literally being fried because the voltage is cranked up so high on just one or two pins. Good job Nvidia!

Vulkan is kind of the bridge between macOS and Windows and even then, it still differs quite drastically between the two environments in how it's delivered.

But given that Apple continues to make solid progress while Nvidia and AMD abuse their market dominance (Intel is still asleep at the switch), it's quite likely that if all things were considered equal, Apple may be ahead now. Again, not that it matters when everything is pushed through a translation layer and poorly optimized for the platform to boot, but if you are comparing raw compute vs power draw, Apple excels.

A shame Apple doesn't buy a gaming studio and develop real games made specifically for Apple silicon. Then they could really showcase the true power of their SoCs.

FWIW, Geekbench scores have historically meant nothing in the real world.

The M6 is the real story here. That's a huge generational jump.

That’s amazing. I wish I could afford the 512GB RAM model when that comes out in October. From what I can tell, I think DeepSeek 4.1 can fit on it, and it’s so fast and efficient and benches really well on DeepSWE. But using the API is a lot more cost-efficient at this point, it’s so freaking cheap. I keep blowing through my Codex usage so I’m currently working out a better solution by having Astra orchestrate with DS4.1 doing the grunt work, because even Sol is eating usage now.