[bigyabai]

I hate to break it to ya, but Apple Silicon isn't in the top 25 highest-performing consumer GPUs. It's probably not even in the top 25 most-efficient either: https://browser.geekbench.com/opencl-benchmarks

[rafram]

That chart shows that M4 achieves 25% of the Geekbench scores of GPUs that pull >10x more power. That's definitely efficient.

[bigyabai]

Are you comparing it with other 3nm GPUs? When you normalize for process, Apple Silicon is definitely not the most efficient raster architecture.

[rafram]

It doesn't seem like Nvidia even has any 3nm GPUs on the market. But sure. When you control for power efficiency, it turns out there's no difference at all!

[bigyabai]

Process is not equivalent to power efficiency. It's a step-change enabling better designs.

Apple and Nvidia both have 5nm and 4nm GPUs. Take those scores and divide it by the TDP, you'll be shocked at the difference design can make.

[wtallis]

Please never divide anything by TDP. Use actual power measurements, unless you're trying to ensure your numbers end up being bullshit. (In particular, any number someone claims is a TDP for an Apple processor is made up, because Apple doesn't publish or specify any quantity remotely similar to TDP.)

[bigyabai]

Okay, then don't divide by TDP. Measure the GPU wattage frame-by-frame and you'll still end up with similar numbers. The point stands.

> because Apple doesn't publish or specify any quantity remotely similar to TDP

1) That doesn't mean that power usage isn't measurable.

2) They actually do, although it's not a perfect breakdown chip-by-chip: https://support.apple.com/en-us/102027

[lunaticlabs]

The big advantage of macs when it comes to GPUs isn't their direct speed, its the unified memory model. If I want to buy a GPU that has 64-128GB of addressable memory, it will cost me an enormous amount and the computer itself will be a server module for racks that is loud and not a consumer PC. You can buy a mac with a unified memory model, and even though its GPU is not on the top rankings, the fact that it can operate on your model in regular memory is what gives it its advantages.

[jiehong]

Neither is their target, they are more in the perf/watt segment.

[bigyabai]

Which is why it's confusing that the M3 Ultra is less efficient than several 130w laptop chips.