Hacker News Reader

Comment by MindSpunk | Hacker News Reader

Taking games designed for desktop GPUs and running them on mobile GPUs with tile-based-deferred-rendering hardware will be a disaster. Mobile GPU designs will choke on modern games as they're designed around hardware features that mobile GPUs either don't have, or that run very slowly.

Peak theoretical throughput for the GPUs you find in ARM SoCs is quite good compared to the power draw, but you will not get peak throughput for workloads designed for Nvidia and AMD GPUs.

View on HN

Isn't the GPU on Apple Silicon machines a tile-based "mobile" GPU design? Many of the hardware features that traditional GPU's have and mobile GPU's lack can be easily "faked" with GPU-side general compute.

But even the most powerful apple silicon GPU is terrible compared to an average Nvidia chip

While I agree with the general point, this statement is factually incorrect - apple's most powerful laptop GPU punches right about the same as the laptop SKU of the RTX 4070, and the desktop Ultra variant punches up with a 5070ti. I'd say on both fronts that is well above the average.

And it will consume almost as much power as the Nvidia GPU to do so.

There is no world where Apple silicone is competing with a 5070ti on modern workloads. Not the hardware and certainly not the software where Nvidia DLSS is in it's own air with AMD just barely having gotten AI upscaling out and started approximating ray reconstruction.

Apple's MetalFX upscaler is pretty similar to DLSS (and I think well ahead of AMD's efforts on this front).

Ray tracing outside of Nvidia is a disaster all round, so yeah, nobody is competing on that front.

Ray tracing support on the newest AMD chips is getting good enough. They are still behind Nvidia but definitely not a disaster anymore

Certainly, nobody would buy an Apple hoping to run triple-A PC games.

But among people running LLMs outside of the data centre, Apple's unified memory together with a good-enough GPU has attracted quite a bit of attention. If you've got the cash, you can get a Mac Studio with 512GB of unified memory. So there's one workload where apple silicon gives nvidia a run for their money.

Only in the size of model it can run, not speed of token generation.

Are there real world game benchmarks for this or are these synthetic tests?

Only a few, because it's not easy to find contemporary AAA games with native macOS ports. Notebookcheck has some comparisons for Assassins Creed: Shadows and Cyberpunk 2077[1]

[1]: https://www.notebookcheck.net/Cyberpunk-AC-Shadows-on-Apple-...

That simply isn't true. I have an RTX 4070 gaming PC and an M4 MacBook Pro w/ 36GB shared memory. When models fit in VRAM, the RTX 4070 still runs much faster. Maybe the next generation M5 chips are faster but they can't be 2-4x faster.

GP said laptop 4070. The laptop variants are typically much slower than the desktop variants of the same name.

It's not just power budget, the desktop part has more of everything, and in this case the 4070 mobile vs desktop turns out to be a 30-40% difference[1] in games.

Now I don't have a mac so if you meant "2-5x" when you said "much faster" well thdn yea, that 40% difference isn't enough to overcome that.

[1]: https://nanoreview.net/en/gpu-compare/geforce-rtx-4070-mobil...

a 4.5k$ M4 Max barely competes with an entry-level laptop with a 4060 which will be around ~1K in FPS in cyberpunk given the same settings. For AI it's even worse - on NVidia hardware you're getting double-digit speeds for FPS for real-time inference of e.g. stable diffusion, whereas on the M2 Max I have you get at best 0.5 FPS

Snapdragon doesn't do tile based deferred rendering the way Apple does (or did). Snapdragon does (or did) a form of tile-based rendering, but it is a completely different design, with completely different performance tradeoffs.

What about the Switch 2 (nVidia's Tegra) line? The one in the Swich 2 is using Ampere architecture.

That should be feasible, no?

Linux support is in a terrible state for Nvidia chips. Not going to happen.

You don't have to use tile-based rendering on these chips anymore. They can directly draw to the entire screen.

You can, but the immediate mode path is slower and uses significantly more power. Mobile GPUs are not good at modern desktop game workflows where significant portions of the frame are compute shaders. They're generally very memory bandwidth starved, and general compute sidesteps most of the optimizations the hardware has made to work around this.

Reply on news.ycombinator.com