Hacker News Reader

Comment by ls612 | Hacker News Reader

The rumors are that the M1 supersized version they are putting in the new Mac Pro is 20 CPU cores and 128 GPU cores, which would place it at a little under twice the performance of a 3090. Not saying that Nvidia can't catch up with a hypothetical 4090 next year but it'll be a tall order.

View on HN

The M1 Max 32 core GPU is ~1/8 the performance of the 3090. 4x the core count should put it half of 3090 performance, not double.

The path forward is simple then. 512 GPU cores.

Dumb question here: Can you get more performance from the same number of cores by giving them more power?

Not a dumb question at all. You can to an extent, but the returns diminish quickly. It isn't linear.

This fact however makes "performance per watt" comparisons misleading between different processors designed for different environments (e.g. M1 vs AMD Desktop). It takes a more power to get that much extra perf, conversely and the more under-appreciate part IMO reducing the speed a little can save a ton of power/heat if the chip is currently running at the higher power portion of the curve.

On a desktop machine most people would want them to tune for performance at quite a substantial power efficiency cost so of course a desktop chip most probably is less power efficient per unit of compute. You don't need to power it with a battery after all and there's heaps more cooling capacity in a bigger form factor so why optimise for that?

That's just not true. 3dmark shows it having a little under half the performance of a 3090 with the largest GPU, with it getting a score of 18,000 vs a 3090 having around 42,000. Scaling up linearly by a factor of 4 would lead to a comparison of 72k vs 42k, or a little less than twice as much.

It's not going to scale linearly though. The Max has 400mb of memory bandwidth. If you attempt to use as much as possible, as single Max chip uses 220-240mb/sec. The remainder is likely because they planned to use multiple chips in later products. But 2 Maxs will completely saturate it... so that's going to be a bottleneck if they try 4 Maxs. And there would be little point going above 4.

For comparison, the 3090 has almost 1TB of memory bandwidth.

Aren’t you assuming they won’t scale the memory bandwidth?

Yes, that's the way that works. To scale memory bandwidth, you would need a new chip. It wouldn't be an M1 Max anymore. If rumors of them using multiple m1 maxs is true, then they have 400mb/sec of memory bandwidth.

> they have 400mb/sec of memory bandwidth

Per chiplet.

If they improve the architecture at the same time, perhaps feasible?

I'll believe it when I see it.

Reply on news.ycombinator.com