Apple's M1 is up to 3.6x faster at training machine learning models | Hacker News Reader

Apple's M1 is up to 3.6x faster at training machine learning models

blog.roboflow.com

17 pointsyeldarb8 comments

Title is very misleading. M1 is faster than a macbook pro without discrete graphics, slower than a macbook pro with discrete graphics. But who is training models on their laptops? Any real (i.e. non-toy) machine learning model requires a desktop and Nvidia GPU, which is several orders of magnitude more powerful.

If you can't afford it (but can afford the newest Macbook Pros?), Google Colab is free. Or use other cloud based GPU providers.

I don't know what Roboflow is, but I wouldn't trust claims they have on their website given the deceptive headline.

I like using my laptop with a 2060 RTX, and it's a lot of fun, as I can travel with it and use it without good internet. I'm not sure how much I'm missing out on though (and probably it would beat the M1).

(Disclaimer: I was involved in the prep for this article so I'm probably biased but) I think the comparison between the two 13" MBPs is the most like-for-like. They're similarly priced and outfitted.

I'd expect the future 16" MBP with Apple Silicon to be more of a like-for-like competitor to the current 16".

All this fuss about a processor when no serious computations get done with a processor.

Since the early 2010s, people stopped buying high end processors in favor of high end gpus.

I genuinely wonder if Intel has stopped trying to Innovate because it wasn't being rewarded.

I'm not sure this result is that interesting. First, there's the assumption that Apple bothered optimizing their proprietary ML library for Intel as much as they did for their own M1 chip:

> Presumably, Apple has already done the work in optimizing CreateML for both for their Intel and M1 devices.

Second, the M1 is faster than the Intel Macbook with an integrated GPU. But the article also notes this, which makes the comparison apples and oranges IMO:

> Notably, training was unable to take advantage of the integrated Intel Iris graphics card to accelerate training but it was able to partially utilize the integrated graphics card on the M1.

Third, the M1 lags the Intel with the discrete Radeon GPU, where the GPU is being utilized. But this is not exactly a surprising result. Nor is it a comparison against the state of the art. What matters more practically is a comparison between an M1 Macbook Pro and a leading competing laptop with a discrete nvidia GPU.

I think you’re right from a technical point of view. Sure, it might be slower than a dedicated GPU, and sure the Intel CPUs in the MacBooks are not the fastest in the world in the first place.

However, the comparison still has value from a practical point of view: the new thing is more than 3 times as fast as the old thing. This matters to the people who use the things more than to people who just want to compare abstract chip performance.

I assume we’ll see other data points, more relevant to some, when high-end ARM MacBook Pros are released in a couple of months.

It's incredibly deceptive headline. But then again I'd expect people doing ML to know better.

But given Apple's marketing reputation, it's all about appealing to customers that have no idea what this means. They won't know, but it's sounds good.

> What matters more practically is a comparison between an M1 Macbook Pro and a leading competing laptop with a discrete nvidia GPU.

100% agree. We would have liked to do that comparison but, as mentioned in the article, the required software is not yet supported on the M1.

Reply on news.ycombinator.com