Hacker News Reader

Comment by pram | Hacker News Reader

There were comparisons to the M1 and M2 in the presentation, and indeed the linked website.

The M3 chips are 30% faster than the M2 chips for efficiency cores and 15% faster for performance cores. The overall performance is still impressive and there is no alternative to M*s when compared with performance per watt. But others have caught up, for example https://www.amd.com/en/products/apu/amd-ryzen-9-7940hs is very good.

> But others have caught up

You mean others now got to use TSMC's 4nm process? These new M3 chips are probably on the 3nm process, Apple is still a generation ahead here.

It looks like Apples aim is to always stay a generation ahead of its competition, I wonder how long they can keep that up since they aren't running their own fabs.

They are literally on the 3nm process. They say so in the video.

Ok, so no probably, they are on 3nm. Anyway my point that they are a generation ahead still stands.

Apple signed a deal with TSMC to purchase nearly all of their available 3nm chips for the next year. Ethical or not, they positioned themselves to ensure almost no competitors could develop on 3nm until they did first since no fab on earth has the scale of TSMC. They could ride this plan for years if it pays them.

How is Apple buying all the production slots for a process in any way unethical? It's not like they're buying them and then burying the chips in a landfill. They paid TSMC's asking price for production slots. AMD or Intel could have bought those same slots but didn't. TSMC has limited capacity at 3nm, it was up for sale, Apple bought it. Where's the ethics question?

I’ve seen argument for ‘business ethics’ before from the losing side. It’s sometimes part of a media campaign. It’s likely cheaper than legal action.

>How is Apple buying all the production slots for a process in any way unethical?

This has been the case on HN for more than 5 years. Intel Fabs used to sell their industry best node to only Intel themselves, and charges a premium for those newer CPU. I guess that is unethical too.

How is Google buying search defaults on all iPhones in any way unethical? It's not like they're buying them and then burying the searches in a landfill. They paid Apple's asking price for search defaults. Microsoft or Brave could have bought those same search defaults but didn't. Apple has a limited amount of search defaults on iPhones, it was up for sale, Google bought it. Where's the ethics question?

not only is outbidding the competition not unethical (as a sibling notes), apple actually is very involved in the early node work etc. a lot of this work is literally done for apple, it is "exclusive games" in the "this game would not have been made without the sponsorship" sense. this literally would not have been brought to market on the same timelines if Tim Apple wasn't signing a couple billion dollars a year to TSMC right upfront.

Apple pays lavishly to support TSMC's early node research, and they get their say in what happens in the R&D process, and very early insight into the node and their say on how it would work for them as they do their rollout. TSMC gets carried through the research phases much faster than their competitors can do, and it's led them to be on an absolute tear starting with 7nm. And they absolutely cannot fill the same level of demand with the same level of R&D funding from any of their competitors.

It's been a healthy, productive long-term partnership, TSMC is maybe the only supplier Apple can't boss around and Apple is certainly a client that is always too big to fire. Doesn't mean every apple product is good (and TSMC can still flub, and their competitors are catching up a little bit) but Apple can move whatever they need to lol, they are masters of supply chain managment. They can cover TSMC's mistakes if needed, and they have insight into exactly what is happening as the node is developed and how they need to maneuver their product stack around to exploit it.

Engineers study designs, CEOs study logistics. Also true of NVIDIA btw lol, they are very logistics-oriented because they make up such a large marketshare. How many companies on the planet are ordering big bulk runs of GDDR? Well, if we are ordering 20% of the planet's GDDR on a fixed timetable then maybe we can get a custom version, micron, right? (9 months later, GDDR5X/6X is born lol)

It is an interesting contrast to Intel - this is almost the same kind of synergistic relationship as intel's own fab and IP side have historically had together. Did intel fail because they had a tight fab-design coupling, or did they fail because they had a rotted internal culture and then the fab slipped a bunch?

Not running their own fabs makes it easier, right? They presumably pay a premium for being the first to use a generation, but not anywhere near as much as the full cost of developing it, since TSMC can sell that capacity to everybody else when they are done with it.

Oryon seems to have caught up in performance despite being on N4 (which is just 5nm++), but it's an ARM design from the guys who made M1 in the first place.

M3 seems like it generally underwhelms. A17 had like a 3% increase in IPC. They didn't discuss battery much and I suspect that's because ramping the clockspeeds to over 4GHz isn't so good for the battery benchmarks.

The most worrying part is the transistor count. M3 Max gets quite a bit higher transistor count, but M3 is only 37B transistors while M1 was 33.7B. Apple and/or N3 absolutely suck here.

Looks like I might be keeping my M1 system for yet another upgrade cycle.

the backstory here is that TSMC N3 is a trainwreck, it's now separated into two different nodes, N3B (for bad) and N3E (enhanced). N3E gets the promised step that was originally for N3, but it only enters volume production next year. Supposedly it will actually bring costs down (and yields up) because this is where some additional steps go EUV. Both TSMC and Samsung have been fucking around with their marketing around nodes to try and say they're first in volume 4nm production, but both are having problems with the final bosses of FINFET at 3nm and after this both TSMC and Samsung do GAAFET and solve a different set of problems. Past that lies... nothing. Hyper-NA seems dead.

In the meantime, M3 is on N3B despite very low yields, which surely applied design pressure to keep size down, and the power gains are not as good, and also the density is worse than promised. Apple also surely feels pressure to keep prices high (bait-tier M3 base option with 8GB lol) and honestly they probably are going to be tough to justify on a performance/efficiency basis compared to very fierce ARMv8 competition (we are now testing the thesis there's no difference lol). Apple still has advantages but man do they take you to the cleaners for the result, a loaded apple laptop is obscene. I chose to go for an older loaded M1 Max instead of waiting for M3, because I could actually get a nice laptop that wouldn't impose limits on a prosumer etc. 1TB is all anyone can afford still and that's really silly.

(SSD prices in particular are absolutely inexcusable lol. Mandate a M.2 NVMe 3.0/4.0/+ port please, EU, it's time. Don't care how it works, slot it into the side or whatever if you want, it can be single-sided 2230 or 2242 if you want (or caddy-loading, the icy dock standard lol), but it's time.)

https://global.icydock.com/products-c5-s48-i0.html

https://global.icydock.com/vancheerfile/images/mb873mp-b_v2/...

I also wonder if losing a bunch of the PA Semi team to whatever startup (it may have been nuvia or tenstorrent lol) may also have hurt apple's velocity on A16/A17. There were a lot of apple silicons before M1, after all. But certainly TSMC is a bunch of the problem here.

I think they'll hustle to refresh it and do M3+ as a fast follow in 6-12 months with N3E, the cost economics are very favorable to jump as soon as there's the volume. That doesn't mean MBP gets refreshed immediately though, they'll ramp on the phones (iphone 16/16 pro, etc) and base-tier M3+ or whatever first.

>the backstory here is that TSMC N3 is a trainwreck, it's now separated into two different nodes, N3B (for bad) and N3E (enhanced).

Without N3B, there will never be N3E.

It was the same with N7. And none of these are new to a new node generation.

> But others have caught up

Then you proceed to link to one that... hasn't? (it's good, yes, but it's not caught up at all)

I note that if those performance numbers are correct it'll still be a lot slower than Intel's Core i9-13980HX laptop CPU. The M2 Max was between 50% and 80% of the speed of the 13980HX on most benchmarks. A 15%-30% uplift will get it closer to the Intel part but still not reach it.

Good point.

However, when they’re that close in performance, I think power usage should be taken into account. It’s a laptop CPU after all. Maybe I’d rather have 80% of the performance at 4x the battery life

Did you account for the M3 Max having four extra performance cores (12 versus 8 for the M2 Max)?

Yeah, but the problem is that 13980HX has 24 cores, 32 threads and incredible 157W Turbo Power. That means it will be really slow when on battery. And when connected to the mains, it'll be as loud as an airplane taking off.

The perf per watt claims still apply.

Reply on news.ycombinator.com