Live data from Hacker News

First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

extremetech.com

41–50 of 83 posts

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#41

I wonder what kind of usage metrics Apple has that led it to reduce (8 down to 6) the number of performance cores and increase (4 up to 6) the number of efficiency cores on the M3 Pro? Or the reduction (25%) in memory bandwidth? Maybe they're trying to get people to buy the Max version? The spec differences are interesting: M3 Pro, 6 efficiency, 6 performance, 18 GPU, 16 Neural cores, 150GB/s memory bandwidth M3 Max…

Reducing memory bandwidth is particularly bad given that most AI work loads are memory bound.

With the memory bandwidth Apple's fully integrated GPU with shared memory is actually pretty good for smaller AI stuff like running and fine tuning llama2-sized models. It gives you a decently fast GPU option with a lot of RAM. NVIDIA cards with similar RAM are a decent fraction of the price of a whole Pro or Max M chip laptop.

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#42

It was a deliberate choice to differentiate between Pro and Max There's a 45% performance uplift between M2 Max and M3 Max in benchmarks. You still get more performance between M2 Pro and M3 Pro. And you get significantly more if you're upgrading from M1 Pro.

Agreed.

Also, I see the discussion about pro vs max being focused on performance, but what about thermals or battery life? Perhaps many professional users would be better off using pro and not max?

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#43

I wonder what kind of usage metrics Apple has that led it to reduce (8 down to 6) the number of performance cores and increase (4 up to 6) the number of efficiency cores on the M3 Pro? Or the reduction (25%) in memory bandwidth? Maybe they're trying to get people to buy the Max version? The spec differences are interesting: M3 Pro, 6 efficiency, 6 performance, 18 GPU, 16 Neural cores, 150GB/s memory bandwidth M3 Max…

Way outside of my field of knowledge but would it be around prioritizing battery life above all else on the pro? Would also nudge people who _need_ the higher power features to not stick with Pro too, obviously.

That confuses me because shouldn't those people just get an M3? and then those who get the M3 Pro want better performance. I don't understand the 'pro' naming scheme means if they're not serious about making it pro.

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#44
post #36

"The first Apple M3 Pro chip benchmarks have leaked! It’s actually slower than the M2 Max based on margin of error. M3 Pro: 15,173 M2 Max: 15,242" An alternate, and in my opinion more accurate and relevant, way of phrasing this would be: "Very cool! The new *mid-level* chips perform as well (within In what world is this not a good thing? Pretty solid annual incremental upgrade. Of course I understand why people are u…

People are picking on that lede tweet somewhat unfairly. Numbers in the rest of the article say that the same benchmark run between the two "pro" variants improves only 6%. And that's actually quite disappointing for a chip that's supposed to be a on a new semiconductor node. Not a lot of people make a laptop purchase decision over 6%. > Look at the annual incremental performance increases seen from AMD and Intel and…

> (10nm++/Intel7, the name changed over the period for marketing reasons)

People like to pick on this change a lot but these names have been marketing for decades, and other fabs (including TSMC and GloFo) named nodes more advanced than they were purely to look better (iirc - GF 14nm was much closer to Intel's 22nm than to 14nm, Samsung 8nm performed way worse than TSMC N7 etc.) - meanwhile "Intel 10ESF" aka "Intel 7" is actually very similar in most metrics to TSMC N7, though is more amenable to higher clocks.

If anyone wants to loose braincells over this, Wikipedia editors are (still) having an epicly long debate over this since the name was changed: https://en.wikipedia.org/wiki/Talk:7_nm_process#Intel_7

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#45
post #12

Live by benchmark results and slick performance marketing, die by benchmark results and slick performance marketing. The market is ruthless; if you aren't delivering more of what you promised made you great last year, you're failing. Really the underlying news here, which the article doesn't really treat with, is that TSMC's 3nm node is sort of a bust. It's not looking like a full node upgrade at all (something that'…

I'd say gains from M2 Max to M3 Max are pretty reasonable. Of course same might not apply to M2 Pro to M3 Pro.

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#46
post #41

I wonder what kind of usage metrics Apple has that led it to reduce (8 down to 6) the number of performance cores and increase (4 up to 6) the number of efficiency cores on the M3 Pro? Or the reduction (25%) in memory bandwidth? Maybe they're trying to get people to buy the Max version? The spec differences are interesting: M3 Pro, 6 efficiency, 6 performance, 18 GPU, 16 Neural cores, 150GB/s memory bandwidth M3 Max…

Reducing memory bandwidth is particularly bad given that most AI work loads are memory bound. With the memory bandwidth Apple's fully integrated GPU with shared memory is actually pretty good for smaller AI stuff like running and fine tuning llama2-sized models. It gives you a decently fast GPU option with a lot of RAM. NVIDIA cards with similar RAM are a decent fraction of the price of a whole Pro or Max M chip lapt…

> most AI work loads are memory bound.

Is that the reason why the number of neural cores stays the same, do you think?

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#47

"The first Apple M3 Pro chip benchmarks have leaked! It’s actually slower than the M2 Max based on margin of error. M3 Pro: 15,173 M2 Max: 15,242" An alternate, and in my opinion more accurate and relevant, way of phrasing this would be: "Very cool! The new *mid-level* chips perform as well (within In what world is this not a good thing? Pretty solid annual incremental upgrade. Of course I understand why people are u…

The difference between the M2 Pro and M2 Max was quite subtle, especially on the CPU department. Same exact CPU cores except the Pro had a binned variant with two P cores disabled. The biggest difference between the two was on the GPU front. If they're testing the CPU, then the M2 Pro and M2 Max should have similar performance (assuming non-binned M2 Pro).

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#48
One of the interesting things I've noticed is that the transistor count for the M3 Pro hasn't gone up much.

For example, the regular went from M1 16B, M2 20B, M3 25B; The Max went from M1 57B, M2 67B, M3 92B. On the other hand, the Pro went from M1 33.7B, M2 40B, M3 37B. For the regular, transistor count is up 56% over the M1 and for the Max it's up 61%. However, for the Pro it's only up 10% over the M1 (and slightly down from the M2).

In terms of the cores, the M1 Pro was either 6+2 or 8+2 Performance+Efficiency cores. The M2 Pro became 6+4 or 8+4 so you still had 6 or 8 performance cores. The M3 Pro is either 5+6 or 6+6 so fewer performance cores.

It just seems like the regular has kept the same 4+4 and gone up a bunch in transistors and the Max has gone from 8+2 to 8+4 to 10+4 or 12+4 (a decent step up in cores and transistors). The Pro, on the other hand, hasn't been given the same treatment. It's still a great processor, but it feels like it isn't quite the value that it was. The M1 Pro offered such huge performance with up to 8 performance cores. The M3 Pro feels like a bit of a weak spot in Apple's lineup where it'd be worth deciding between the M3 Max or the regular M3 depending on your needs.

Still, I love Apple Silicon. It's amazing having a laptop that just feels so good to use that I don't think about raw performance numbers. Things just work without pain or heat/fans. I know that's hard to quantify, but my M1 MacBook Pro has felt so different from all the computers I've used before. I was using a 28W Intel 2020 MacBook Pro and my M1 MacBook Pro just felt completely different in terms of performance. I guess part of that is the raw performance with an M1 getting single-core performance that is 85% faster and multi-core performance that's 2.8x faster, but it's also that Apple did it at power/thermal level that means I never hear a fan and my laptop never feels warm. Maybe a big part of it is just that Apple Silicon was such a huge improvement that I'm still in awe of it. Maybe part of it is that I haven't seen what Intel has done since 2020 (and they've been catching up after years of mediocrity). Still, for the first time I feel like I'm waiting on non-CPU stuff for a reason to upgrade (like an OLED or microLED display or something).

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#49

"The first Apple M3 Pro chip benchmarks have leaked! It’s actually slower than the M2 Max based on margin of error. M3 Pro: 15,173 M2 Max: 15,242" An alternate, and in my opinion more accurate and relevant, way of phrasing this would be: "Very cool! The new *mid-level* chips perform as well (within In what world is this not a good thing? Pretty solid annual incremental upgrade. Of course I understand why people are u…

Yet another way of phrasing:

The six performance cores in the M3 Pro are on par with the eight performance cores in the M2 Max.

That’s a per-core performance increase of 1/(6/8)=+33% (ignoring the efficiency cores because I have no idea how they contribute to performance).

Re: First Apple M3 Pro Benchmarks Show Surprisingly Low Multi-Core Gains over M2 Pro

#50
post #36

"The first Apple M3 Pro chip benchmarks have leaked! It’s actually slower than the M2 Max based on margin of error. M3 Pro: 15,173 M2 Max: 15,242" An alternate, and in my opinion more accurate and relevant, way of phrasing this would be: "Very cool! The new *mid-level* chips perform as well (within In what world is this not a good thing? Pretty solid annual incremental upgrade. Of course I understand why people are u…

People are picking on that lede tweet somewhat unfairly. Numbers in the rest of the article say that the same benchmark run between the two "pro" variants improves only 6%. And that's actually quite disappointing for a chip that's supposed to be a on a new semiconductor node. Not a lot of people make a laptop purchase decision over 6%. > Look at the annual incremental performance increases seen from AMD and Intel and…

> People are picking on that lede tweet somewhat unfairly. Numbers in the rest of the article say that the same benchmark run between the two "pro" variants improves only 6%. And that's actually quite disappointing for a chip that's supposed to be a on a new semiconductor node. Not a lot of people make a laptop purchase decision over 6%.

The M3 Pro has been neutered - the normal M2 Pro* was 8 Performance + 4 Efficiency cores (same as the Max) whereas the M3 Pro is just 6 P + 6 E cores.

If you want the full complement of CPU cores on the M3, you have to get the Max variant.

*There was a special 'low' end M2 pro that only had 6 + 4 cores.

Post reply on HN