Apple: M1 vs. M2

[Cmaier]

 

[theorist9]

So we have a 20% increase in GB MT, as compared with Apple's 18%, which means GB essentially matches Apple's figures.

And that we've gotten this with just a 12% increase in ST means the M2's efficiency cores are much improved over the M1's, as rumors predicted:

However, much of that 18% increase in MT performance might be due to significantly increased performance of the efficiency cores in the M2 vs the M1, in which case the increase in SC performance could be <18%.

 

[Colstan]

It's notable that from the first Geekbench result, the M2 is running at 3.49Ghz, compared to the M1's 3.2Ghz. There must have been additional overhead in the Avalanche P-cores for higher clocks, as @Cmaier predicted.

 

[Andropov]

Faster SC than Alder Lake's mobile chips, and just a smidge slower SC than the desktop version. On a passively cooled ultrabook. Wow.

 

[Cmaier]

It's notable that from the first Geekbench result, the M2 is running at 3.49Ghz, compared to the M1's 3.2Ghz. There must have been additional overhead in the Avalanche P-cores for higher clocks, as @Cmaier predicted.

Didn’t I say 3.5? I was wrong.

 

[Cmaier]

Faster SC than Alder Lake's mobile chips, and just a smidge slower SC than the desktop version. On a passively cooled ultrabook. Wow.

Imagine what they can do in the same power envelope after a node shrink.

 

[Colstan]

Imagine what they can do in the same power envelope after a node shrink.

At first glance, the M2 appeared to be a relatively minor update, but these new benchmarks suggest that it's an impressive upgrade, and Apple wasn't fudging the performance numbers. It does make me curious about the M3 generation, because rumors suggest that it will be a more substantial architectural overhaul, combined with a full node shrink. As tempting as M2 is, I'm still standing by my current plan to wait for M3. Until then, I'll still be using an Intel Mac...like a savage.

 

[Cmaier]

At first glance, the M2 appeared to be a relatively minor update, but these new benchmarks suggest that it's an impressive upgrade, and Apple wasn't fudging the performance numbers. It does make me curious about the M3 generation, because rumors suggest that it will be a more substantial architectural overhaul, combined with a full node shrink. As tempting as M2 is, I'm still standing by my current plan to wait for M3. Until then, I'll still be using an Intel Mac...like a savage.

You can’t have a “Apple is doomed because all the good engineers went to Nuvia” narrative unless you also pretend that M2 was nothing more than shortening a few wires and is really just M1+.

 

[casperes1996]

You can’t have a “Apple is doomed because all the good engineers went to Nuvia” narrative unless you also pretend that M2 was nothing more than shortening a few wires and is really just M1+.

I mean, it sounds reductionist and I'm sure it still took a lot of work, but is that not still kinda what happened? Clock speed bump and some extra cache. As far as I can tell, Avalanche and Blizzard (not considering the rest of the SoC) doesn't seem to be much more than that as far as I can tell. That can still require substantial re-architecting to achieve; No clue the effort required to go from Firestorm and Icestorm to Avalanche Blizzard, but the numbers seem to reflect "just" higher clocks and more cache, as far as I can tell. In particular for Icestorm->Blizzard.

I don't believe in the Apple is doomed because the engineers left thing, but you can also still believe that and just think that M2 and potentially more ahead was already well enough planned out that it hasn't started mattering yet but is still to come.

Just some devil advocacy

 

[casperes1996]

At first glance, the M2 appeared to be a relatively minor update, but these new benchmarks suggest that it's an impressive upgrade, and Apple wasn't fudging the performance numbers. It does make me curious about the M3 generation, because rumors suggest that it will be a more substantial architectural overhaul, combined with a full node shrink. As tempting as M2 is, I'm still standing by my current plan to wait for M3. Until then, I'll still be using an Intel Mac...like a savage.

I have an M1 Max and soon also an M1 (tomorrow), but I do and will still mostly live on Intel. Just because of the machines I have covering different cases. Intel iMac for most my work where I sit at the desk anyway and the 5K screen is nice, M1 Max laptop on the go where it shines and. soon an M1 Mini as an auxiliary device for server and testing duties - The Intel iMac still offers a great experience. And can double-duty as a bootcamp gaming machine with its Radeon Pro 5700XT. - In short; There's still value in a good Intel Mac too, even though the Apple Silicon devices are fantastic.

 

[Cmaier]

I mean, it sounds reductionist and I'm sure it still took a lot of work, but is that not still kinda what happened? Clock speed bump and some extra cache. As far as I can tell, Avalanche and Blizzard (not considering the rest of the SoC) doesn't seem to be much more than that as far as I can tell. That can still require substantial re-architecting to achieve; No clue the effort required to go from Firestorm and Icestorm to Avalanche Blizzard, but the numbers seem to reflect "just" higher clocks and more cache, as far as I can tell. In particular for Icestorm->Blizzard.

I don't believe in the Apple is doomed because the engineers left thing, but you can also still believe that and just think that M2 and potentially more ahead was already well enough planned out that it hasn't started mattering yet but is still to come.

Just some devil advocacy

I believe blizzard upped its IPC quite a bit. And I believe they redid the pipelines on avalanche entirely in order to be able to clock them higher. Both are clearly entirely new netlists and entirely new RTL. I’m not sure what people are expecting if “new cores with new physical design based on new logical design that allows 10% higher clock rate and which ups IPC on the efficiency cores” is not enough of a change for them.

 

[Andropov]

I mean, it sounds reductionist and I'm sure it still took a lot of work, but is that not still kinda what happened? Clock speed bump and some extra cache.

Possibly true for the high performance cores, but the efficiency cores must have gotten quite a bit more than a clock speed bump and some extra cache, I think? Anandtech reported a median +23% performance increase on the efficiency cores, while clock only went up by ~8%.

 

[Colstan]

After having the first Geekbench numbers for M2 CPU, we've now got GFXBench numbers for the M2 GPU.

In comparison to M1, the M2 GPU is...

Aztec Ruins Normal Tier: 45.6% faster
Aztec Ruins High Tier: 42% faster
1440P Manhattan 3.1.1: 43.7% faster
Car Chase: 32.7% faster
T-rex: 44.4% faster

 

[Cmaier]

After having the first Geekbench numbers for M2 CPU, we've now got GFXBench numbers for the M2 GPU.

In comparison to M1, the M2 GPU is...

Aztec Ruins Normal Tier: 45.6% faster
Aztec Ruins High Tier: 42% faster
1440P Manhattan 3.1.1: 43.7% faster
Car Chase: 32.7% faster
T-rex: 44.4% faster

But how does it do on chess benchmarks?

 

[Andropov]

After having the first Geekbench numbers for M2 CPU, we've now got GFXBench numbers for the M2 GPU.

In comparison to M1, the M2 GPU is...

Aztec Ruins Normal Tier: 45.6% faster
Aztec Ruins High Tier: 42% faster
1440P Manhattan 3.1.1: 43.7% faster
Car Chase: 32.7% faster
T-rex: 44.4% faster

That's even better than expected. One thing I'd always like to know with these kind of benchmarks is whether they're using TBDR-optimized rendering or if it's the same for all platforms.

 

[Colstan]

But how does it do on chess benchmarks?

I know this is going off on a tangent, but Chess has actually been diagnostically useful. Back when I had my first Intel Mac mini, I replaced the socketed Core Solo CPU with a Core 2 Duo. The original Core Solo was 32-bit only, while the C2D was 64-bit. However, there were almost zero applications compiled for 64-bit. Still, I wanted to make sure that I could run 64-bit binaries, and the only application that supported 64-bit was...the Chess program that came with OS X. I launched Chess and checked Activity Monitor to make sure the new CPU upgrade would default to running a 64-bit executable, and it was successful. That was the first time I've ever used Chess, it was also the last time. Still, I can conclusively state that, at one point in time, a Chess program actually had a useful function. I haven't been able to say that since then.

 

[casperes1996]

I believe blizzard upped its IPC quite a bit. And I believe they redid the pipelines on avalanche entirely in order to be able to clock them higher. Both are clearly entirely new netlists and entirely new RTL. I’m not sure what people are expecting if “new cores with new physical design based on new logical design that allows 10% higher clock rate and which ups IPC on the efficiency cores” is not enough of a change for them.

Right - I don't know where I heard it from, but I heard the Blizzard cores got a big clock improvement that with the cache bump would pretty much cover its improvement. I haven't honestly looked that much into them outside of the media "M2 Disappointing?!" stuff that was everywhere for a while.
That said I think the answer to what people wanted; Well really just comes down to the P cores. While the E cores serve an important role and can help with both efficiency and overall system performance, I think people focus mostly on the P cores when doing performance benchmarks and whatnot, especially the ones that focus heavily on single threaded numbers comparing that against Alder Lake and whatnot. And people want a single P core to beat a full turbo ~200W 12900K overclocked thermals be damned. Oh but also fit in a fanless MBA of course. That's the answer to what they want I think

Possibly true for the high performance cores, but the efficiency cores must have gotten quite a bit more than a clock speed bump and some extra cache, I think? Anandtech reported a median +23% performance increase on the efficiency cores, while clock only went up by ~8%.

See above - I thought we only had clock speed went up a lot more on Blizzard than Avalanche covering a lot of it

After having the first Geekbench numbers for M2 CPU, we've now got GFXBench numbers for the M2 GPU.

In comparison to M1, the M2 GPU is...

Aztec Ruins Normal Tier: 45.6% faster
Aztec Ruins High Tier: 42% faster
1440P Manhattan 3.1.1: 43.7% faster
Car Chase: 32.7% faster
T-rex: 44.4% faster

Do we know if that's 8v10 or 7v10 or?

I know this is going off on a tangent, but Chess has actually been diagnostically useful. Back when I had my first Intel Mac mini, I replaced the socketed Core Solo CPU with a Core 2 Duo. The original Core Solo was 32-bit only, while the C2D was 64-bit. However, there were almost zero applications compiled for 64-bit. Still, I wanted to make sure that I could run 64-bit binaries, and the only application that supported 64-bit was...the Chess program that came with OS X. I launched Chess and checked Activity Monitor to make sure the new CPU upgrade would default to running a 64-bit executable, and it was successful. That was the first time I've ever used Chess, it was also the last time. Still, I can conclusively state that, at one point in time, a Chess program actually had a useful function. I haven't been able to say that since then.

Ey, Chess can be useful for more than that too! It's all open source and can often demonstrate neat features of macOS updates like when Game Center first came to Mac, Chess had it and we could go see its open source implementation to see how it both used Game Center and even allowed Game Center to integrate with the custom email based multiplayer system too if the email address you played with was associated with a known account in Game Center to give them the right icon and all

 

[theorist9]

At first glance, the M2 appeared to be a relatively minor update, but these new benchmarks suggest that it's an impressive upgrade, and Apple wasn't fudging the performance numbers. It does make me curious about the M3 generation, because rumors suggest that it will be a more substantial architectural overhaul, combined with a full node shrink. As tempting as M2 is, I'm still standing by my current plan to wait for M3. Until then, I'll still be using an Intel Mac...like a savage.

Me too. My 2019 i9 iMac is fine for now, and I like to see a significant bump when spending the money for new gear, so I'll probably wait for the M3 as well.

And it will have to be real money. I used to do calculations that needed a lot of RAM, but more recently I haven't, and I've thus left the iMac at 32 GB RAM. But if my work changes and I need lot of RAM again, I can upgrade to 128 GB (for <$400). I won't have that flexibility with Apple Silicon. I'd have to pay for the high RAM without knowing if I'll need it or not, just in case. In addition, there are problems using the iMac as a display with either AirPlay or Luna, so I'd also need to buy a Studio Display.

And, finally, the program that gives me the longest runtimes (Mathematica) is still much slower on AS than Intel, even though they're on their second native AS build, perhaps because they haven't yet developed fast AS replacements for the Intel math libraries. Maybe by next year that will change.

 

[Colstan]

Do we know if that's 8v10 or 7v10 or?

I was wondering the same thing about the number of GPU cores, but it's not stated on GFXBench's results site, nor on GeekBench, assuming that it's the same device being tested. For now, we can just consider this a teaser benchmark, until release hardware gets properly reviewed. Even so, with only preliminary tests, the M2 looks overall promising compared to the M1.

 

[casperes1996]

Me too. My 2019 i9 iMac is fine for now, and I like to see a significant bump when spending the money for new gear, so I'll probably wait for the M3 as well.

And it will have to be real money. I used to do calculations that needed a lot of RAM, but more recently I haven't, and I've thus left the iMac at 32 GB RAM. But if my work changes and I need lot of RAM again, I can upgrade to 128 GB (for <$400). I won't have that flexibility with Apple Silicon. I'd have to pay for the high RAM without knowing if I'll need it or not, just in case. In addition, there are problems using the iMac as a display with either AirPlay or Luna, so I'd also need to buy a Studio Display.

And, finally, the program that gives me the longest runtimes (Mathematica) is still much slower on AS than Intel, even though they're on their second native AS build, perhaps because they haven't yet developed fast AS replacements for the Intel math libraries. Maybe by next year that will change.

Out of curiosity what kinds of calculations are you doing in Mathematica and if another tool could do the same thing faster is that viable as an alternative for your needs?