Apple M2 Pro & M2 Max Chip Analysis 

I'd like to see analysis of cache subsystem. Because AMD Zen 4 uses 32 kB instruction L1$ ....... while Apple M1/M2 uses monstrous 192 kB L1i$ ..... which is 6x fold more!!! Intel went from 32 kB to 48 kB ..... so 1.5x fold more than AMD ....... but still Apple has 4x fold more. It looks like Apple discovered at micro-architecture level something nobody else did.

@@richard.20000 I read somewhere and it says that they have such large L1$ because they chose to allocate such large area for the cache

@@tanthokg There must be reason why x86 is stuck at 32+32 kB L1 cache for 20 years (since 2003 Pentium-M Banias) while Apple multiplied size of L1 to record breaking size in just few years. Apple M2 has ........ 54% higher IPC than Zen 4 and Raptor lake (54% higher performance at same GHz in Geekbench 6). That's why Apple M2 has similar performance at 3.5 GHz like x86 at 5.5 GHz. Maybe because M2 has 6x ALU units .... Zen 4 only 4x ALU. Maybe because M2 has 2x Branch units ... Zen 4 only 1x Branch Unit Maybe because M2 can execute 6+2= 8 instructions per clock .... Zen 4 only 4+1= 5 instr / clock. Have you heard about Zen 5 rumors having huge 30% IPC increase? Well, this still wouldn't be enough to beat an old 2022 Apple M2. And M3 will be released this fall so Zen 5 will need to fight M3 and M4. And cheap license ARM core Cortex X4 has 33% higher IPC than Zen 4 and Raptor Lake. X4 also has record breaking architecture: it has 8xALU units + 3x Branch units = 11 instr / cycle. That's world record even the L1 cache is just 64 + 64 kB.

@@richard.20000 on paper..but actual real performance...M2 still lag behind intel aderlake

@@D0x1511af Same way as wide AMD K7 Athlon XP 2100+ at 1.53 GHz lagged behind high speed demon Pentium 4 at 2.4 GHz. We all know from history that high speed demon is the wrong way and Intel lost so much market in favor of AMD. Intel finaly abandoned P4 in favor of wider and lower clocked Core2Duo (even Apple switched from PowerPC to Core2Duo at that time). This time all x86 CPUs are just as high speed power-hogs as P4 was (Intel and AMD both consume >200W which is crazy). The right philosophy like Core2Duo is Apple CPU today. M2 might be lagging small percentage in overall performance however it is multiple times better in everything else (56% higher IPC than Zen 4 means at least 5 years advantage in architecture, also power consumption has M2 Ultra CPU around 80W). Do Intel or AMD have last level cache shared with GPU? No. Apple has L3 cache shared with GPU and NPU. ARM Cortex X4 has L4 cache also shared for all SoC processors. And it goes on and on.... Intel and AMD are dinosaurs who cannot keep the pace with ARM. They employs hundreds of engineers to invent sophisticated predictors to be able mitigate that horrible CISC x86 variable length encoding (while ARM do not spend single minute or single transistor on that because ARM is modern RISC by design - every instruction has fixed length 4 bytes). RISC was invented after CISC for a reason (1984 Intel's RISC i860 was designed after 1978 x86, same for AMD their first inhouse CPU was Am29000 based on RISC, AMD glued x86 decoder and 1st AMD's inhouse x86 was born - K5). RISC is better. Every new ISA in last few years is RISC based (ESP32) or even Chinese 64-bit Loongson is RISC. When CPU does not need run as much legacy SW as x86 does, then you always choose superior RISC. It has more performance and less power consumption at the same time (thanks to modern ISA w/o old garbage). And emulate x86 SW if you really need. That' exactly what Apple does now.

He is aiming for high yield instead of high volume

Ya may be he should open a website or a telegram group where he share with us stuff for futher analysis for his curios subscribers

@@rishavpapaji5349 That's what patreon is for ;)

Thanks! BTW: love your name ;)

i think the whole point of the e-cores in the first place was to be able to have an efficient cpu core for use during standby. x86 cpus are notoriously awful at that sort of thing and even intel's e-cores haven't solved that because that's not what they were meant for.

9:20 I think he said "N5P", unless he misspoke somewhere else in the video.

@@NootNoot. he corrected himself later in the video. But I hadn’t gotten to that part yet.

@@chidorirasenganz Haha, all good

I think I say „5 nanometer“ and „N5P“ at different points in the video. Both is correct (if you disregard that it’s not actually 5nm at scale), but as you mentioned, N5P is the specific process used, so technically more correct ;)

I was looking for the core #20, which would have been labeled as "core 19". There are only 19 physical cores, core0 to core18.

Good question! I think you are most likely correct with 6P + 4E, since 8P + 2E would be very similar in performance and the P-cores are much larger and thus more prone to silicon defects.

I think Apple is in their own AAC+HEVC camp and would like to remain there too. Maybe AAC is the reason they don't want to give any ground to AV1 (with Opus as audio codec?)

@@VADemon Apple is part of the Alliance for Open Media, the group who created AV1. So it's not a "give ground to" thing

@@kirby0louise Maybe not, but in a similar issue Apple added WEBM (VP9) support *looks up* only 2 years ago.

They could probably use the NPU to accelerate ray-tracing

@@meru_lpz I doubt it. All the NPU does is accelerate matrix math. If RT acceleration could be done that way, then Nvidia/AMD/Intel wouldn't even bother with RT cores, just shaders and tensor cores would suffice.

No, I dont think so, both should have similar ST performance.

Maybe the increase use of 3D in different creative industries? And maybe the increasing gpu rendering, that would be my guess

The pictures or die-shots show the silicon structures on the chip and while we can't see through the chip, some structures are very easy to recognize, including the GPU cores. I've made another mock-up for you, I think it should be very clear in this picture: i.imgur.com/KASpdAh.png Can you see all the other GPU cores? And the difference to the silicon structures where we would expect core #20? The GPU area is bright compared to the rest of the M2 Pro chip, GPU misc & interconnect areas are grey scale and I have marked one complete GPU core + the area which is not a GPU core.

I honestly can't imagine that at all. I think its mostly space efficiency.

both intel and AMD have already taken performance crown in laptop chips in some cases even more than double the performance of M1 pro. what they(intel and AMD) are now trying to do is make gains in performance per watt scenario. they've come close especially AMD 7000 laptop chips and Intel ratptor lake at 45 watts. but still Apple's chips have that performance per watt advantage. Apple knows that from this point onwards they cannot compete with both AMD and Intel in RAW performance so this time around in their Presentation they didn't compare it with either of those companies' latest chips rather only claimed Performance per watt. In another Graph Apple only compared their chips' RAW performance with a 5 year old intel Chip.

No AV1 support afaik, HEVC (H.265) is supported tho.

@High Yield Yeah would love it even on a software level for exporting for youtube. Funnily, Apple is one of the founders behind AV1.🤷‍♂️

A die-shot pf M2 Max is in the video, but as mentioned, Apple most likely photo shopped it to remove the D2D interconnect. I'm also curious about the rumored M2 Extreme. How do you think connecting four M2 Max would work? Like a daisy chain? Latest rumors say it got cancelled.

@@HighYield exactly. Apple's die shot is edited. Yes it would work like a jigsaw if the M2 Max has two D2D interconnects. M1 Max D2D interconnect was discovered before apple announced it ever had one. It was discovered by real die shots as soon as the M1 Max was released.

how is it handling the ML workloads?

@@seanwfindley for createml it's running well. For tensorflow-metal it's utilizing the GPU nicely. Wish ANE was used for training and not just when you run coreml models.

@@seanwfindley I wish there was a good tool to really measure GPU scheduling and utilization. That way I can gather data when I benchmark my macs. I'm sure apple has some in-house tools, but I don't think they'll ever give it to the public :(

@@woolfel Would you say the new macbook pro could be useful for pricing models or predicting shortages, some based on news/text analytics?

@@seanwfindley if you're just doing pricing models, I don't believe you need to use Tensoflow or pyTorch. Regular regression or stats software like Mathematica is "probably" sufficient. If you need a lot of fast memory to analyze data and use GPU, a maxed out M2Max might be a fit. You'd have to try it. If there's anything I've learned in 20 years of programming, nothing beats a real test to validate whether my assumes are correct or complete garbage :) I know neural networks is the HOT thing, but there's plenty of old boring techniques that get the job done.

Games apple...

Then you are a exception, since most ppl are not stressing the NPU that much. I’ve never heard of “OCR”, can you tell me what it is?

@@HighYield OCR is Optical Character Recognition. I'm using Apple's utilities in Apple's Shortcut App to read words on the screen. Yes, maybe I am an exception, but all I'm doing is running Apple's software.

You can't. The RAM is ball soldered directly onto the SoC and the storage isn't conventional M.2 storage but rather discrete flash chips controlled by the SoC's storage controller. You have to spec out what you need when you purchase. There are a couple of experimental attempts at aftermarket upgrades but they require specialized tooling.

@@Longlius sounds quite bad I ended up going for something that was on stock let's see how long it holds up.

@@andraslibal I mean depending on workload, you'll probably be fine. MacOS is quite efficient with Apple's own hardware. I've yet to find anything that my M2 Max/32 GB/1 TB really struggles to do.

@@Longlius I do scientific simulations and I wrote my own code for them it is a fine laptop for it of course the heavier workload is on desktops and computational servers. It is also fine for presentations etc. I just wonder how many years before I need to upgrade I think the cycle is getting shorter.

Since the 3nm delay scrambled Apples timeline, we might even get M3 before the A17 this time. But ofc this is pure speculation on my side.

@@HighYield I think they won't release M3 before A17 if it is based on it, because they would need all the 3nm silicon available for their iPhones which is their biggest business. However if M3 is based on A16 then they might release it before.

Apple will find a way to avoid increase in battery life and faq up thermals anyway.

@veled veled i don't think so

They don't have to have unused silicon parts to bin their chips? If the fully enabled die has 19 cores they can bin 3 cores and make a one tier down SKU with 16 cores (which is what they do, there's a config of the M2 Pro with 10 CPU cores instead of 12 and 16 GPU cores instead of 19). Likewise with M2 Max, while they don't bin down the CPU on that there's 30-core GPU option and 38 at the higher end.

@@utubekullanicisi Oh didn't know they had these, thanks

Maybe because Apple likes money? :p

The M2 Max goes up to 96GB RAM and 8TB storage. But of course at Apple prices.

No, you need to satisfy your porn addiction. Anything with internet access will do.

I should have used € instead, missed opportunity!