Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
[CPU only] Apple M1/2(Max/Ultra) (TSMC 5nm) vs AMD Zen 4 (TSMC 5nm) - Technical Analysis
- Thread starter exoticSpice
- Start date
- Sort by reaction score
I can't find a test replicating 30k in ECO 65, but PCWorld did run the 7950x in ECO 65 and got 28,655, which is really impressive (it should be noted that max draw at ECO 65 is 88w, versus ~70 on the M1 Ultra).
These are new cores on an enhanced version of 5nm compared to the 2 year old cores and base 5nm of the M1 Ultra, but extrapolating from the M1 -> M2 multicore gains (~14%) and adding the 2 extra e-cores rumored pretty heavily for the M2 Pro etc., would suggest an M2 ultra at about 28,300 (23841*1.14+1100 (550 per e core, roughly). If the Ultra has 8 e cores (the rumored 4 from M2 Max, doubled) that number may be more like 29,400. If that's close to correct then it's very interesting how similar the results from AMD and Apple may look. Cinebench may be a poorly optimized test for Apple Silicon of course. In Geekbench the M1 Ultra and the full-fat 230w 7950x are essentially tied in multicore.
That said, AMD and Intel despite the pandemic have been putting out new processor generations on a yearly cadence recently. I strongly suspect an M2 Ultra will be about on par with the 7950x in multicore at similar power consumption, but at the moment it's only a suspicion because the chip simply doesn't exist yet. I'm really hoping there's not typically going to be 18 mos. between each M series generation going forward, as there was between M1 and M2.
EDIT - Interesting, as a side note, that in all these tests the 7600X does so poorly in comparison. At full power (TDP 105, actual consumption closer to 144) the 7600X is only getting just above 15k in Cinebench R23 Multi. That's 8k less than the Ultra at over twice the power draw?
Last edited:
Love GN reviews. He pulled no punches on this release. Not sure why AMD chose to boost to tjmax, and the power draw when all cores active is atrocious, though it is fast.
Apple has been employing the same strategy (or playing the same trick) for over a decade now. Most users aren't aware of it perhaps.
I think AMD will face an uphill battle on consumer education as most ppl are "trained" for decades to believe higher temperature is "bad, poor design or lower quality."
It's perfectly fine as long as the chip operates <= TjMax.
A valid point or perhaps not. A common misnomer for knowledgeable Apple users here is to compare OEM systems (e.g. Apple computers) against DIY PCs.
DIY PC market users are empowered with many tunnables (and still tonnes of other tunnables hidden) through BIOS or other means. You can push the power to the limit if you want or you don't have to if you're wiser.
For the review I saw, Zen4 is very efficient. For example, the flagship 7950x, tune down the power cap to 60% (around 140W) of the stock limit, you retain 95% of the performance.
I was going based on GN's review. Others chose different apps to do a power comparison on and got lower numbers.
GN said that they worked non-stop to get these reviews out so I assume that they will go back and run other benchmarks over the next week or two. I guess it was a sprint for the review channels and sites.
Could someone give me the TL;DR of the following stats of M1 vs. M2?
Energy consumption, idle:
Energy consumption, ramped up:
Single Thread improvement:
Multithread improvement:
GPU Improvement:
Energy consumption, idle:
Energy consumption, ramped up:
Single Thread improvement:
Multithread improvement:
GPU Improvement:
Marketing.
According to Anandtech:
AMD Zen 4 Ryzen 9 7950X and Ryzen 5 7600X Review: Retaking The High-End
Insane score for the 65W 7950X.Marketing.
According to Anandtech:
View attachment 2080965
AMD Zen 4 Ryzen 9 7950X and Ryzen 5 7600X Review: Retaking The High-End
www.anandtech.com
Zen 4 is a very power efficient CPU architecture.
Beats everything from Intel (105W-125W) at half the power limit.
Too bad 7950x ECO 65 mode, the sweet spot for single and multi-core performance efficiency, and ECO 105 are minimally covered on day 1. 7950x manages to destroy both M1 Ultra and Intel Alder Lake 12900k in single and multi-core performance. Now looking forward to Nvidia RTX4000 vs AMD RX7000 dGPU comparisons.
7950x ECO 65 170W total system power consumption
Cinebench R23 28,655
168.6 points per watt
75% performance at 52% total system power consumption vs Plaid
105% performance at 47% total system power consumption vs Intel 12900k
7950x ECO 105 240W total system power consumption
Cinebench R23 34,300
142.9 points per watt
90% performance at 73% total system power consumption vs Plaid
126% performance at 67% total system power consumption vs Intel 12900k
7950x Plaid 330W total system power consumption
Cinebench R23 37,973
115.0 points per watt
139% performance at 92% total system power consumption vs Intel 12900k
Intel 12900k 360W total system power consumption
Cinebench R23 27,283
75.8 points per watt
7950x ECO 65 170W total system power consumption
Cinebench R23 28,655
168.6 points per watt
75% performance at 52% total system power consumption vs Plaid
105% performance at 47% total system power consumption vs Intel 12900k
7950x ECO 105 240W total system power consumption
Cinebench R23 34,300
142.9 points per watt
90% performance at 73% total system power consumption vs Plaid
126% performance at 67% total system power consumption vs Intel 12900k
7950x Plaid 330W total system power consumption
Cinebench R23 37,973
115.0 points per watt
139% performance at 92% total system power consumption vs Intel 12900k
Intel 12900k 360W total system power consumption
Cinebench R23 27,283
75.8 points per watt
Last edited:
Marketing.
According to Anandtech:
View attachment 2080965
AMD Zen 4 Ryzen 9 7950X and Ryzen 5 7600X Review: Retaking The High-End
Performance seems to be a big deal in the hobbyist world as I think that most of these are sold to builders and not OEMs. I think that performance used to be the primary metric but some care about cost and efficiency. Apple has shown that you can have performance and efficiency and seems to be doing really well with it. The thing is that Intel competes directly with AMD on performance and they both seem pretty committed to it.
Since rumors say that gaming notebook CPUs (Dragon Range) are power-limited desktop CPUs (Raphael), some reviews have focused on benchmarks with a 65W limit to predict the performance of future gaming notebooks with Zen 4.
I have to a
Apple sillicon in desktop is meh. pointless
I have to agree with you. Eco mode at 65Watts the 7950X destroys the M1 Ultra. The AMD machine will also be modular.
Apple sillicon in desktop is meh. pointless
You can limit the 7950X to 65 Watts and still pref better than M1 Ultra while being efficient. Laptop Zen 4 parts will be on TSMC 4nm and they will be super efficient.
Does it perform better than the Ultra on all workloads?
The M1 Ultra is overkill for my needs. Frankly so is the M1 Max. I guess we'll see what kind of laptop parts they wind up with but I think that Apple will be on 3 nm the first half of 2023 and that should be an interesting reset. One complaint that I've been seeing in the reviews is the cost of MB, CPU, Cooler, PSU and RAM and that the 6 core models are not really value parts as they have been in the past. I actually do not plan any x86 builds for quite some time because prices for so many components has gone up a lot. I'm also fairly annoyed at the situation with EVGA and nVidia.
And I like macOS and have given up on Hackintosh as it's just easier to buy Macs these days.
So far the most surprising bit to me is Zen 4 seems a tock over Zen 3 (which is a tock over Zen 2). Zen 4 still has the "out of fashion by now" 4-wide decoder for the past decade or so:
Perhaps 5/6-wide decoding is reserved for Zen5/6 (no pun intended). I dare not imagine how much performance uplift to look forward to in two years time.
Both M2 & Zen4 demonstrated from different angles the importance of process node. Round 2 will be 3nm M2 vs 4nm mobile Zen4 in 2023 as other folks said.
Perhaps 5/6-wide decoding is reserved for Zen5/6 (no pun intended). I dare not imagine how much performance uplift to look forward to in two years time.
Both M2 & Zen4 demonstrated from different angles the importance of process node. Round 2 will be 3nm M2 vs 4nm mobile Zen4 in 2023 as other folks said.
It doesn't. The geomean SPEC2017 scores for 7950X (full TDP) are 92.5 and 100.9 for int and fp tests, respectively. For M1 Max (also from anandtech benchmarks) its 53.3 and 81.07. Double that for the Ultra. So Zen4 is about the same performance in int tests and considerably slower for fp tests. This is also reflected by Geekbench where 7950X and Ultra score about the same. Zen4 CPUs are of course faster in single core in almost every test. Of course, 7950X uses much more power.
mi7chy's raving about the CB23 scores in 65W mode is just mi7chy being mi7chy. As already discussed on these forums many many times, CB23 represents best case scenario for AMD and worst case scenario for Apple. It's a SIMD-focused benchmark with long dependency chains, so it's particularly well suited for x86 with it's higher frequencies and SMT, where Apple will struggle to keep the (already lower throughput) backend busy. And the result will be similar for pretty much any SIMD throughout test. M1 and Zen4 can do the same amount of SIMD operations per clock, but M1's clock is simply lower. And finally, in CB23 Apple suffers from the fact that it uses an x86 SIMD optimised library where x86 SIMD semantics is emulated on top of ARM SIMD (a recent patch to Embree library claims to improve performance on M1 by 8%, but it's not integrated into C23 and I'd expect the codegen to still be suboptimal on ARM).
To put this in perspective, a power-restricted 7950X seems to be somewhere between 30% and 50% faster than the Ultra (score varies by benchmark) at the expense of using ~30% more power (per #204 ). It's a shame that nobody really publishes the package power during these tests, that's bad tech journalism (see also edited note). So no, Zen4 is not as efficient as M1 across the board, but the difference is probably just below 20% at the same power consumption per core, which is a great result for AMD nevertheless. For desktop use, there is no doubt that 7950X is a better choice, even with it's higher power consumption.
Edit: just learned something about AMD's TDP. Turns out AMD defines TDP differently than Intel. For Intel the TDP is the target maximal power consumption under depending sustained multicore workload. For AMD it's a more or less an arbitrary number that is calculated from the spread between case and ambient temperatures. Anyway, 65W TDP on Zen4 corresponds to a package power of 88 watts.
Last edited:
The tick should be Zen 5, which will launch in 2024. AMD has said that Zen 5 will be a major redesign and leakers have revealed that it will be a hybrid microarchitecture with larger caches and some fixed-function hardware for AI.
Well avoid Zen 5 and wait for Zen 6. Those who buy Zen 5 are beta testers
What are you talking about? There is only a Cinebench score, which is badly optimized for Apple or ARM in general. You cannot possibly generalize based on Cinebench!
This is particularly obvious in floating point SPEC, where the unrestricted 7950X is only 25% faster than M1 Max despite using much more power.
AMDs TDP definition is totally bogus and cannot be used to determine the actual package power. The situation with Intel is much better, as TDP actually refers to actual package power under load.
Indeed the Cinebench performance penalty for ARM CPUs is quite huge - which is no surprise as Embree is open source and just a quick glance reveals the issue.
Apparently people do not learn and still take Cinebench as gospel when it comes to comparing x64 CPUs with ARM64 CPUs.
Last edited: