I'm trying to learn about ISAs.
And, after taking a class and watching a whole bunch of RISC vs CISC videos, I still don't know 'why' Apple Silicon is faster at a lower power consumption than Intel's chips.
Can someone tell me?
And, after taking a class and watching a whole bunch of RISC vs CISC videos, I still don't know 'why' Apple Silicon is faster at a lower power consumption than Intel's chips.
Can someone tell me?
Apple have access to the newest and best fabbing processes from TSMC. That alone can put a chip a generation ahead in terms of efficiency. Intel foundries have been struggling for years and AMD only get TSMC's scraps a few years later.
Increasing clock speed is the easiest way to increase performance, but power increases quadratically with clock speed so it's very inefficient. Apple clock their processors pretty low and instead focused on increasing instructions-per-clock, which is generally more efficient but requires more die area. As an example, a lot of noise was made about the M1's 8-wide decoder, twice as wide as contemporary x86 chips, which is an important bottleneck for IPC.
Increasing instructions-per-clock requires much bigger CPU cores, and the dies Apple use for their SoCs are famously enormous. While other chip manufacturers are very conscious of performance/$, Apple's vertical integration gives them wider margins and their customers are much less sensitive to value. So Apple can afford to spend a lot more on bigger chips that are on par in performance but have much greater efficiency.
This is the money shot. Margins on Android phones are often in the single digit percentages, they have even had periods where they’re negative. Qualcomm or Mediatek could easily make phat chips but there would be almost no market for them.
I am very curious what the actual performance of the Snapdragon X is going to be. Laptop margins are often just as thin, so it can’t be too expensive a chip either..
I'm sceptical of Snapdragon X's real world performance until these things actually ship. There's a lot of big promises made. Intel promises big too.
[1]https://videocardz.com/newz/snapdragon-x-series-chips-cost-o...
Nothing magic.
But the deep integration with better margins and higher end product prices allows Apple to have more expensive CPUs than the competition.
What do you consider legacy? Can you point out how much it does affect die area and the instruction decoder?
If I'd take a guess it's not more than maybe 2 or 3%.
But I do think the benefits are in a faster and easier decoder, and easier memory interfaces etc.
Not a hardware engineer, from a software engineer dropping legacy code makes development of new features easier and more cost efficient, so I can concentrate on other things. But as I've said, I'm no hardware engineer.
When AMD/intel have moved to the same or comparable nodes, they’re not catching up for perf/watt. Lunar Lake will be on the same N3E as M4, so we’ll have to see how it compares.
They are though? If you compare the Apple and AMD processors that are both on e.g. TSMC 5nm, the AMD ones have if anything better performance per watt. Compare the 7945HX3D to any Apple CPU on the same 5nm process with a similar TDP and the 7945HX3D will generally have similar if not better performance.
But people keep doing comparisons right after Apple is the first to release a CPU on a new TSMC process, and naturally if you don't compare like with like, the newer process is more efficient.
Another reason people keep getting confused is that AMD isn't afraid to go back and release updated models of older cores on their original process nodes. So for example the 7730U is from 2023, it's just a year old, but it's Zen 3 on the 7nm process they've been using since 2019. Meanwhile the "real" 2023 release was Zen 4 on 4nm (e.g. 7840U), but don't pick the wrong one to do your comparison.
Sounds interesting, do you have numbers - s I get this argument often, that on the same node, Apple is so much better.
https://nanoreview.net/en/cpu-compare/apple-m1-ultra-vs-amd-...
They both have a similar TDP, a similar number of cores and the performance on most tests is similar or better for the 7945HX3D.
OT: I think it also shows the bad iGPU compared to Apples offerings, so too many x86 laptops have additional GPUs, which reduces battery life (we'll see what Strix brings)
Also both have 16 performance cores, but the Ryzen is faster in multicore. The Ryzen is one year later b/c due to node access/ costs. Doesn't look like "far behind".
Discrete GPUs obviously use more power because their use case is when the iGPU isn't fast enough. Hypothetically you can make a discrete GPU with the same performance as an iGPU and it would be no more power hungry. These have occasionally been available, basically monitor cards for CPUs without an iGPU etc. But the ones with a lot of fans and VRMs have them because they do indeed use a lot of power -- which isn't the same as having poor performance per watt, because they're also really fast.
Pointing to TDP as if it meant anything compared to actual power measurements is usually a mistake. It doesn’t mean anything anymore. I really hate it as much as anyone else, I love architectural comparisons/etc... but you have to look at actual measurements from actual devices in a particular workload. Which, again, is very unfortunate since most reviewers don't do that, certainly not for more than one or two workloads.
Cinebench R23 is also not a particularly great test compared to SPEC2017, Geekbench 6, or Cinebench 2024, and most vendors still haven’t switched over even to the newer CB2024. Generally the older tests tend to give x86 a bit of a boost because of the very simple test scenes/lack of workload diversity.
(I was looking at this when M3 Max MBPs came out and 7840HS was the current x86 hotness etc... just very few reviewers with direct CB2024 numbers yet and fewer with actual power measurements, let alone anyone having any other workload benched... geekerwan and notebookcheck seem to be the leaders in doing the actual science these days... save us, Chips And Cheese...)
Geekbench 6 (Single-Core)
M1 Ultra 2436
Ryzen 7945HX3D 2890
Gamer Nexus has several videos on the topic around PBO and Eco mode.
https://images.anandtech.com/doci/16084/Power%20-%2015W%20Co...
That's why I said: you have to look at the actual power measurement and not just the box spec, because Apple generally undershoots the box-spec for cpu-only tasks, and AMD always exceeds it substantially due to boost. Obviously if you give the AMD processor twice the power it's going to be competitive, that's not even in question here. The claim was, more efficient than Apple - and you simply cannot assess that with the box TDPs, or even cTDPs, because x86 processors make a mockery of the entire concept of TDP nowadays.
(and it didn't use to be like that - 5820K and 6700K would boost to full turbo under an AVX2 all-core load within the rated TDP! The "boost power is not the same thing as TDP" didn't come around until Ryzen/Coffee Lake era - accompanied by ploppy like "electrical watts are not thermal watts" etc)
https://www.anandtech.com/show/8426/the-intel-haswell-e-cpu-...
https://www.anandtech.com/show/9483/intel-skylake-review-670...
edit: notebookcheck has cinebench R15 at 36.7 pt/w for M3 Max and 33.2 pt/w for 7840HS, and 28 pt/w for Pro 7840HS (probably a higher-cTDP configuration). Obviously CB R15 is miserably old as a media benchmark, but perhaps ironically it might be old enough that it's actually flipped around to being more representative for non-avx workloads lol.
CB24 MT at 50% higher for M3 Max and CB R23 at 11% higher (shows the problem there with x86 processors on R23). Looking back at Anthony's review-aggregator thing... they're using the highest score for 7840HS I would assume (it's 1k points higher than median on notebookcheck) and also they're probably comparing box TDPs, which is where the inversion comes from here (from parity to a 10%-ish lead on efficiency). Because actual measurements of the M3 Max, from actual reviewers, have CB 2024 at around 51W all-core, and it almost doubles a 7840HS's score there.
https://youtu.be/BpeVdHwaHWU?t=644
And looking at it a little further even that nanoreview aggregator thing doesn't claim 7840HS is more efficient... they rate the M3 Max as being 16% more efficient. Seems that is a claim Anthony is imputing based on, as I said, box TDP...
edit: I’m also not even sure that nanoreview article has the right TDPs for either of them in the first place…
"We don't have comparisons because nobody checks but Apple is better" is not convincing to me.
This is more Intel than AMD. AMD will generally stick to the rated TDP unless you manually reconfigure something.
> and usually not reached unless it’s a mixed cpu-gpu load like gaming.
For mobile chips you'll generally hit the rated TDP on anything compute-bound in any way, simply because the rated TDP is low enough for a single core at max boost to hit it. And you'll pretty much always hit the rated TDP on a multi-threaded workload regardless of if the GPU is involved because power/heat is the limit on how high the cores can clock in that context, outside of maybe some outliers like very low core count CPUs with relaxed (i.e. high) TDPs.
> you have to look at actual measurements from actual devices in a particular workload.
But then as you point out, nobody really does that. Combined with the relative scarcity of parts made on the same process, direct comparisons of that kind may not even exist. Implying that the people saying the Apple processors are more efficient on the same process are doing so with no real evidence.
no, it's generally the opposite. Intel tends to exceed TDP substantially due to boost, and AMD actually exceeds it by greater margins than Intel.
This is consistent across both desktop and laptop, but AMD's mobile SKUs are actually allowed to exceed TDP by even larger margins than the desktop stuff.
https://images.anandtech.com/doci/16084/Power%20-%2015W%20Co...
https://www.anandtech.com/show/16084/intel-tiger-lake-review...
But the TDP is also configurable. It's one of the main differences between different model numbers with the same number of cores which are actually based on the same silicon. The difference between the 4800U and the 4900H is only the TDP and the resulting increase in clock speed. But the TDP on the 4900H goes up to 54W.
Whereas the TDP on the 4800U they tested there is configurable even within the same model, from 10-25W. And then we see it there using a sustained 20-25W, which is perfectly within spec depending on how the OEM configured it. And there is presumably a setting for "use up to max power as long as not thermally limited", which is apparently what they used, and then combined it with a 15W cooling solution. Which is what you see clearly in the first graph on the same page:
https://images.anandtech.com/doci/16084/Power%20-%2015W%20Co...
It uses more power until the thermal solution doesn't allow it to anymore.
In the second test it can sustain a higher power consumption, maybe because the test is using a different part of the chip which spreads the heat more evenly instead of creating a hot spot. But this is all down to how the OEM configured it, regardless of what they advertised. That CPU is rated for up to 25W and is only using 22. Obviously if the OEM configures it to use more, it will.
Geekbench 6 (Single-Core)
M1 Ultra 2436
Ryzen 7945HX3D 2890I meant to ask for benchmarks across the power spectrum (measured by actual pull not OEM TDP) and off charger. That’s where the x86 processors down clock heavily to keep power in check, and M series pulls ahead because it doesn’t follow suite.
The M series, even at launch, weren’t the king of outright perf, but they were for power to perf.
The problem with asking for benchmarks that actually measure power draw is that somebody would have had to have done them, which hardly anybody does, and specifically on the few CPU models that were made on the same process. So if that's what you want to see, where are yours? Find a comparison that actually measures power draw between two CPUs with similar performance/TDP on the same process, like the M1 Ultra and the Ryzen 7945HX3D.
> That’s where the x86 processors down clock heavily to keep power in check, and M series pulls ahead because it doesn’t follow suite.
That's not really how either of them work. CPUs from any vendor will run at lower clocks under threaded workloads in order to meet their power budget. This is also why the high core count CPUs actually have the best performance per watt -- the cores are clocked lower which is more efficient but you still get the performance from having more of them. And then "race to sleep" etc.
Lunar Lake will be on N3B process.
AMD and Intel can do what Apple does and Apple can do what AMD and Intel do, but each company chooses the path they believe it makes more sense for them.
Because the power consumption grows at least quadratically with the clock frequency, having the same speed at a lower clock frequency provides better energy efficiency.
Meanwhile, the difference in IPC (instructions per clock cycle) between Apple and the others has been reduced, but Apple still has a healthy advantage.
The different microarchitecture explains the energy efficiency advantage for single-threaded tasks.
On the other hand, for multi-threaded tasks, the power consumption per chip area is limited by cooling and the energy required to do a quantity of useful work (which is done by switching a certain number of logic gates) is mainly determined by the parameters of the CMOS process used to make the CPU die (assuming that the CPU designers are not incompetent).
So for multi-threaded tasks the better energy efficiency of the Apple CPUs has been determined mostly by the fact that they have been the only company able to use the latest and greatest TSMC CMOS process.
Apple has the benefit of controlling the OS, so no scheduler problems.
Intel came later but seems to make good advances, if you look at e-Cores in Lunar Lake.
I predict AMD will go the same route (way bezond 5c), now that x86 also focuses on performance/watt and not peak performance (if you watch the announcements).
CPUs with big and little cores were not useful for the old "14 nm" manufacturing processes used by Intel for Skylake and its derivatives, because in those processes the power consumption of a core and its speed could be varied in a very large range, e.g. for the extremes of the supply voltage, the clock frequency of a laptop CPU core could vary between 0.6 GHz and 4.5 GHz, corresponding to a variation of the power consumption per core of even 50 times.
Modern processes have more constrained ranges for the optimal operating parameters of the transistors, so it has become better to split the cores into several classes that are designed with different trade-offs between power consumption and performance, instead of attempting to use the same core both at very low and at very high power consumptions.
The decision to use 2 or more kinds of cores to cover the entire curve of performance vs. power consumption, when you can no longer cover it by varying the operating parameters of a single kind of core, is a no-brainer.
While for the CPUs with a small number of cores, like for smartphones, tablets and light laptops, the existence of multiple kinds of cores is needed for optimum energy efficiency, the CPUs for desktops or big laptops with a great number of medium-size cores, like Raptor Lake, have them for a different reason. The big Intel cores, in the quest for maximum single-thread performance, have poor performance per chip area. The medium-size cores have better performance per area, so in the same limited chip area you can put more cores, increasing the multi-threaded performance beyond what would be obtained by filling the same area with big cores.
Nevertheless, AMD has a better strategy than Intel, because their big cores with high single-thread performance and their "compact" cores with better performance per chip area are logically equivalent, avoiding the software incompatibility problems caused by Intel.
Only if it brings at least the same perf/watt results. We'll see what Lunar Lake vs. Strix does.
Jim Keller (and others) worked on Apple CPUs and on AMDs Ryzen.
So when leaving Apple, he (they) became stupid?