A Dive into Ray Tracing Performance on the Apple M1
willusher.io
willusher.io
Edit: The article does compare the ray tracing capabilities of the M1 GPU to a RTX 2070. The RTX 2070 obviously crushes it, as the author points out the obviously unfair comparison. The RTX card uses way more power, is much largerx and has hardware acceleration for ray tracing.
I am interested in seeing if Apple adds hardware acceleration for RT to future chips (or if they will support external AMD GPUs).
Isn't RTX card using less power per ray according to this benchmark? This also applies to ray/size unit I think.
People have been rendering on any CPU they could get their hands on for the last forty years and there is no such thing as "software accelerated ray tracing". That would be like someone saying their car does engine accelerated driving.
Is this a particular term/concept, or do you just mean non-hardware-accelerated ray tracing?
If the former, what does it mean for something to be "software-accelerated?"
If the latter, I apologize if this comes off as nit-picking.
Though I do understand the desire to be pedantic, because ISAs like this and their SIMD instructions can kinda blur the lines between what constitutes as "hardware accelerated". That's why I try to avoid it as much as possible for the sake of simplicity.
GPUs require acceleration because the memory latency on GPUs is pretty large, caches are small, and GPUs have such huge compute that waiting around for the "next = bvh-tree->left" operation takes a huge amount of time.
By making dedicated acceleration structures and hardware, that latency can be mitigated, and the huge compute capabilities of GPUs can be unleashed.
----------
In contrast, CPUs are already really good (relatively) at pointer-dereferencing loops. CPUs have less compute and more memory-optimizations (relative to a GPU anyway).
so their own raytracers are not real-time either
I don't know. Desktops simply tend to have a much longer lifespan and so far every new laptop I owned in the past 10 years outperformed my then desktop PC for that precise reason (at the very least in certain aspects like performance per watt).
What I'm rather afraid of is other manufacturers adopting Apple's insane vendor lock-in and total hardware lock-down. Any failure in an M1 system results in complete data loss and a logic board replacement - RAM on SoC, SSD soldered to the board, etc. No more choice and zero upgradability - the oldest components in my current desktop are 12+ years old, but the rest was gradually upgraded (SSD, GPU, CPU, additional RAM, HDD replacements, etc.)
Such things won't be possible in brave new world of SoCs...
People aren't going to be buying new desktops every 1-2 years, but many of them will buy new GPU's, or cases, or fans etc. The best part is that we get to upgrade the components that will actually provide better performance. For instance, why would I upgrade my entire machine just for a better graphics chip, if my SSD, CPU and RAM are still in tip top shape, short answer is I wont.
DDR4 as an example was released 7 years ago and DDR5 is still yet to be released not less a required upgrade for a new CPU.
There were a couple of exceptions, I just got lucky that old stuff was still around to buy.
PC's have always been this way and sometimes you get lucky that you can use the same slot for a long time, other times you dont.
For me, if I wanted the latest CPU I would need to upgrade my mobo and CPU. But not my RAM, my GPU ($1500), power supply or my SSD's. If I was on a Mac and one of these parts broke, I would need to actually buy all of the internals again which are soldered to the main board.
In your case, if nothing fails in your machine both options don't impact you at all. Whereas if a part does fail you might be up for a more expensive repair bill than previous
Intel® Core™ i7-1165G7 by default is an 15W TDP CPU, especially true if you are spreading load across all cores. ( 28W if you are on single Core TurboBoost ) It is on 10nm SuperFin ( equivalent to ~TSMC 7nm ), Quad Core and 8 Thread, Not sure about its AVX 512 clock speed limitation I dont have time to dig that up.
The Apple M1 is ~ 24W TDP, 8 Core 8 Thread, with 4 being High Efficiency Core on TSMC 5nm.
You are looking at a ~60% higher TDP number. The results are still impressive, but needs these additional context before it is consumed.
I will also not be surprised if A15 has a major GPU uArch change. ( Something similar to IMG B-Series )
Intel chips have much higher peak and sustained actual power usage than their TDP indicates. At the wall, those “15w” systems are actually much closer to 40-60w on sustained loads.
They dont. PL2 only last at best 30 secs. Not Sustained Loads. Laptop depending on heatsink config sustain at either 15W at base clock or 28W at slightly higher clock.
The M1 Mac mini under full load is 39W as an official Apple Published Figure [1].
Not sure why people are downvoting my original post. But I guess that is the state we are in for hardware topic on HN.
https://www.anandtech.com/show/16252/mac-mini-apple-m1-teste...
For comparison, Panther Lake 1165G7 NUC uses as much as 85w (70w sustained) when maxing out its load. The author even notes how they'd beefed up the cooling.
https://nucblog.net/2021/02/panther-canyon-i7-nuc-review-nuc...
The last one I’ve seen that had it was a 8700K with dual 1080’s DRT monstrosity.
That's news to me. I try to keep up to date with laptops, and all the reviewers have pretty accurate power consumption measurements. Even just using the internal reporting reveals quite a few details [1].
Even a "12W" TDP 11th gen Intel CPU can draw up to 50W in turbo mode [2].
The 11th gen "45W" laptop CPUs have a PL2 power draw of 135W [3], which, thermals and VRMs permitting, can be held indefinitely according to the specs (which is what makes Intel's whole TDP-rating useless indeed).
So yes, it's very hard indeed to estimate newer laptop's power consumption, but for entirely different reasons. It's primarily the cooling and OEM's choice of how to implement the very loosely defined (Intel!-) specs that define power draw. If you have excellent cooling, there's nothing to stop an 11800H from drawing 135W for minutes at a time... (which can be measured, both internally and from the wall no problem).
AMD CPUs on the other hand mostly adhere to the published power rating (give or take 20% - again, cooling permitting and configurable by the OEM).
[1] random example: https://bit.ly/34xLKpM
[2] https://www.anandtech.com/show/16084/intel-tiger-lake-review...
[3] https://www.itworldcanada.com/article/eight-core-intel-tiger...
Also, Intel for the love of God sort out your product names, use a proper system rather than /dev/random Lake
Suffixes are the main thing on Intel's systems.
U -- Ultrabook class 15W.
M -- Previously the standard laptop class 25W.
H / HQ / etc. etc. -- High-performance laptop: 45W. Q for quad-core but that doesn't seem to be a big deal anymore.
T -- Power-optimized Desktop
No suffix -- Standard Desktop
K -- Unlocked Desktop.
-----
The i5 / i7 is a marketing name, and more about MSRP than anything else. You can mostly ignore it. i7 and i9 are more expensive, while i3 and i5 are cheaper.
The last 3 digits are highly-specific codes within a family. The other digits are the generation number: 8 for 8th generation, 11 for 11th generation. For example, 2700K means 2nd generation Unlocked Desktop processor.
----
I personally start with the suffix: if you know someone's looking for a battery life longer than 4-hours, I go with the U-class, then I pick out an MSRP for them (usually i7, but i5 or i3 if they care more about budget).
If the battery life under 2-hours is fine, I push for H-class for the higher performance (though these are usually on 17in laptops).
I stay up to date and make sure that I buy a current-generation, or at worst a "last generation" product if I know that who I'm talking to cares about price. That means that I will recommend a 11th generation i7 to most people, but if people care about price, maybe a 10th generation i7 or a 11th generation i5.
Xeon has its own naming convention that was changed maybe 5ish years ago? (Whenever they went from E3 / E5 / E7 designation to Xeon Bronze / Silver / Gold / Platinum). There are also naming schemes for the Atom-line, but these aren't commonly bought by consumers.
This also results in situations like for example selling working 10 cores as 8 cores (with cores turned off) to fulfill market needs in case of shortages ... shortages of "faulty" dies. In some cases methods are available to "upgrade" you chip again. tl;dr how to abuse chip binning (or how to void your warranty)
As I understand it, the ability to re-enable parts isn't really a thing any more. It was true historically that you used to be able to re-enable cores, but most modern chips have the components fused off at a hardware level.
The problem is that their actual product naming is such a massive mess that people prefer to learn the code names instead of trying to figure out what, exactly, a i7-1165G7 is and how it relates to, say, a i5-11400H?
Combine this with the fact that they have completely different processor families for different segments concurrent with these, there are going to be hundreds of SKUs in total on the market at any point.
I don't know if there is any sort of naming system that can salvage that to make any intuitive sense. The actual product names mostly serve to function as keys to search ARK.
They actually have a guide explaining all the bits. It doesn't really matter since most of it is arbitrary and marketing (including µarch rebadging same as the GPU vendors), so even within a generation it tends to tell you very little in and of itself.
Then it became inconvenient so generations and µarchs got disconnected (because marketing, and then more marketing as they had to introduce µarch refreshes because they couldn't move through their plans)
https://www.intel.com/content/www/us/en/products/sku/213805/...
Retina displays, multitouch on touchscreens, and now ARM processors. The notable thing is that Apple makes the enabling incremental change (to a point where the product is good enough for mass market) seem like a step function.