Yes Houdini is a great chess engine and much slower on Apple devices than on Intel or AMD CPUs.
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.
A bunch of new mac studio benchmarks dropped
- Thread starter dugbug
- Start date
- Sort by reaction score
It's interesting to me that even with that massive cooling system (at least it seems massive in the images), they're still running the M1 Max at the same performance levels.
I'd read an article once upon a time that suggested that among the optimizations in the M1 was that it was only designed to run at one clock rate. I only remember seeing it once, and I can't find it now, so that leads me to believe it was merely speculation.
In general, though, if you limit the number of variables you're designing for you can make a thing simpler and more optimized. Apple doesn't really have a need to bin their processors for performance in the way Intel does, they're differentiating by complexity. So the idea of optimizing for one frequency doesn't seem absurd...
I'd read an article once upon a time that suggested that among the optimizations in the M1 was that it was only designed to run at one clock rate. I only remember seeing it once, and I can't find it now, so that leads me to believe it was merely speculation.
In general, though, if you limit the number of variables you're designing for you can make a thing simpler and more optimized. Apple doesn't really have a need to bin their processors for performance in the way Intel does, they're differentiating by complexity. So the idea of optimizing for one frequency doesn't seem absurd...
Last edited:
Hasn't Geekbench been verified as correlating extremely well with the industry-standard SPEC benchmarks that have been used for decades?
A little searching brings up a Medium post from NUVIA, where they tested SPECint 2006 and 2017 against Geekbench 5 on a bunch of CPUs and found a near-perfect linear correlation for both single-threaded and multi-threaded scores (less than 1% error):
Yep. As I have been saying in many threads as a game developer - it’s about marketshare. The hottest game right now (Elden Ring) runs great on my GTX 1080 at 1080p resolution. Steam reports most used GPUs on Windows are 1060 and 1650. The M1 Ultra is at least better than those.
Don't think socket sTRX4 ITX motherboard exists. Socket is too fat plus they tend to have eight if not four memory sockets.
Smallest was micro-ATX but it was for Threadripper 2000 and 1000 series.
https://www.asrock.com/mb/AMD/X399M Taichi/index.asp
Last edited:
I wonder if this is partially the chip shortage issue. Pumping out M1 Max chips that can be in several products vs only the Mac Studio.
There is an Asrock server board that can take a 64 core Epyc 7763 .
ASRock Rack > ROMED4ID-2T
Great post. Of course it will be ignored by those looking to post chess benchmarks and software that is unoptimised for Apple Silicon.Hasn't Geekbench been verified as correlating extremely well with the industry-standard SPEC benchmarks that have been used for decades?
A little searching brings up a Medium post from NUVIA, where they tested SPECint 2006 and 2017 against Geekbench 5 on a bunch of CPUs and found a near-perfect linear correlation for both single-threaded and multi-threaded scores (less than 1% error):
There is an Asrock server board that can take a 64 core Epyc 7763 .
ASRock Rack > ROMED4ID-2T
www.asrockrack.com
Too bad single socket Epyc 7763 is slightly slower and double the cost compared to Threadripper 3990x. Where the 7763 really shines is 1.87x (closer to perfect 2x) performance scaling across dual socket compared to 1.4x scaling from M1 to M1 Pro/Max 1.4x doubling of CPU performance cores on same chip.
https://openbenchmarking.org/test/pts/blender
https://openbenchmarking.org/test/pts/stockfish
Last edited:
Threadripper isn't a high volume product. Ryzen 9 16 cores is good enough for a wider range of folks. Like folks are pointing out it is a $2-6K processors. The size of the market really isn't that large.
The binding is to a OEM vendors firmware so Lenovo buyers could sell to someone wanting to upgrade their Lenovo. That will keep the 'used' prices high but it won't stop them completely.
Long , long term AMD could do an unlocked , "non Pro" version, but I suspect there probably won't be much demand for one. ( at that point the mainstream Ryzen 9 would be Zen 4 or 5 and more cost effective. )
The Threadripper is a refactored Epyc SoC. The point of "getting ahead" is making a better Epyc not a Threadripper. When there is a large excess or materials for a Epyc packages than AMD can get another "side hustle product" out of the 'golden sample' chiplets that can clock higher as fall out from doing the Epyc work.
Two fold problem at the moment. First, Not enough wafer starts to go around. So it makes about zero sense to sell a slightly lower margin Threadripper when AMD can make more money selling it as an Epyc. AMD has been holding the Threadripper 5000 back for almost a year. It isn't like they didn't prepare the product. They just didn't have enough product to sell.
When the wafer supply chain is back to running in excess of demand and AMD is still throttling Threadripper releases then will have much more evidence of a 'lazy' problem.
Secondly, while Threadripper has been on hold AMD has deployed stuff like VCache on top end of Ryzen and on Epyc. Dell/HP/Lenovo sell top end workstations based on dual socket Xeon SP. Other than scarcity if Epycs there isn't much to stop them from doing a 'workstation' by taking an Epyc board and putting it into a tower. IF had some HPC code wanted to do 'deskside' that would be viable using an Epyc VCache solution while waited to see if VCache Threadripper every showed up.
AMD has been moving on Epyc update progresssion. It is the trickle down to Threadripper that is small. But also not all that necessary because AMD isn't blocking the mainstream Ryzen from steady progressions. Threadripper is just in a small Goldilocks zone between them. Given AMD has Nvidia GPU 'war' to fight also. It isn't a stratically important Goldilocks zone. Even so, Apple is still behind AMD in workstation SoC without even putting tons of effort.
Yes, M1 only scales horizontally, by increasing the number of clusters. The clusters themselves are “locked” and will perform the same in all M1 chips. Some folks familiar with chip design speculated that this is how Apple achieves their superior power efficiency: by optimizing the very layout to max out at a relatively low speed. These chips can scale down very well (I remember seeing that M1 has a power level with something really ridiculous like 50mhz), but it cant be pushed higher even if there is enough thermal headroom. This is very different to x86 designs.
I get it, 50MHz seems silly because the static power is probably dominating the dynamic power at that point on a modern chip, but it just made me remember how stupid fast our current systems are. The original Mac ran a single core 68000 with a pitiful instructions per clock metric at a clock rate just under 8MHz and it seemed to run Excel just fine in its day.
The original Mac had 128K of RAM, the new Mac Studio has 128GB of RAM-- more than a million times more. ?
What on earth are we doing with all that extra computing power? Refreshing Facebook...
Turns out though that some things remain the same. When I went to double check the clock rates, I found this little tidbit from Burrell that I'd forgotten about the original Mac:
"The RAM is triple-ported; this means that the 68000, screen-displaying hardware, and sound-output hardware have periodic access to the address and data buses, so that the video, the sound, and the current 68000 task appear to execute concurrently."
Turns out the unified memory architecture is taking us back to the future...
Last edited:
Initially thought this was a sardonic reference to the legendary stockfish benchmark, but turns out there’s a Houdini chess engine that is influenced by stockfish. Learn something new every day
But yeah, the 3D programme, not the other one.
@JimmyjamesEU posted this in another thread but it seems that M1 Ultra will be outperforming 3090 as Apple claimed. Here are some results from rendering Disney's Moana in Redshift:
2x 2080ti = 34m:17s
M1 Max = 28m:27s
Single 3090 = 21m:45s
2x 3090 = 12m:44s
My guess for M1 Ultra 48c is 18m:58s and for 64c 14m:13s.
arstechnica.com
arstechnica.com
arstechnica.com
2x 2080ti = 34m:17s
M1 Max = 28m:27s
Single 3090 = 21m:45s
2x 3090 = 12m:44s
My guess for M1 Ultra 48c is 18m:58s and for 64c 14m:13s.
Peek Performance: March 8th Event [Event discussion begins on p6!]
My BTO Mac Studio has just switched to "preparing to ship" :eek: Maybe I'll get it sooner rather than later? I have a hold on my credit card, but it's still at 'Processing'. No hold for the display yet though. I had to spend an hour on the phone with Apple on Saturday to get mine cleared...
arstechnica.com
Peek Performance: March 8th Event [Event discussion begins on p6!]
My BTO Mac Studio has just switched to "preparing to ship" :eek: Maybe I'll get it sooner rather than later? I have a hold on my credit card, but it's still at 'Processing'. No hold for the display yet though. I had to spend an hour on the phone with Apple on Saturday to get mine cleared...
arstechnica.com
Peek Performance: March 8th Event [Event discussion begins on p6!]
My BTO Mac Studio has just switched to "preparing to ship" :eek: Maybe I'll get it sooner rather than later? I have a hold on my credit card, but it's still at 'Processing'. No hold for the display yet though. I had to spend an hour on the phone with Apple on Saturday to get mine cleared...
arstechnica.com
@JimmyjamesEU posted this in another thread but it seems that M1 Ultra will be outperforming 3090 as Apple claimed. Here are some results from rendering Disney's Moana in Redshift:
2x 2080ti = 34m:17s
M1 Max = 28m:27s
Single 3090 = 21m:45s
2x 3090 = 12m:44s
My guess for M1 Ultra 48c is 18m:58s and for 64c 14m:13s.
Peek Performance: March 8th Event [Event discussion begins on p6!]
My BTO Mac Studio has just switched to "preparing to ship" :eek: Maybe I'll get it sooner rather than later? I have a hold on my credit card, but it's still at 'Processing'. No hold for the display yet though. I had to spend an hour on the phone with Apple on Saturday to get mine cleared...Peek Performance: March 8th Event [Event discussion begins on p6!]
My BTO Mac Studio has just switched to "preparing to ship" :eek: Maybe I'll get it sooner rather than later? I have a hold on my credit card, but it's still at 'Processing'. No hold for the display yet though. I had to spend an hour on the phone with Apple on Saturday to get mine cleared...Peek Performance: March 8th Event [Event discussion begins on p6!]
My BTO Mac Studio has just switched to "preparing to ship" :eek: Maybe I'll get it sooner rather than later? I have a hold on my credit card, but it's still at 'Processing'. No hold for the display yet though. I had to spend an hour on the phone with Apple on Saturday to get mine cleared...
From the thread pretty much, but this is one of the situations where the memory makes a significant difference. It does kind of raise the question of how often you are going to be trying to render something as large as the Moana set locally*, but also points to how things like the Blender BMW or classroom benchmarks don't really reflect the advantages of all that sweet sweet memory. The M1 Ultra will be great for rendering large scenes and, more importantly, the time to first pixel should be much faster as it doesn't have to move things from RAM to GPU memory.
Looking forward to seeing some benchmarks in the next few days (but suspect we'll all be arguing over unrepresentative Blender benchmarks
)* Possibly, as USD becomes a thing, splitting scenes up into individual shots becomes less of a necessity and you'd just have one usd with all the things in, so displaying and rendering large scene would become more common at least as far as film and tv goes (feature animation would probably benefit most from this approach, hence Pixar creating usd).
That makes not much sense. The "pushing higher" is a process and physical design property not an architecture property in the first place - this is unless you assume that the M1 is already operating at the maximum performance corner - max Voltage, lowest VT, maximum buffered, highest performance cells - which is not the case.
I am not a CPU engineer, so I don’t know how all this works on the technical level, but it is a fact that M1 Firestorm cores are limited to 3.2ghz (which they achieve running at approx 5W). This upper limit is the same for every M1 based product, be it the passively cooled Air or an iPad or a high-end workstation like the Studio. If Apple had a technical capability to push these CLUs past 3.2ghz on their desktop systems, they would have likely done. Another striking observation that M1 does not use frequency-based binning at all - in stark contrast to x86 CPUs. All M1 cluster performance characteristics are identical no matter which product we look at. The only binning occurs, again, at the horizontal level, by tweaking the cluster size.
I have no idea how Apple does it. The 3.2ghz limit could be a physical property of this CPU design, or it could merely be a statistical common ground most manufactured chips can sustain, or it could be a business decision on Apples side (a weird one). Regardless of the technical reason, this is what we have, and this is what I meant in my post: M1 relies exclusively on horizontal scaling, the cores themselves do not exhibit any performance scaling/binning across products. This is true for the CPU and the GPU equally.
I think it might be a design choice. As you say it allows them to easily scale horizontally while maintaining power efficiency and they get to reuse the same basic cluster design over and over again - just cut and paste. Obviously literally in the case of Max -> Ultra. I know there is probably more to it than that but it makes sense. It simplifies the design. Maybe there are other considerations like not having to bin for tight frequency tolerances increases yields but I don't know if they're anywhere near TSMC's limits on that anyway. Maybe?
Intel and AMD also use same cores - and even same dies - for different products, by allowing different configurations at various performance and power levels. Apples approach is obviously a design choice, but it’s different enough from the mainstream practice to ask why they do it this way. Is it an arbitrary restriction (the chips can run faster but Apple limits them), is it a physical restriction of the design, is it a statistical restriction? At any rate, it’s more sane and transparent than what happens elsewhere.