Intel Alder Lake vs. Apple M1

[Gnattu]

No chance it was gonna be $999. I was pleasantly surprised that it was only $1599 and not $1999.

Oh, and the $1599 actually INCLUDES a STAND, unlike the other one

 

[cmaier]

I believe they called M1 Ultra 'the last chip of the M1 family' during the keynote. Anyhow, the M1 Pro/Max/Ultra only supports two-die interconnect, so there can't be a 4-die M1 Max. That'd only be possible with a redesigned chip.

I said something similar in The Other Place, though I noted that it’s possible they *could* do it if they designed an interposer die that slots between the four SoC die, assuming that they didn‘t scrimp on things (e.g. Allowing only 1 bit for CPU id, or whatever).

My guess is that the fabric is some sort of synchronous crossbar with FIFOs to handle clock skew or offset, and not a clock-forwarded bus or the like. If so, that makes it more difficult to allow for 2 more die, though not impossible.

That said, I‘m sure we have to wait for M2 before we see something like 4 SoC’s tiled.

 

[Andropov]

4 M2 Max’s??

Could be. It's impossible to know until some M2 hardware gets released. I not sure they need it (the 2x Max is already pretty powerful). They'd also have to change the geometry of the interconnect to be able to make a tight 4x configuration.

Oh, and the $1599 actually INCLUDES a STAND, unlike the other one

But it can't rotate to be vertical

 

[Gnattu]

But it can't rotate to be vertical

You have to pay $400 more to be able to do that😂

 

[leman]

No chance it was gonna be $999. I was pleasantly surprised that it was only $1599 and not $1999.

Well, something like Dell U3223QE can be had for $900... And of course, Apples display is better, looks nicer and also has a webcam, but almost 80% higher price point hurts…

 

[MayaUser]

And now to answer your question definitively: I love the display but I can’t justify the price. I would buy it if it were $999

its an @999 display with an @599 stand included
So go ahead and take the display only, tell them to keep the stand

 

[Andropov]

You have to pay $400 more to be able to do that😂

I wish the $999 stand from the Pro Display XDR kept that possibility. I don't like the VESA mounts.

 

[leman]

its an @999 display with an @599 stand included
So go ahead and take the display only, tell them to keep the stand

I don’t even want the stand… I’ll use a wall mount. Alas, the stand is “free” 😂

That said, I‘m sure we have to wait for M2 before we see something like 4 SoC’s tiled.

They would also need to find a different RAM arrangement. Can’t have four chips surrounded by RAM like now, unless… two chips on one side of the PCB and two on the other one? 😁

 

[Rigby]

The 16-core PC they mentioned is exactly the Alder Lake CPU from their footnote:

Which begs the question what benchmarks they were referring to. Assuming that the M1 Ultra scales ideally at 2x compared to M1 Max, and they haven't increased the maximum clock (meaning the single thread performance will be the same as M1 Max), we can extrapolate the approximate Geekbench results:

M1 Max: ~1700 single-thread, ~12000 multi-thread
M1 Ultra est.: ~1700 ST, ~24000 MT

For comparison:
i9-12900K: ~1900 ST, ~18000 MT

Nice bump in multi-threaded performance, but you pay a steep price. I wonder what the yields for the M1 Ultra are. Going from the M1 Max to the M1 Ultra Studio adds $2000. Even if you take off a few hundred bucks for 32 GB additional RAM and 512 GB additional SSD, that leaves well over $1000 just for essentially adding an additional M1 Max ...

 

[cmaier]

I don’t even want the stand… I’ll use a wall mount. Alas, the stand is “free” 😂



They would also need to find a different RAM arrangement. Can’t have four chips surrounded by RAM like now, unless… two chips on one side of the PCB and two on the other one? 😁

That seems to me how they’d have to do it, but of course that raises some weird latency issues potentially. Would certainly be awkward getting the signals across all those die to get to RAM on the other side.

 

[Gnattu]

Which begs the question what benchmarks they were referring to. Assuming that the M1 Ultra scales ideally at 2x compared to M1 Max, and they haven't increased the maximum clock (meaning the single thread performance will be the same as M1 Max), we can extrapolate the approximate Geekbench results:

M1 Max: ~1700 single-thread, ~12000 multi-thread
M1 Ultra est.: ~1700 ST, ~24000 MT

For comparison:
i9-12900K: ~1900 ST, ~18000 MT

Nice bump in multi-threaded performance, but you pay a steep price. I wonder what the yields for the M1 Ultra are. Going from the M1 Max to the M1 Ultra Studio adds $2000. Even if you take off a few hundred bucks for 32 GB additional RAM and 512 GB additional SSD, that leaves well over $1000 just for essentially adding an additional M1 Max ...

It's for a doubled GPU too, so yes, you are literally buying "two M1 Max system" and paying the price.

 

[Andropov]

I said something similar in The Other Place, though I noted that it’s possible they *could* do it if they designed an interposer die that slots between the four SoC die, assuming that they didn‘t scrimp on things (e.g. Allowing only 1 bit for CPU id, or whatever).

My guess is that the fabric is some sort of synchronous crossbar with FIFOs to handle clock skew or offset, and not a clock-forwarded bus or the like. If so, that makes it more difficult to allow for 2 more die, though not impossible.

That said, I‘m sure we have to wait for M2 before we see something like 4 SoC’s tiled.

Yeah, I was just reading your thread there. Spot on prediction, as always. I'll have to read a lot more to understand what some of the jargon means, though.

They would also need to find a different RAM arrangement. Can’t have four chips surrounded by RAM like now, unless… two chips on one side of the PCB and two on the other one? 😁

They could make a comically long N-die package, and they'd have two sides free

Terrible latency between the farthest cores too, though.

 

[mr_roboto]

I said something similar in The Other Place, though I noted that it’s possible they *could* do it if they designed an interposer die that slots between the four SoC die, assuming that they didn‘t scrimp on things (e.g. Allowing only 1 bit for CPU id, or whatever).

My guess is that the fabric is some sort of synchronous crossbar with FIFOs to handle clock skew or offset, and not a clock-forwarded bus or the like. If so, that makes it more difficult to allow for 2 more die, though not impossible.

That said, I‘m sure we have to wait for M2 before we see something like 4 SoC’s tiled.

T6001 aka M1 Max is designed to link exactly 2 die and no more. Hector Martin (the Asahi Linux project lead) found the smoking gun a few months ago while reverse engineering the T6000/T6001 interrupt controller - it has provisions for multi-die systems, but the die count is limited to 2.

irqchip/apple-aic: Add support for AICv2 [LWN.net]

lwn.net

 

[mi7chy]

Which begs the question what benchmarks they were referring to. Assuming that the M1 Ultra scales ideally at 2x compared to M1 Max, and they haven't increased the maximum clock (meaning the single thread performance will be the same as M1 Max), we can extrapolate the approximate Geekbench results:

M1 Max: ~1700 single-thread, ~12000 multi-thread
M1 Ultra est.: ~1700 ST, ~24000 MT

For comparison:
i9-12900K: ~1900 ST, ~18000 MT

Nice bump in multi-threaded performance, but you pay a steep price. I wonder what the yields for the M1 Ultra are. Going from the M1 Max to the M1 Ultra Studio adds $2000. Even if you take off a few hundred bucks for 32 GB additional RAM and 512 GB additional SSD, that leaves well over $1000 just for essentially adding an additional M1 Max ...

That seems low for 12900k.

https://www.techspot.com/review/2351-intel-core-i9-12900k/

12700KF
https://www.techspot.com/review/2352-intel-core-i7-12700kf/

 

[MayaUser]

So everything needs to be written in metal to take advance of that M1 ultra...otherwise it will work at 32 gpu core ?!

 

[Andropov]

That seems low for 12900k.

Geekbench, not Cinebench.

So everything needs to be written in metal to take advance of that M1 ultra...otherwise it will work at 32 gpu core ?!

When did they said that?

 

[leman]

That seems low for 12900k.

https://www.techspot.com/review/2351-intel-core-i9-12900k/

12700KF
https://www.techspot.com/review/2352-intel-core-i7-12700kf/

Nobody cares about Cinebench. It’s relevance for real world performance is zero.

 

[leman]

So everything needs to be written in metal to take advance of that M1 ultra...otherwise it will work at 32 gpu core ?!

No, if it works today it with any M1 it will work with Ultra.

 

[oz_rkie]

Looks like a pretty impressive m1 ultra launch but at a very very steep premium. As good as the m1 max and ultra chips seem, they do sort of expose one of the downsides of the SoC based approach. While this is great for consumers who will make full use of both the CPU and GPU resources, its a sub optimal situation for consumers that are only interested in making use of the amazing CPU resources but have no use of so many GPU cores, you still end up paying an insane premium for it, if you want the m1 ultra. $7500 for the decked out m1 ultra option in AUD prices, pretty insane.

The display though. Its impressive looking and I am sure the image quality will be great but $2500 (AUD) for a 27" 5k Monitor? Not even 32", its not even mini-led. Would people be interested in buying this purely as a good monitor option? I mean the speakers and the camera are alright additions I guess but if you just need a good monitor, there are so many options in 2022 especially with the incoming OLED options this year.

 

[Gnattu]

That seems low for 12900k.

https://www.techspot.com/review/2351-intel-core-i9-12900k/

12700KF
https://www.techspot.com/review/2352-intel-core-i7-12700kf/

Why do you like Cinebench so much? The rendering library Cinebench uses is called Embree, a library developed by Intel and only have AVX native implementation. To use it on arm64, you have to use an avx-to-neon translation layer and it hurts the performance and will result in a lower score. The "arm64 native" is not the same level native as it is on x86, which has hand-crafted SIMD optimization.

 

[leman]

Why do you like Cinebench so much? The rendering library Cinebench uses is called Embree, a library developed by Intel and only have AVX native implementation. To use it on arm64, you have to use an avx-to-neon translation layer and it hurts the performance and will result in a lower score. The "arm64 native" is not the same level native as it is on x86, which has hand-crafted SIMD optimization.

Well, that’s easy to answer. Because Cinebench is one of the few benchmarks that favors x86.

 

[mi7chy]

So everything needs to be written in metal to take advance of that M1 ultra...otherwise it will work at 32 gpu core ?!

Same also applies to CPU to utilize 20 threads.

 

[MayaUser]

Looks like a pretty impressive m1 ultra launch but at a very very steep premium. As good as the m1 max and ultra chips seem, they do sort of expose one of the downsides of the SoC based approach. While this is great for consumers who will make full use of both the CPU and GPU resources, its a sub optimal situation for consumers that are only interested in making use of the amazing CPU resources but have no use of so many GPU cores, you still end up paying an insane premium for it, if you want the m1 ultra. $7500 for the decked out m1 ultra option in AUD prices, pretty insane.

The display though. Its impressive looking and I am sure the image quality will be great but $2500 (AUD) for a 27" 5k Monitor? Not even 32", its not even mini-led. Would people be interested in buying this purely as a good monitor option? I mean the speakers and the camera are alright additions I guess but if you just need a good monitor, there are so many options in 2022 especially with the incoming OLED options this year.

remember how much the mac pro costs....with afterburner and after look at the mac studio
Amen

 

[Rigby]

Why do you like Cinebench so much?

Presumably for the same reason that Apple fans simply discount it.

The rendering library Cinebench uses is called Embree, a library developed by Intel and only have AVX native implementation. To use it on arm64, you have to use an avx-to-neon translation layer and it hurts the performance and will result in a lower score. The "arm64 native" is not the same level native as it is on x86, which has hand-crafted SIMD optimization.

There is still no evidence that this causes the rather big discrepancy. Apple is active in the Embree project, so nothing prevents them from "hand-crafting" optimizations for the M1. I think it's more likely that Apple's SIMD instructions (which are only half as wide as Intel's AVX 2) are simply not as fast, perhaps because Apple prefered to spend the chip real estate on other things (such as their highly specialized video encoders).

 

[mi7chy]

Why do you like Cinebench so much?

All the news sources I follow don't include Geekbench garbage so just a habit seeing lesser garbage Cinebench R23. Actually prefer the industry standardize on Blender. What dinosaur still renders with CPU instead of GPU?