Alder Lake

[Joelist]

So they MAY have a chip that MAY be 10% faster on synthetic benchmarks while drawing over three times the power and not a word on graphical performance?

 

[Yoused]

So they MAY have a chip that MAY be 10% faster on synthetic benchmarks while drawing over three times the power and not a word on graphical performance?

Alder Lake is not a UMA-design SoC: Intel is expecting most use cases to employ an eGPU, so graphical performance is less likely to be intrinsic. In addition, an article from August suggested that consumer-level CPUs will have AVX512 and AMX (matrix logic for ML) but it will be fused off (non-reachable) except in enterprise/server models.

In other words, comparing the rich capabilities of M1 to the baseline capabilities of AL is like comparing a fisherman's pontoon boat to an Alaska Marine Highway multi-vehicle ferry with staterooms: the former might be faster across the lake, but getting a lot of people and cars up the inside passage may not be a thing it can accomplish efficiently on its own without help.

 

[jbailey]

Intel is reportedly telling motherboard/bios vendors to disable unsupported AVX-512 instructions in a microcode update. Previously Anandtech reported that you could disable the efficiency cores which don’t have AVX-512 hardware and enable AVX-512 in the BIOS for the performance cores.

 

[Andropov]

Leaked Alder Lake 12900HK benchmark.

Looks like its 2-3% faster than M1 Max in single core, 10-15% faster than M1 Max in multicore (judging by Geekbench results). PL2 limited to 115W, but the exact energy efficiency is uncertain as of now. Doesn't look very promising though, looks like they aimed for surpassing the competition in performance regardless of power consumption. For example, it's 30% faster than last year's AMD 5900HX at more than twice the power consumption (and the AMD chip is running on slower RAM too).

 

[Cmaier]

Leaked Alder Lake 12900HK benchmark.

Looks like its 2-3% faster than M1 Max in single core, 10-15% faster than M1 Max in multicore (judging by Geekbench results). PL2 limited to 115W, but the exact energy efficiency is uncertain as of now. Doesn't look very promising though, looks like they aimed for surpassing the competition in performance regardless of power consumption. For example, it's 30% faster than last year's AMD 5900HX at more than twice the power consumption (and the AMD chip is running on slower RAM too).

That’s what the evidence seems to be pointing to. They just moved to the right on the perf/power curve.

 

[Yoused]

For example, it's 30% faster than last year's AMD 5900HX at more than twice the power consumption (and the AMD chip is running on slower RAM too).

Sounds like that would put Ryzen pretty close to M1 in perf/power – at least until you factor in the GPU. I think I read that the M1 GPU can pull as much as 100 watts at times.

 

[Andropov]

Lol, I went back to the source to check their numbers and couldn't figure how they got the '30% faster than 5900HX'. Finally it looks like what they did was:
- Assume that ALL CineBench tests (R15 and R23 specifically) use the same CPU package power.
- Intel 12900HK got 2697 points in CineBench R15 (plugged in).
- Intel 12900HK used an average of 65.4W in CineBench R23 (plugged in).
- AMD 5900HX got 2086 points in CineBench R15 (plugged in).
- AMD 5900HX used an average of 29.3W in CineBench R15 (plugged in).

The problem is that there's no data for how much CPU package power the 12900HK used in Cinebench R15, just on Cinebench R23. IF both Cinebench tests average about the same CPU package power: Intel 12900HK gets ~40 Cinebench R15 points per watt, AMD 5900HX gets ~70 Cinebench R15 points per watt. So +29% faster using +123% more power (not 23% more, 123% more).

That's a big if, though.

Also, all scores get cut in half (literally) when unplugged. So, on battery, the 12900HK barely outperforms the regular M1 (not Pro/Max).

 

[Cmaier]

Lol, I went back to the source to check their numbers and couldn't figure how they got the '30% faster than 5900HX'. Finally it looks like what they did was:
- Assume that ALL CineBench tests (R15 and R23 specifically) use the same CPU package power.
- Intel 12900HK got 2697 points in CineBench R15 (plugged in).
- Intel 12900HK used an average of 65.4W in CineBench R23 (plugged in).
- AMD 5900HX got 2086 points in CineBench R15 (plugged in).
- AMD 5900HX used an average of 29.3W in CineBench R15 (plugged in).

The problem is that there's no data for how much CPU package power the 12900HK used in Cinebench R15, just on Cinebench R23. IF both Cinebench tests average about the same CPU package power: Intel 12900HK gets ~40 Cinebench R15 points per watt, AMD 5900HX gets ~70 Cinebench R15 points per watt. So +29% faster using +123% more power (not 23% more, 123% more).

That's a big if, though.

Also, all scores get cut in half (literally) when unplugged. So, on battery, the 12900HK barely outperforms the regular M1 (not Pro/Max).

Intel is going to look pretty dumb when these things are out in the wild and we can run real tests and put numbers out there.

 

[Souko]

To the power consumption. My M1 Pro 8 core (6+2) draws about 20.5W average (6 cores 20W and 2 effeciency cores 0.5W) average in Cinebench R23 and score was 9540. (CPU only - no RAM, no GPU,..) So M1 Pro/Max with 10 core CPU will draw about 27W...

In low power mode it was about 12.2W CPU only (probably around 16W for 10-core) and score for 8core M1 Pro was 7904 for me.

Power consumption was measured using Terminal - sudo powermetrics.

 

[Cmaier]

To the power consumption. My M1 Pro 8 core (6+2) draws about 20.5W average (6 cores 20W and 2 effeciency cores 0.5W) average in Cinebench R23 and score was 9540. (CPU only - no RAM, no GPU,..) So M1 Pro/Max with 10 core CPU will draw about 27W...

In low power mode it was about 12.2W CPU only (probably around 16W for 10-core) and score for 8core M1 Pro was 7904 for me.

Power consumption was measured using Terminal - sudo powermetrics.

Welcome aboard, by the way!

 

[Nycturne]

Sounds like that would put Ryzen pretty close to M1 in perf/power – at least until you factor in the GPU. I think I read that the M1 GPU can pull as much as 100 watts at times.

As far as I know, that would be full package power with both CPU and GPU loaded. That said, I've yet to get the GPU on an M1 Max to pull the 55-60W Apple claims outside of a couple specific benchmark scenarios. Lots of underutilization in the (admittedly small and Rosetta-dependent) selection of stuff I've tried so far.

 

[Souko]

Welcome aboard, by the way!

Thanks!

As far as I know, that would be full package power with both CPU and GPU loaded. That said, I've yet to get the GPU on an M1 Max to pull the 55-60W Apple claims outside of a couple specific benchmark scenarios. Lots of underutilization in the (admittedly small and Rosetta-dependent) selection of stuff I've tried so far.

OK, so I ran some tests for GPU. I have MBP14" base, 14-core GPU. Power consumption was measured using Terminal - sudo powermetrics.
I ran 3D Mark Wild Life Extreme, unlimited and GFXBench Metal - Aztec Ruins 1440p, offscreen. Average power draw was 19.5W for only GPU in both test. So estimate is around 45W for 32-core M1 Max.


I got 47W max on M1 Pro with GFXBench + Cinebench R23 (21W GPU + 20W CPU + 2W DRAM + 4W other).

I would say maximum power that M1 Max takes is about 81W. (27W CPU + 45W GPU + 4W DRAM (not sure if 32GB is consuming two times 16GB power) + 4W other (again not sure about it + 1-2W for maximum (not average) power draw.)) In some specific tests probably more.

 

[casperes1996]

In the very first post here, Cmaier talked about the AVX-512 capability of the P cores. Well, not just is it disabled in BIOS, Intel has said they're looking into ways of ensuring it cannot be enabled at all, even if you disable the E-cores. Don't know why they would potentially want to do that but they've said it, so. Think I read it on Anandtech, may have been GN or Hardware Unboxed though.

I think Alder Lake will also be rather interesting on the Linux front as patches roll in to use Intel's Thread Director thingy and see how the Linux kernel might utilise it and perform AMP scheduling - Has the Windows NT Kernel ever had to deal with different core types like this? The ARM cores Windows have supported in the past have they even been big.Little? Just looked up the Snapdragon 8c and it looks like that's just 8 cores of one type.

Do we know for a fact that enabling AVX-512 in the BIOS requires disabling the E-cores, or is it just that Windows can't handle it both being supported and not at the same time in the scheduler (yet)? Because in that case Linux could catch illegal AVX-512 on E cores and schedule those jobs on P cores though there's a lot of decisions to be made regarding the logic for moving threads, their permanency on P cores after encountering AVX-512 for the first time and potentially not for a long time again after that and so on, but still

 

[Cmaier]

In the very first post here, Cmaier talked about the AVX-512 capability of the P cores. Well, not just is it disabled in BIOS, Intel has said they're looking into ways of ensuring it cannot be enabled at all, even if you disable the E-cores. Don't know why they would potentially want to do that but they've said it, so. Think I read it on Anandtech, may have been GN or Hardware Unboxed though.

I think Alder Lake will also be rather interesting on the Linux front as patches roll in to use Intel's Thread Director thingy and see how the Linux kernel might utilise it and perform AMP scheduling - Has the Windows NT Kernel ever had to deal with different core types like this? The ARM cores Windows have supported in the past have they even been big.Little? Just looked up the Snapdragon 8c and it looks like that's just 8 cores of one type.

Do we know for a fact that enabling AVX-512 in the BIOS requires disabling the E-cores, or is it just that Windows can't handle it both being supported and not at the same time in the scheduler (yet)? Because in that case Linux could catch illegal AVX-512 on E cores and schedule those jobs on P cores though there's a lot of decisions to be made regarding the logic for moving threads, their permanency on P cores after encountering AVX-512 for the first time and potentially not for a long time again after that and so on, but still

I can’t think of any heterogeneous cores that the NT scheduler has had to deal with previously. I may be forgetting something. It’s a bit of an interesting computer science problem figuring out how you would dispatch threads so as to avoid the problem, unless you have some facility in the OS where threads that need AVX-512 are marked somehow.

 

[mr_roboto]

In the very first post here, Cmaier talked about the AVX-512 capability of the P cores. Well, not just is it disabled in BIOS, Intel has said they're looking into ways of ensuring it cannot be enabled at all, even if you disable the E-cores. Don't know why they would potentially want to do that but they've said it, so. Think I read it on Anandtech, may have been GN or Hardware Unboxed though.

One possible reason which has come to my mind: they may be enhancing yield by not scrapping parts with defects somewhere in the upper 256 bits of the AVX execution units. Wouldn't look good if some random subset of the customers who disabled all E cores to enable AVX512 had it blow up in their face.

The way they're disabling it is a microcode update, BTW.

 

[Colstan]

So, over at the "other place" mother ship, the early results for i9 mobile Alder Lake are trickling in. The short version is that, yes, the i9 does outperform the M1 Max...by about 4%. The tradeoff is that the i9 runs at 100w, spiking to 140w, while the Max tops out at 40w. On battery, the i9 gets six hours of video playback, while the Max gets 21. The MSI laptop tested is over an inch-thick and weighed one-third more than the 16-inch MacBook Pro.

Is is just me, or did Intel intentionally fine tune this particular i9 just to edge out the M1 Max in performance, forget all other considerations? A 4% performance improvement isn't something that the typical user is going to notice in day-to-day use, but heat, battery life, weight, energy bills, and fan noise are.

As a desktop user, most of those issues are irrelevant to me, except fan noise. I optimize for quiet computing because I have sensitive hearing, so Apple Silicon is a huge bonus, in that regard. Of course, if you are primarily a macOS user, Alder Lake and the x86 side are mainly a curiosity, but the horse race is an amusing side show.

 

[Yoused]

i9 is the 12K Xeon-ish. Most realistic notebooks will have an i5 or maybe i3, which will probably have better stamina at the cost of less mega-juice but be no real match for a '20 M1. If Apple produces a 20 core SoC with a ridiculous GPU (say, 64 cores) for the desk, Alder Lake will be knee deep in trying to keep up.

 

[Andropov]

Is is just me, or did Intel intentionally fine tune this particular i9 just to edge out the M1 Max in performance, forget all other considerations? A 4% performance improvement isn't something that the typical user is going to notice in day-to-day use, but heat, battery life, weight, energy bills, and fan noise are.

Yeah, I think it's been pretty obvious that the i9 12900HK was specifically designed and tuned to beat the M1 Max, no matter what.

 

[Cmaier]

Looks like taking battery sizes into account that alder lake essentially ties M1 Max but needs around 2.5-3x the power to get there, and with battery life around ¼ (not quite sure how the numbers work for that, but we’ll get more data soon). Add in fan noise differential to boot.

 

[casperes1996]

Looks like taking battery sizes into account that alder lake essentially ties M1 Max but needs around 2.5-3x the power to get there, and with battery life around ¼ (not quite sure how the numbers work for that, but we’ll get more data soon). Add in fan noise differential to boot.

The only device I know that has been tested with 12900HK is that Raider laptop. According to Anandtech the Nvidia 3080 Laptop GPU in it, is not very good at power-gating when not in demand, which contributes to power draw as well.
Though on the flipside of that, even though it's a 17" display, it's only 1080p and Anandtech normalises battery tests at 200 nits, but the display can't go much above that either. So the MacBook Pro display probably draw quite a bit more power, at least when the dimming zones aren't mostly off or something. There's also the consideration that M1 Max power is often reported as package power, which includes the memory and Intel numbers would not.

So when talking about the battery life, well, many more aspects than just the CPU at play. Spinning the fans harder isn't free either I guess