Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

[Tigerick]

As Hot Chips 34 starting this week, Intel will unveil technical information of upcoming Meteor Lake (MTL) and Arrow Lake (ARL), new generation platform after Raptor Lake. Both MTL and ARL represent new direction which Intel will move to multiple chiplets and combine as one SoC platform.

MTL also represents new compute tile that based on Intel 4 process which is based on EUV lithography, a first from Intel. Intel expects to ship MTL mobile SoC in 2023.

ARL will come after MTL so Intel should be shipping it in 2024, that is what Intel roadmap is telling us. ARL compute tile will be manufactured by Intel 20A process, a first from Intel to use GAA transistors called RibbonFET.





Intel Core Ultra 100 - Meteor Lake



As mentioned by Tomshardware, TSMC will manufacture the I/O, SoC, and GPU tiles. That means Intel will manufacture only the CPU and Foveros tiles. (Notably, Intel calls the I/O tile an 'I/O Expander,' hence the IOE moniker.)

 

[igor_kavinski]

Could you please clarify what you mean by this? Also, note that the intention is to run it with TDP max limit set to approximately 100 W.

Regarding 385K, what fixes are expected?

Read the section "Quick and simple tips for enthusiasts" here: https://www.tomshardware.com/pc-com...-set-world-records-and-pushed-it-to-the-limit

Just 100W TDP limit? Intel desktop architecture is too power hungry to give great performance at just 100W. Maybe if you went with non-K 285: https://www.intel.com/content/www/u...-36m-cache-up-to-5-60-ghz/specifications.html

But 9700X would be much better and you gain AVX-512 with it for "free".

385K, I think they may raise the NGU, ring and D2D clocks so it performs better. They might improve the IMC even more so that it can run DDR5-10000 with XMP just like that (they usually improve the IMC with a CPU refresh. Raptor Lake is able to do 8000+ while Alder Lake barely goes to 7000 MT/s RAM speed). And of course, 385K will have the 48 TOPS NPU.

 

[511]

Read the section "Quick and simple tips for enthusiasts" here: https://www.tomshardware.com/pc-com...-set-world-records-and-pushed-it-to-the-limit

Just 100W TDP limit? Intel desktop architecture is too power hungry to give great performance at just 100W. Maybe if you went with non-K 285: https://www.intel.com/content/www/u...-36m-cache-up-to-5-60-ghz/specifications.html

But 9700X would be much better and you gain AVX-512 with it for "free".

385K, I think they may raise the NGU, ring and D2D clocks so it performs better. They might improve the IMC even more so that it can run DDR5-10000 with XMP just like that (they usually improve the IMC with a CPU refresh. Raptor Lake is able to do 8000+ while Alder Lake barely goes to 7000 MT/s RAM speed). And of course, 385K will have the 48 TOPS NPU.

It's not raptor lake at 100W ARL is the better option vs Zen 5 simply due to the fact that AMDs package power takes quite a bit of power budget the only advantage with zen 5 is AVX-512 for transcoding Intel has quick sync as well.
You can upgrade Zen 5 to 6 but Intel doesn't have upgrade path.

 

[MS_AT]

The intention is to use it mainly for SW compilation and video transcoding.

Cinebench does not correlate well with these to be honest. Generally speaking for code compilation due to homogeneous cores 9950x will give you less headaches. But this depends on how exotic your environment might be. Ideally you would be able to test your workload on both and pick the one that fits better.

If you are considering limiting yourself to 100W, then maybe consider Framework Desktop (or just a mobo) with Strix Halo, it solves the IOD idle power draw issue, and gives you a chance to run small models for code completion etc. On the other hand it is fairly closed platform.

 

[Fjodor2001]

Just 100W TDP limit? Intel desktop architecture is too power hungry to give great performance at just 100W. Maybe if you went with non-K 285: https://www.intel.com/content/www/u...-36m-cache-up-to-5-60-ghz/specifications.html

Not according to this which I posted earlier:

As can be seen, it's quite the opposite. 285K performs very well below 100 W. Better than e.g. 14900K / 9950X(3D). In fact 100 W seems to be the sweet spot for 285K. That was why I raised it as a suitable option for my use case to begin with.

 

[Fjodor2001]

Cinebench does not correlate well with these to be honest. Generally speaking for code compilation due to homogeneous cores 9950x will give you less headaches.

For SW compilation, why do you think the heterogeneous P/E cores in 285K would cause a problem? Anything you can link to?

For video transcoding, which is my other main use case, 285K performs very well at least according to this:

If you are considering limiting yourself to 100W, then maybe consider Framework Desktop (or just a mobo) with Strix Halo, it solves the IOD idle power draw issue, and gives you a chance to run small models for code completion etc. On the other hand it is fairly closed platform.

In my case I'm not interested in soldered CPU solutions. And I don't think there is any socketed solution available for Strix Halo. Which is unfortunate, since I agree it would likely be a good solution for my use case otherwise (low idle power consumption, good MT performance at ~100 W TDP (but not sure if better than 285K?), and also very nice iGPU as a bonus ).

 

[alcoholbob]

Not according to this which I posted earlier:

As can be seen, it's quite the opposite. 285K performs very well below 100 W. Better than e.g. 14900K / 9950X(3D). In fact 100 W seems to be the sweet spot for 285K. That was why I raised it as a suitable option for my use case to begin with.

This is 100W at the chip, not 100W at the wall. The DVLR resistor is soaking up a good 30-40W. An actual apples to apples comparison would be to just add 30-40W to those graph numbers.

 

[alcoholbob]

Read the section "Quick and simple tips for enthusiasts" here: https://www.tomshardware.com/pc-com...-set-world-records-and-pushed-it-to-the-limit

Just 100W TDP limit? Intel desktop architecture is too power hungry to give great performance at just 100W. Maybe if you went with non-K 285: https://www.intel.com/content/www/u...-36m-cache-up-to-5-60-ghz/specifications.html

But 9700X would be much better and you gain AVX-512 with it for "free".

385K, I think they may raise the NGU, ring and D2D clocks so it performs better. They might improve the IMC even more so that it can run DDR5-10000 with XMP just like that (they usually improve the IMC with a CPU refresh. Raptor Lake is able to do 8000+ while Alder Lake barely goes to 7000 MT/s RAM speed). And of course, 385K will have the 48 TOPS NPU.

Problem with raising NGU, D2D, and ring to bring it up to “competitive” with the Raptor Lake is it ends up losing power efficiency to Raptor Lake. De8bauer had to go balls to the wall with Direct Die cooling and maxing out everything just to barely beat a stock 14900KS. The thing is the moment you tune load lines and undervolt a 14th gen processor, you can cut power consumption by a good 30-35% while gaining another ~20% in performance. Which means all that screwing around with Arrow Lake with expensive CUDIMM and overly-complex tuning was a waste of time, it's still 20% slower and uses significantly more power (when you add the additional power losses from DVLR).

 

[igor_kavinski]

De8bauer had to go balls to the wall with Direct Die cooling and maxing out everything just to barely beat a stock 14900KS.

I think that was before the microcode and BIOS updates. Currently, some users with good 360 AIOs (not custom cooling) are getting 39 seconds in y-cruncher while my 9950X3D's best so far is 43 seconds at 7200C34. Not sure about games performance. 285K can't be an all round performer due to its poor L3 latency but maybe 385K can break that curse.

 

[alcoholbob]

I think that was before the microcode and BIOS updates. Currently, some users with good 360 AIOs (not custom cooling) are getting 39 seconds in y-cruncher while my 9950X3D's best so far is 43 seconds at 7200C34. Not sure about games performance. 285K can't be an all round performer due to its poor L3 latency but maybe 385K can break that curse.

Yeah for productivity Raptor Lake can't compete with ARL because of the weak e-cores, that's not going to change at any time. Also Arrow Lake has an advantage of having less threads, as not a hole lot of software can even saturate 32 threads, so it can hit higher numbers in productivity benchmarks that are more thread limited. The microcode update from my understanding was mostly to fix gaming performance. The overall uplift was low, because the baseline uplift was mainly just making sure the balanced power profile in Windows matched the high performance profile. However some games got a huge uplift, like Cyberpunk 2077.



Cyberpunk is basically the best game Arrow Lake has to show off right now. This looks great on paper, until you realize you can tune both Arrow Lake and Raptor Lake. And if you tune both (even giving ARL the leg up with much faster and more expensive RAM), a 14900K still ends up being quite a bit faster a 285K, only trailing the 9800X3D by a few percentage points.

 

[511]

This is 100W at the chip, not 100W at the wall. The DVLR resistor is soaking up a good 30-40W. An actual apples to apples comparison would be to just add 30-40W to those graph numbers.

DLVR soaks at high power it sucks lot less at low power basically the amount it sucks is proportional to the voltage hence the power.

 

[MS_AT]

For SW compilation, why do you think the heterogeneous P/E cores in 285K would cause a problem? Anything you can link to?

From personal experience: we are exploring a move to Windows on Windows containers to host our builds, where one of the isolation methods simply fails to use the P cores at all. While the other one behaves just fine (a similar story to https://github.com/microsoft/Windows-Containers/issues/397 but I am not sure what they are building there). Certain hypervisors will refuse to use P cores until you disable E cores (like VMware afaik) what could be or not be a problem depending on your use case.

Now, more esoteric problems are when the scheduler will decide to put a long running process that is dependency for next thing in job chain on E-cores for some strange measure, what can be observed from task manager and can bottleneck the whole build. I have just recently moved from 11850H to 13850HX and the clean build time of one of our targets increased from 150 to 300s. At the same time LLVM build time decreased, but the two are using different build systems and LLVM build is "heavier" what might make it easier to hide such dependencies (I have so far excluded AV interference but did not have time to investigate further).

Of course with 285K, the difference between P and E cores is smaller, so such mishaps might be less pronounced and being on desktop the power management config is a lot simpler so your mileage may vary. I guess if you plan on using fairly standard setup you will be perfectly fine, but ideally you would be able to test before committing to either setup.

For video transcoding, which is my other main use case, 285K performs very well at least according

They are not testing at 100W though.

 

[igor_kavinski]

I have just recently moved from 11850H to 13850HX and the clean build time of one of our targets increased from 150 to 300s.

Could be the horrible DDR5 RAM latency affecting the performance. Could also be really awful thermal throttling (one of the most oft-repeated complaints about Intel HX CPUs). May want to add a beefy laptop cooler with one of those huge fans to rule out the latter or maybe try running the workload in power saving mode with the laptop plugged in.

One other option is to use Process Lasso. It's really, really good at pegging workloads to specific cores (you can even force them to run on only HT threads).

 

[MS_AT]

Could be the horrible DDR5 RAM latency affecting the performance. Could also be really awful thermal throttling (one of the most oft-repeated complaints about Intel HX CPUs). May want to add a beefy laptop cooler with one of those huge fans to rule out the latter or maybe try running the workload in power saving mode with the laptop plugged in.

It's not a game for DRAM latency to matter so much, and since LLVM build, which I noted as heavier, is noticeably faster on the new machine, then whatever thermal throttling there is (and it's there, I have measured ) does not overcome sheer core advantage. I suspect the problem lies somewhere in the power management configuration and Dell drivers, but currently I don't have time to read through the documentation. Basically few of our custom programs responsible for generating intermediate files for the build seem to be always put on e-cores what penalizes the overall duration. Since these are old internal programs, might be we would need to recompile them and put some hints for Windows scheduler. Anyway debugging this here was not the purpose, it was just an illustration that you might run into corner cases with heterogeneous CPUs

 

[igor_kavinski]

Since these are old internal programs, might be we would need to recompile them and put some hints for Windows scheduler.

Instead of recompiling, if they are in a script, you can use the affinity command here: https://www.windowsdigitals.com/how-to-set-process-cpu-affinity-priority-permanently-in-windows-10/

 

[Fjodor2001]

This is 100W at the chip, not 100W at the wall. The DVLR resistor is soaking up a good 30-40W. An actual apples to apples comparison would be to just add 30-40W to those graph numbers.

But is this not (roughly) the same for all the compared CPUs? So the main take way still remains, i.e. that 285K performs well at 100 W TDP, and better than the CPUs it's compared to in that test.

 

[Fjodor2001]

They are not testing at 100W though.

Yeah, I know. But I figured it should give at least some indication of how well the CPU performs with video transcoding in general. If it performs well with video transcoding at peak performance / max TDP (as used in that test), and it has a sweet spot at around 100 W (after which more watts does not give much more perf), then I figure it's likely to have good video transcoding perf and perf/watt at 100W also.

I've not been able to find any comparison of video transcoding performance at ~100 W between 285K / 14900K / 9950X(3D). But if someone has found any such comparison, please share. Ideally for x265 video transcoding, since that's probably what I'll do most of.

 

[igor_kavinski]

I've not been able to find any comparison of video transcoding performance at ~100 W between 285K / 14900K / 9950X(3D). But if someone has found any such comparison, please share. Ideally for x265 video transcoding, since that's probably what I'll do most of.

I think the situation seems simple for you. 14900K is too power hungry and possibly unreliable in the long term. 9950X(3D) has too high idle power draw for your taste. So 285K is it.

 

[Sgraffite]

What about a 7945HX? Low power and inexpensive. Not sure about idle power.

https://www.amazon.com/MINISFORUM-Motherboard-BD795i-SE-PCIe4-0/dp/B0DQ8WXMKP/ref=sr_1_4?sr=8-4

 

[511]

What about a 7945HX? Low power and inexpensive. Not sure about idle power.

https://www.amazon.com/MINISFORUM-Motherboard-BD795i-SE-PCIe4-0/dp/B0DQ8WXMKP/ref=sr_1_4?sr=8-4

7945HX is rebranded 7950X get a 275HX Instead if you want Mobile CPU.

 

[Sgraffite]

7945HX is rebranded 7950X get a 275HX Instead if you want Mobile CPU.

What options are there for a 275HX in a desktop?

 

[511]

What options are there for a 275HX in a desktop?

Nothing as of rn 🥲 you can only get it in a laptop

 

[gdansk]

What about a 7945HX? Low power and inexpensive. Not sure about idle power.

https://www.amazon.com/MINISFORUM-Motherboard-BD795i-SE-PCIe4-0/dp/B0DQ8WXMKP/ref=sr_1_4?sr=8-4

I'm pretty sure at 100W the 285 should be faster for many workloads.

If anyone has a 285(K) and is eager to demonstrate its strengths here is a 100W PPT 7940HX to compare to https://browser.geekbench.com/v5/cpu/22976049

 

[Sgraffite]

It probably would, but no way are you beating it in price. If that is important.

 

[alcoholbob]

What about a 7945HX? Low power and inexpensive. Not sure about idle power.

https://www.amazon.com/MINISFORUM-Motherboard-BD795i-SE-PCIe4-0/dp/B0DQ8WXMKP/ref=sr_1_4?sr=8-4

Idle power isn't amazing with ARL or Zen 5, but sustained performance is good within a TDP window. However the actual power bill at the end of the year won't necessarily be great if the PC spends most of the day idling or running low power tasks. It's basically just a build for SFF for the sake of SFF so peak power stays within the cooling envelope. One thing Raptor (and Alder) Lake is really good at that chiplet based designs like Zen 5 or ARL aren't good at is actually low load performance. With RPL there's very little difference in package power and temps between idle versus low load scenarios like typing a spreadsheet or browsing the web, but this is actually the worst performance efficiency range for chiplet CPUs as Zen and ARL will spike up in temps and power consumption here. So if you do casual productivity tasks all day, funny enough Raptor Lake ends up being the most efficient here.

 

[Sgraffite]

Idle power isn't amazing with ARL or Zen 5, but sustained performance is good within a TDP window. However the actual power bill at the end of the year won't necessarily be great if the PC spends most of the day idling or running power tasks. It's basically just a build for SFF for the sake of SFF so peak power stays within the cooling envelope. One thing Raptor (and Alder) Lake is really good at that chiplet based designs like Zen 5 or ARL aren't good at is actually low load performance. With RPL there's very little difference in package power and temps in low load scenarios like typing a spreadsheet or browsing the web, but this is actually the worst performance efficiency range for chiplet CPUs as Zen and ARL will spike up in temps and power consumption here. So if you do casual productivity tasks all day, funny enough Raptor Lake ends up being the most efficient here.

I get that. You would need to idle at least a fair amount of decades I would think to make up the initial cost difference with your power bill.