Question Intel Mont thread

[igor_kavinski]

https://store.acer.com/en-us/aspire-3-laptop-a315-510p-3905

First Gracemont laptop available for sale.

Anybody got disposable $500 to buy and test this laptop?

 

[511]

hmm https://chipsandcheese.com/p/skymont-in-desktop-form-atom-unleashed ?

There is no SpecInt Graph 🙁btw I forgot about the Ring Skymont article from Chips and 🧀 🤣.

 

[Doug S]

I'd be surprised if the ARM64 and x86/64 schedulers weren't pretty separate.

I don't know how Microsoft does things but in Linux there aren't dozens of different schedulers for all the architectures it supports. It is a single scheduler that supports them all (well two schedulers if you count the default and the real time scheduler, and maybe there are others you can choose for special cases like batch loads) You can make that work if you have tunables that can be set up to represent the CPU, i.e. stuff like NUMA boundaries, number of SMT threads, different core types with "weights" for their relative capabilities and so forth.

Given that the Windows kernel in its earlier incarnations as Windows NT has supported a lot of different architectures and its original architect was big on building things for portability I'd be surprised if Microsoft doesn't do something similar. Of course both ISAs running the same scheduler doesn't mean it would work as well, its comes down to whether they willing to devote effort to do fine tuning on the niche platform. They are probably devoting more resources towards insuring their scheduler is performing optimally when a new mass market x86 platform arrives like say Zen 5.

 

[soresu]

I don't know how Microsoft does things but in Linux there aren't dozens of different schedulers for all the architectures it supports. It is a single scheduler that supports them all

I'd be surprised if MS does things the same way.

They are famously very lumbering about making significant changes to their scheduler, so it's clearly not exactly the well oiled machine that Linux has.

Also the thing about Linux is that it works on OSS best principles of reusing as much code as possible to reduce bloat and increase efficiency.

Have you ever known MS to miss a chance to bloat Windows?

 

[MS_AT]

They are famously very lumbering about making significant changes to their scheduler, so it's clearly not exactly the well oiled machine that Linux has.

I am afraid the greatness of Linux scheduler is overstated by the same degree Windows one is understated.

Also the thing about Linux is that it works on OSS best principles of reusing as much code as possible to reduce bloat and increase efficiency.

Have you ever known MS to miss a chance to bloat Windows?

Your Linux will not be bloated if you build and configure it yourself You can't do that with Windows.

Also then you have articles like this:

谈谈Linux与ITMT调度器与多簇处理器 | David Huang's Blog

So your mileage may wary, and I wouldn't be so quick to dismiss Windows kernel just because Windows Home edition is the test vehicle for enterprise users. I mean the execs decision make the final experience worse, but it's not necessarily the kernel people fault. Not to mention things that Windows got right, like DLLs vs shared libraries, etc.

There is no SpecInt Graph

You mean score vs power? Yes, my bad, was thinking you wanted to see general scores

 

[soresu]

Your Linux will not be bloated if you build and configure it yourself You can't do that with Windows.

I mean, you can do it with Windows (or at least it used to be possible), it's just a helluva lot harder than with Linux.

 

[511]

Well I actually have a answer File windows 11 install but that can't turn the react native search bar into native UI 🙁.

 

[moinmoin]

Not to mention things that Windows got right, like DLLs vs shared libraries, etc.

How is the versioning nightmare of DLLs something Windows got right?

 

[MS_AT]

How is the versioning nightmare of DLLs something Windows got right

No interposition. No lazy binding with GOT open for writes for program duration. Of course you can opt-out from defaults. I am new to the subject, might be I missed some thing, but when trying to get up to speed I have heard rather negative opinion on current defaults.

 

[DavidC1]

Also @Thunder 57 Raichu leaked a while ago about 20% over Darkmont for arctic wolf.

Raichu was not expecting much for Skymont actually.

As I said, performant yes in perf/mm² metric.

But in perf/watt they are significantly worse than the *Cove cores.

While the *mont cores were originally targeted for a more balanced PPA type efficiency along the lines of Cortex A7x/7xx, since Gracemont they are targeted at adding compute density per mm² specifically in order to increase core counts per package.

Skymont seem noticeably better on power efficiency part looking at Lunarlake results.

Also, 30%+ better in Int and 60-70% better in FP per clock cannot be ignored, if Gracemont is the experience you had with the most modern cores.

As to underselling it I've actually used Gracemont laptop only a couple of years ago as a work computer and it was miserable even doing basic Google apps, so I've got good reason to be unimpressed as the presentation for the 6 wide Tremont sounded very impressive too.

Intel said 32% per clock over the previous generation so if people were expecting something way beyond that, then the fault is on them. Also the core size was under 1mm2 so it indeed was impressive.

 

[Io Magnesso]

Raichu was not expecting much for Skymont actually.

Skymont seem noticeably better on power efficiency part looking at Lunarlake results.

Also, 30%+ better in Int and 60-70% better in FP per clock cannot be ignored, if Gracemont is the experience you had with the most modern cores.

Intel said 32% per clock over the previous generation so if people were expecting something way beyond that, then the fault is on them. Also the core size was under 1mm2 so it indeed was impressive.

By the way, which one is better, Skymont or Cortex-A725?
Personally, I think Skymont is the most powerful of the Little Core in 2024.
Skymont has excellent PPA and efficiency from the point of view of architecture.

It can be applied to any size from Lunar Lake's LP-E Core to Arrow Lake's e-core. I think there is also scalability I'm looking forward to Clearwater Forest, equipped with Improved Skymont, Darkmont

 

[DavidC1]

The Clearwater Forest version of Darkmont isn't indicating any architectural enhancements.

Sierra Forest version of Crestmont stayed with the 5-wide backend as Gracemont while the client version expanded it to 6. I think the slight differentiation between server and client is continuing here.

 

[511]

The Clearwater Forest version of Darkmont isn't indicating any architectural enhancements.

Sierra Forest version of Crestmont stayed with the 5-wide backend as Gracemont while the client version expanded it to 6. I think the slight differentiation between server and client is continuing here.

17 Cycle L2 Darkmont vs 19 cycle L2 for skymont is a bid difference

 

[igor_kavinski]

17 Cycle L2 Darkmont vs 19 cycle L2 for skymont is a bid difference

Woah rad dude! (Keanu Reeves can't believe it!)

Seriously, that's great if it actually happens. Latency is the destroyer of our dreams.

 

[LightningZ71]

That L2 latency needs to be amazing given that it's shared between 4 cores and the number of hops after it to get to the L3 in the Xeon products...

 

[DZero]

By the way, which one is better, Skymont or Cortex-A725?
Personally, I think Skymont is the most powerful of the Little Core in 2024.
Skymont has excellent PPA and efficiency from the point of view of architecture.

It can be applied to any size from Lunar Lake's LP-E Core to Arrow Lake's e-core. I think there is also scalability I'm looking forward to Clearwater Forest, equipped with Improved Skymont, Darkmont

Skymont feels that is more like an X2 or X3 core tier instead of an A tier core.

Still waiting for A730 to see how much improves.

 

[DavidC1]

17 Cycle L2 Darkmont vs 19 cycle L2 for skymont is a bid difference

11% latency difference being called big is a hyperbole. It'll result in maybe 2% in the optimistic scenario. Also, it's not a core change. When ring bus enabled L3 latency to go from 33 cycles to 25 cycles on Sandy Bridge, that was decent. "Big" is if you halve the latency. There are many instances in CPU where you can literally double something but get a fraction of a fraction of gains. That's why the inverse square law applies without adding new features brought on by innovation.

 

[igor_kavinski]

谈谈Linux与ITMT调度器与多簇处理器 | David Huang's Blog

Most of these problems have probably been fixed by now. I have asked him for an update though.

 

[igor_kavinski]

The Intel Core 3 N355 Update Over the Core i3-N305

The Intel Core 3 N355 is an 8-core processor, a 2025 update to the 2023 era Intel Core i3-N305. We get into the N355 performance advantage

www.servethehome.com

I respect engineers but when they are complicit in watering down RAM performance, I am left feeling pretty miffed: https://forums.anandtech.com/thread...d-against-core-i7-12700.2610597/post-41497748

 

[igor_kavinski]

Intel isn't gonna bother with such tricks on such a low end part. Also, if the TDP is low that's going to keep all core clocks relatively low as well. Still I do expect it'll be a faster CPU overall than Alderlake/Twinlake-N.

Yeah something like 4 GHz. Not the 5 GHz we deserve

 

[igor_kavinski]

Next year it might actually be worth spending $300 on a Wildcat Lake laptop.

 

[511]

Wildcat lake will have 2 Integrated TB4 btw and manufacturers will find a way to cheap out

 

[DZero]

Intel should bring dual channel back to that tier.

 

[511]

Single Channel DDR56400 iirc

 

[hemedans]

Skymont feels that is more like an X2 or X3 core tier instead of an A tier core.

Still waiting for A730 to see how much improves.

in lunar lake its more of A than X, it scale below 1W and match Apple E cores (M2).

 

[coercitiv]

in lunar lake its more of A than X, it scale below 1W and match Apple E cores (M2).

In LL is operating at reduced clocks and without access to proper L3. The SLC is mostly there to save on power, not much faster than memory. That being said, Skymont is probably hard to pin to X or A roles because it's somewhere in between even when placed in the desktop cluster.