News Intel Meteor Lake Architecture Added to Linux Kernel

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
>> " I think it would be appropriate to believe Meteor Lake will be noticeably quicker vs Intel's future 11th and 12th Gen CPUs. "

So you think the 13th gen will be at least somewhat faster than the ones which came before it? Really going out on a limb there, aren't you?
 

spongiemaster

Admirable
Dec 12, 2019
2,273
1,277
7,560
>> " I think it would be appropriate to believe Meteor Lake will be noticeably quicker vs Intel's future 11th and 12th Gen CPUs. "

So you think the 13th gen will be at least somewhat faster than the ones which came before it? Really going out on a limb there, aren't you?
Certainly true, but let's also keep in mind after generation 6 (skylake), generation 7 (Kaby Lake), generation 8 (coffee lake), generation 9 (coffee lake refresh), and generation 10 (comet lake) weren't faster at the architectural level.
 

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
Certainly true, but let's also keep in mind after generation 6 (skylake), generation 7 (Kaby Lake), generation 8 (coffee lake), generation 9 (coffee lake refresh), and generation 10 (comet lake) weren't faster at the architectural level.
I won't debate that -- but since the author prefaces his conclusion by referencing Intel's process change, rather than any architectural IPC improvements -- that's rather a moot point, isn't it? I'd struggle harder to find a safer bet than concluding Intel will see substantial performance gains from the by-then extremely mature 7nm node.
 

Gomez Addams

Prominent
Mar 4, 2020
53
27
560
>> " I think it would be appropriate to believe Meteor Lake will be noticeably quicker vs Intel's future 11th and 12th Gen CPUs. "

So you think the 13th gen will be at least somewhat faster than the ones which came before it? Really going out on a limb there, aren't you?

My opinion is that's not a safe assumption at all. If they are going to spend (waste) chip real estate on a low-power core I have very little anticipation of seeing any performance improvement to speak of.
 

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
My opinion is that's not a safe assumption at all. If they are going to spend (waste) chip real estate on a low-power core I have very little anticipation of seeing any performance improvement to speak of.
If we can find someone to hold the stakes, are you willing to bet on that? :)
 

spongiemaster

Admirable
Dec 12, 2019
2,273
1,277
7,560
I won't debate that -- but since the author prefaces his conclusion by referencing Intel's process change, rather than any architectural IPC improvements -- that's rather a moot point, isn't it? I'd struggle harder to find a safer bet than concluding Intel will see substantial performance gains from the by-then extremely mature 7nm node.
Your original quote doesn't contain anything about the process change. After reading the conclusion, none of it really makes any sense. There isn't going to be an 11th gen desktop generation, so not sure what they are referring to there. When was the last time a process change made any performance improvements for Intel? It used to be that smaller nodes meant higher clock speeds. Now, smaller nodes basically mean better power efficiency and that's about it. Here's an Ivy Bridge overclocking guide from 2012.


"For 5GHz for instance, it is possible to OC to 5GHz with 1.4v on air:
5.3GHz is my maximum validation on air: "

That was on 22nm. We're still basically at those clock speeds when overclocking with Comet Lake and Intel is clearly struggling to even get to 5Ghz with 10nm. Without a fundamentally different ISA, I doubt will ever see even 6Ghz regardless of how small the node get.
 

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
When was the last time a process change made any performance improvements for Intel? It used to be that smaller nodes meant higher clock speeds. Now, smaller nodes basically mean better power efficiency and that's about it. Here's an Ivy Bridge overclocking guide from 2012....

(URL): "For 5GHz for instance, it is possible to OC to 5GHz with 1.4v on air ...

That was on 22nm.
A few points:
a) 22nm is only one node behind 14nm. I won't count 10nm as we both know its essentially broken at present. But Intel's 7nm node will likely actually be TSMC's node, no? And that one is working very well ... and by 2022 will be extremely mature.
b) You're conflating maximum OC speeds with actual release speeds. The 22nm node debuted in the mid 3-Ghz range, IIRC, whereas the 14nm node is in the low 4s. That's a rather healthy bump in clocks for just a single node jump.
c) Obviously, even if one discounts IPC improvements, there is more to performance than clock frequencies. Additional cache ram and extra cores are two of the more obvious alternatives.
 

spongiemaster

Admirable
Dec 12, 2019
2,273
1,277
7,560
A few points:
a) 22nm is only one node behind 14nm. I won't count 10nm as we both know its essentially broken at present. But Intel's 7nm node will likely actually be TSMC's node, no? And that one is working very well ... and by 2022 will be extremely mature.
Well, that's 2 nodes that aren't faster. Just because that counters your argument doesn't mean can just ignore them. What is your prediction for all core overclock for Intel's 7nm? Do you think they'll get higher than the 5.1-5.2 that Comet Lake does? I bet they don't.
b) You're conflating maximum OC speeds with actual release speeds. The 22nm node debuted in the mid 3-Ghz range, IIRC, whereas the 14nm node is in the low 4s. That's a rather healthy bump in clocks for just a single node jump.
If you comparing what the nodes are capable of, you compare overclock to overclock. With each successive generation, Intel is using more sophisticated boosting algorithms to get CPU's closer to the maximum clocks out of the box leaving less and less overclocking headroom each time. Since Ivy Bridge, Intel has switched from a paste TIM to a soldered one. Should the node get credit for the higher standard clocks that provides? I would argue, no.
c) Obviously, even if one discounts IPC improvements, there is more to performance than clock frequencies. Additional cache ram and extra cores are two of the more obvious alternatives.
Neither of those necessarily requires a smaller process.
 

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
Well, that's 2 nodes that aren't faster.
Just one node: Intel's 10nm. The 14nm node was demonstrably superior.

With each successive generation, Intel is using more sophisticated boosting algorithms to get CPU's closer to the maximum clocks out of the box leaving less and less overclocking headroom each time.
That's part of it, sure. But you're also attempting to compare different architectures across different nodes, which can't be done reasonably. Haswell, Skylake and its successors added a large number of architectural features which affected overclocking. Had Intel done a simple shrink, it's certain there would have been substantially more headroom. In any case, you're still focusing on overclocking, when the real issue -- for nearly all users, at least -- is performance right out of the box. That was the point of my original post, and what, I assume at least, you're attempting to refute, no?
 
Linux Kernel 5.10 adds support for Intel's Meteor Lake architecture.

Intel Meteor Lake Architecture Added to Linux Kernel : Read more

Hybrid designs are an interesting idea that work well in cell phones for keeping power in check.

HOWEVER most power consumption comes from AVX like SIMD/MIMD instructions. These are most often used during decoding/encoding processes. These hybrid designs will fall FLAT on their face when it comes to things like Adobe Photoshop/Elements/Premiere, AI training, and more.
 

samopa

Distinguished
Feb 12, 2015
202
55
18,660
Hybrid designs are an interesting idea that work well in cell phones for keeping power in check.

HOWEVER most power consumption comes from AVX like SIMD/MIMD instructions. These are most often used during decoding/encoding processes. These hybrid designs will fall FLAT on their face when it comes to things like Adobe Photoshop/Elements/Premiere, AI training, and more.

You have to remember that these Tiger Lake, Alder Lake, and Meteor Lake are mobile CPUs. They're designed to be used in mobile environment (AKA laptop, notebook, tablet, etc), so the power is premium commodity and hybrid approach is make sense here.

Hybrid design is less important in desktop CPU, because the nature that always connected to the power make little sense to utilize less powerful core.
 

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
Hybrid design is less important in desktop CPU, because the nature that always connected to the power make little sense to utilize less powerful core.
Less important, certainly, but it still makes sense, given that desktop cpus are today limited by how much heat they can dissipate more than any other factor. If your hybrid core gives you half the performance for 1/4 the power, that means you can pack in four times the cores and get double the performance at the same thermal envelope.
 

samopa

Distinguished
Feb 12, 2015
202
55
18,660
Less important, certainly, but it still makes sense, given that desktop cpus are today limited by how much heat they can dissipate more than any other factor. If your hybrid core gives you half the performance for 1/4 the power, that means you can pack in four times the cores and get double the performance at the same thermal envelope.

If that is the case, that the low power core give you half performance with 1/4 power than the high power one, it is better to replace all the high power one with lower power core, so for a certain budget of power (i.e. 100 watt) you will get twice the performance than using all high core, or 1.5x the performance of hybrid one (assuming 50:50 ratio of high:low core), so I stand correct that make little sense to use hybrid design in desktop space
 

spongiemaster

Admirable
Dec 12, 2019
2,273
1,277
7,560
You have to remember that these Tiger Lake, Alder Lake, and Meteor Lake are mobile CPUs. They're designed to be used in mobile environment (AKA laptop, notebook, tablet, etc), so the power is premium commodity and hybrid approach is make sense here.

Hybrid design is less important in desktop CPU, because the nature that always connected to the power make little sense to utilize less powerful core.
Only Tiger Lake is a mobile CPU. Alder Lake and Meteor Lake are both desktop CPU's.
 

spongiemaster

Admirable
Dec 12, 2019
2,273
1,277
7,560
If that is the case, that the low power core give you half performance with 1/4 power than the high power one, it is better to replace all the high power one with lower power core, so for a certain budget of power (i.e. 100 watt) you will get twice the performance than using all high core, or 1.5x the performance of hybrid one (assuming 50:50 ratio of high:low core), so I stand correct that make little sense to use hybrid design in desktop space
That only make sense for highly threaded applications. What happens when you're running a program that only uses one thread? Now you're stuck on a really slow core. That's why you need a balance of both. High powered larger cores for low threaded apps, and then more efficient smaller cores to boost highly threaded work loads. It's not just saving power, it's lowering cooling requirements, and fewer transistors leading to smaller dies which has multiple benefits.
 
  • Like
Reactions: TJ Hooker

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
If that is the case, that the low power core give you half performance with 1/4 power than the high power one, it is better to replace all the high power one with lower power core, so for a certain budget of power (i.e. 100 watt) you will get twice the performance than using all high core, or 1.5x the performance of hybrid one (assuming 50:50 ratio of high:low core), so I stand correct that make little sense to use hybrid design in desktop space
In addition to @spongiemaster 's excellent point above, you also have to remember that the low-power cores in a hybrid design usually lack some of the advanced instruction set of a more modern core, such as AVX, TXS, and the like. On some applications, this doesn't impact performance at all, but on others it can be crippling.
 

samopa

Distinguished
Feb 12, 2015
202
55
18,660
That only make sense for highly threaded applications. What happens when you're running a program that only uses one thread? Now you're stuck on a really slow core. That's why you need a balance of both. High powered larger cores for low threaded apps, and then more efficient smaller cores to boost highly threaded work loads. It's not just saving power, it's lowering cooling requirements, and fewer transistors leading to smaller dies which has multiple benefits.

If now still exist a program that only used ONE thread, all you need just ONE high power core, then the rest of it would be low power cores, but that is not the case, is it ? Why AMD still cram many high power core in the next Zen 3 ? Maybe also in Zen4 ?

@Endymio :
https://www.tomshardware.com/news/i...architecture-to-support-avx-avx2-and-avx-vnni
The article above show that the low power core (Gracemont) is support advanced instruction set, including AVX, AVX2, and AVX-VNNI instructions, The advanced instruction set not supported is AVX-512 (only ?)
 

Endymio

Reputable
BANNED
Aug 3, 2020
725
264
5,270
If now still exist a program that only used ONE thread, all you need just ONE high power core, then the rest of it would be low power cores, but that is not the case, is it ?
If everyone bought PCs to run just one single application, it very well might be. But those pesky consumers actually want to run a variety of different programs on their PC, some of which may be primary or entirely single-threaded, and some of which are not.

The article above show that the low power core (Gracemont) is support advanced instruction set, including AVX, AVX2, and AVX-VNNI instructions, The advanced instruction set not supported is AVX-512 (only ?)
Did you read the entire article? To quote:

"Intel’s small low-power cores for client system-on-chips have always featured rather decent functionality, but have never supported the instructions required for various high-performance computing or media encoding/decoding workloads to minimize their sizes and power consumption. This is going to change with upcoming Gracemont cores..."

The low-power cores always play a catch-up race with their higher brethren. Gracemont will support most of AVX now, but won't have AVX-512, nor will support the AMX (matrix) extensions. At some point in the future, a Gracemont successor likely will-- but, by then, the high-power cores will have additional extensions of their own.
 

spongiemaster

Admirable
Dec 12, 2019
2,273
1,277
7,560
If now still exist a program that only used ONE thread, all you need just ONE high power core, then the rest of it would be low power cores, but that is not the case, is it ? Why AMD still cram many high power core in the next Zen 3 ? Maybe also in Zen4 ?
If you ranked thread counts by how many applications use that many, 1 thread would probably still come in first. That's missing the overall point. At any one time you're likely to have more than one program running at a time or programs that max at 2 or more threads, so you still need more than one large core. That's why with Alder Lake we're looking at 8 large cores which will cover the vast majority of desktop workloads and then 8 smaller cores to efficiently boost higher threaded applications.