Intel Comet Lake Thread

[AtenRa]

I think I either said it wrong, or you misunderstood. Yes, single core performance matters, and its always getting better. But in relation to multi-core, that is becoming MORE useful at a faster rate, and the need for it, and the actual hardware is stepping up multi-core performance quicker than single core.

So, in regards to Comet lake, its single core is most likely not much faster than 9900k and its multi-core compared to its competition is falling behind faster. Also, its thermal envelope is getting worse at a higher rate than before, and FAR behind its competition. Thus, I think its a horrible release. Its a joke IMO.

From the Core i7 10700F and bellow, I believe this is the best release of Intel the last 2-3 years.
They have competitive CPUs from the $100 up to $350 range in any metric (ST, MT, Price etc etc).
Only consumption is not competitive but in Desktop DIY this is not so much of a problem.

 

[Markfw]

From the Core i7 10700F and bellow, I believe this is the best release of Intel the last 2-3 years.
They have competitive CPUs from the $100 up to $350 range in any metric (ST, MT, Price etc etc).
Only consumption is not competitive but in Desktop DIY this is not so much of a problem.

Thats a way to look at it. But when you can buy the same power for the same price(and thats giving comet lake the benefit of the doubt before benchmarks and retail prices), but one uses twice the power, it makes no sense to buy it. Hence my conclusion.

 

[AtenRa]

Thats a way to look at it. But when you can buy the same power for the same price(and thats giving comet lake the benefit of the doubt before benchmarks and retail prices), but one uses twice the power, it makes no sense to buy it. Hence my conclusion.

Thats true for the top 8-10 Core SKUs, but the 4-6 cores are not that much more power costly than the competition.

 

[wilds]

The 10 series does look like a fun overclocking platform. I like that Intel is adding these nifty features like per-core HT toggling. Intel XTU being a requirement is a bummer as it is on Windows only.

I want to see the new XTU features in action for an overclock. I am curious on how many 'weak' or 'strong' cores are in the 10 core on average. Intel is giving us a lot more control and I like it.

 

[RetroZombie]

Disagree. Single core performance will always matter, even in servers.

Disagree. Multi core performance is what it matters now, specially in servers.
And amd right now is the one that have more cpu cores.
You would be partially right if intel single thread performance would be 2x or more times faster than amd but it's not, and it's actually slower.
This is no longer the year 2000.

 

[gdansk]

Disagree. Multi core performance is what it matters now, specially in servers.
And amd right now is the one that have more cpu cores.
You would be partially right if intel single thread performance would be 2x or more times faster than amd but it's not, and it's actually slower.
This is no longer the year 2000.

Really depends what you're running. If you're unfortunate enough to run software with per-core licensing, you'll think the opposite.

 

[uzzi38]

Disagree. Single core performance will always matter, even in servers. There are always serial tasks that cannot be decomposed to run in more than one thread and will bottleneck the system. I haven't seen a game yet that doesn't peg out one core at 100% (rendering task?).

1. Depends on the use case for the server.

2. Pretty sure Watch Dogs is just that (or at least, there was a similar game that would scale perfectly across threads, I just forget the name), and Doom Eternal absolutely is.

You do need to take an entirely different approach to how games handle rendering tasks currently though. Instead of a main thread that assigns and dispatches tasks to worker threads you need to instead create a list of tasks that need to be competed and threads will automatically retrieve tasks and create new ones, adding those new ones onto the queue for other threads to pick up when needed.

This way of handling thread scheduling in games means that any task can be performed on any single thread without having to call-back to the main thread.

(PS, "render threads", which is the main thread in which any render requests must come from are done for past DX11. Both DX12 and Vulkan both heavily emphasise render task scheduling on any thread you like afaik.)

 

[Markfw]

Really depends what you're running. If you're unfortunate enough to run software with per-core licensing, you'll think the opposite.

I don't want to go too far off-topic, but since you mentioned servers.... Rome is virtually tied with Intel on single core tasks, but they have twice the cores, and 1/2 the power usage per core. So single core licensing does not matter. The power budget in a datacenter is very proportional to the cooling requirements, and that is a big part of the budget, so again, another big non-no for Intel, and a big win for Rome.

Edit: twice the power usage means 4 times the total electric bill due to A/C. (approximately)

 

[RetroZombie]

I like that Intel is adding these nifty features like per-core HT toggling.
Intel is giving us a lot more control and I like it.

More time to waste is more likely.

Really depends what you're running. If you're unfortunate enough to run software with per-core licensing, you'll think the opposite.

Well in that case buy a lower core count and save money, not all roads lead to the 64 core epyc.

 

[uzzi38]

Really depends what you're running. If you're unfortunate enough to run software with per-core licensing, you'll think the opposite.

If you're stuck with per core licensing, you either still want Rome (either the 7F32/7F52) or are willing to wait for Milan in a few months :^P

Intel don't have a in single threaded performance in servers. Even when Ice Lake-SP launches, you're better off going Milan.

 

[gdansk]

Intel don't have a in single threaded performance in servers. Even when Ice Lake-SP launches, you're better off going Milan.

I'm curious, how do you know that to be true? It's all based on speculation that Zen 3 will be higher IPC and/or clock higher than Sunny Cove on 10nm+.

 

[uzzi38]

I'm curious, how do you know that to be true? It's all based on speculation that Zen 3 will be higher IPC and/or clock higher than Sunny Cove on 10nm+.

So far Ice Lake has displayed worse per-core performance and efficiency than Comet Lake-based silicon and is barely above Whiskey Lake at best. Even if Zen 3 only brings about a 10% improvement (which is uh, pretty pessimistic) it will still easily outperform Sunny Cove.

Also, which +s are we talking here? The original +s (CNL being 10nm, ICL being 10+ etc etc) or the CNL-is-pre-10nm +s (where suddenly CNL is not 10nm but is instead 'pre-10nm')? If you mean the latter, there is absolutely 0 reason to believe that Ice Lake-SP is on a newer process than Ice Lake-U.

 

[killster1]

strange everyone came to this thread to talk about how great amd is ;/

 

[gdansk]

So far Ice Lake has displayed worse per-core performance and efficiency than Comet Lake-based silicon and is barely above Whiskey Lake at best. Even if Zen 3 only brings about a 10% improvement (which is uh, pretty pessimistic) it will still easily outperform Sunny Cove.

Also, which +s are we talking here? The original +s (CNL being 10nm, ICL being 10+ etc etc) or the CNL-is-pre-10nm +s (where suddenly CNL is not 10nm but is instead 'pre-10nm')? If you mean the latter, there is absolutely 0 reason to believe that Ice Lake-SP is on a newer process than Ice Lake-U.

I don't pretend to have any idea about future products. But I insist that per-core performance *is important* (for some of us). It may well be that Milan will be better at that too, but we shouldn't say per-core performance doesn't matter since it's 2020.

 

[Markfw]

strange everyone came to this thread to talk about how great amd is ;/

No, at least I started to talk about how bad the product is versus "the competition" and give reasons as to why. Then it started to go a little more OT to discuss the general hardware of Intel. and then vs the competition on that. Its only logical to discuss how a new release relates to its competition . Just that its sort of not just about cometlake sometimes.

 

[uzzi38]

I don't pretend to have any idea about future products. But I insist that per-core performance *is important* (for some of us) and that we to look at that nearly indepently of multi-core performance. It may well be that Milan will be better at that too, but we shouldn't say per-core performance doesn't matter since it's 2020.

I didn't say it wasn't.

I said it depends on your usecase (in servers especially). Per core performance is just as important as core count, as both are needed for performance.

That being said, when it comes to general-purpose servers with software that can easily scale with cores/threads (which is most on the server side of things), would I rather take a <5% improvement in 1T performance over a fraction of power draw, or over 2x the core count? Not a chance.

 

[RetroZombie]

strange everyone came to this thread to talk about how great amd is ;/

Well what do you want, intel just did a paper launch of the 10th gen cpus.
It seams these days even rebranding is a difficult task for intel.

 

[Markfw]

Well what do you want, intel just did a paper launch of the 10th gen cpus.
It seams these days even rebranding is a difficult task for intel.

Exactly. Anandtech had an article on them, but benchmarks ? Hardware ? nooooooo.

 

[uzzi38]

Exactly. Anandtech had an article on them, but benchmarks ? Hardware ? nooooooo.

Performance reviews are still under embargo.

 

[Markfw]

Performance reviews are still under embargo.

Then why the release ? The advertising ? the big slides that Anandtech used ? Thats not a paper release ? And I did not even see a date of ANYTHING beyond what they presented, like "we will have benchmarks on (pick a date)"

 

[OriAr]

Review embargo is up on May 13th

 

[Markfw]

Review embargo is up on May 13th

Just curious how you know that, since I did not see it anywhere officially.

 

[uzzi38]

Then why the release ? The advertising ? the big slides that Anandtech used ? Thats not a paper release ?

Pretty much.

It's something of a standard now for AMD and Intel both. Announcement with SKU list, pricing etc you name it, and then 2-3 weeks later you get oerformance reviews.

Something something more marketing opportunities, something something more news cycles yadda yadda. Point is just to stay in the news as much as possible to keep the eyes on you instead of the other guy.

Same reason why AMD have another embargo set for Thursday despite the 3100 and 3300X embsrgo being so recently.

 

[piokos]

Anything with 16 threads should last 10 years, at least being awesome at first then over time being passed around some friends and family perhaps like i always do with my upgrades. Will certainly be outdated but 16 threads and 16gb of ram is gonna be all you prob will ever need for a browser/htpc box.

Except: what if cores become much faster - either by getting much higher frequencies (kind of in development) or higher IPC / new instruction sets (kind of certain)?
I mean: what if in 2025 we start to get cores twice as fast as what we use today. Or 3-4 times as fast near the end of the decade?
Software will evolve to utilize the new single-thread potential and suddenly you'll end up with a CPU that looks great in rendering benchmarks, but for some weird reason seems less responsive than a modern mobile SoC.

Seriously, don't plan to use a CPU bought today for the next 10 years. Not with all the tech that we might see in this decade.

Your argument was that the i9-9900 managing 10c at 2.8 GHz "wasn't that bad". It is that bad. The competition has 16c products able to run within the same power budget at the same clock speeds with better IPC, and 12c products that can run at significantly higher speeds with better IPC. Even look at the 8c parts. Intel's 65w 8c parts have a 2.9GHz base clock. AMD's 8c 65w part has a 3.6Ghz clock. Stock-for-stock, the 3700X will deliver significantly higher sustained throughput with the factory HSF.

I absolutely haven't said Intel's desktop products match AMD desktop products today. They can't. Fab node advantage is too significant.
What I meant is: 9900 is a good CPU. And it is. Maybe not for "sustained throughput", but it'll likely do well in real life: with a lot low-thread software being used.
But of course there are limits to what clever engineering and well tuned boosts can give you. And Ryzens' raw density advantage will prevail in things like rendering or batch processing. No way around it.

I never said you CAN'T use those chips for general computing, but they are a waste. Most people don't buy 8-core i9 CPUs solely for web browsing. They'll get mainstream i5 and i7 parts.

I've already admitted that mentioning web browsing was a mistake and it derailed this discussion. Don't think about it. A big chunk of consumers edits photo/video, transcodes videos, learns data analysis and so on. And that's just on the consumer side of story, which is the minor target market for these desktop CPUs (at least for Intel).

Thats a way to look at it. But when you can buy the same power for the same price(and thats giving comet lake the benefit of the doubt before benchmarks and retail prices), but one uses twice the power, it makes no sense to buy it. Hence my conclusion.

Well, there are a few reasons why this buy makes sense. Some are obvious: IGP and other technicalities, better support in software etc. Some consumers will find something that they value more than performance or efficiency. And for some AMD will simply not be an option right now (I'm in that group albeit maybe just for few more months).

But the key reason is: supply. You forget how much CPUs are being sold globally. AMD just can't make that many.
Intel knows that. So once the first wave of DIY buyers swallow all the supply of fast, efficient Zen2 SoCs, Intel can sell their slower, more expensive products to the rest. In the same way you go to a grocery store after work and have to buy a slightly beaten, ugly fruits - knowing that someone who gets up early took all the nice ones and paid the same price.

As AMD gets more of TSMC wafers, Intel will have to become more competitive. For now they still can ask as much as they ask. And that's why their profit margin is higher than AMD's.
To put it simply: humanity needs to buy 500 mln x86 processors in 2020 and Intel is the only company that can make 400 mln of those.
So there is some competition (especially since AMD is active in the most lucrative markets: DIY desktops and servers), but for the most part Intel remains a monopolist.
That's worth remembering for the next time you see someone saying that "Intel made a mistake by keeping foundries", "no one will buy from Intel at these prices" or "OEMs are biased and don't buy from AMD".

 

[piokos]

Single-core performance is getting less useful every day.

This is not correct and I have to comment since it's coming from a mod on a respected website. You should know better. Or ask someone.
Single-thread performance will always be important. It's the foundation of computing.

The reason why today we see CPUs evolving more in core count than in per-core performance is very simple.
Having a multi-core CPU in a PC is a fairly young idea. They've only been around for ~15 years.
We're not saturated on the core count yet, so the easiest way to make a CPU faster today is by adding cores. But this won't last forever.

And of course we have technologies that will let us make CPUs running at 10GHz or safely operate well over 100*C.
But they are more expensive than adding cores, so they have to wait for their turn.


The bolded portion above is a moderator call-out,
and is NOT allowed.

If you have an issue with moderation, you create a thread
in moderator discussions.

AT Mod Usandthem