Discussion Intel current and future Lakes & Rapids thread

[DrMrLordX]

The competition will be extremely fierce from now on because Intel won't be stuck waiting on process nodes again to bring new architectures to market.

Intel has no high-volume nodes coming after 10ESF for potentially a very long time. Unless the TSMC rumour is true, your statement is just not correct.

 

[A///]

What's that based on? We would have had competition if AMD hadn't massively fumbled their way into the latter half of the 2000s and hardly innovating in the 2010s first half due to lack of funds from their prior second half of the 2000s mistakes, plural implied.

The cheese blog reads like some teenage boy's adventure fantasy story.

 

[Racan]

Intel has no high-volume nodes coming after 10ESF for potentially a very long time. Unless the TSMC rumour is true, your statement is just not correct.

Yes, I should have mentioned it, I had TSMC in mind.

 

[DrMrLordX]

What's that based on?

You referring to my post? Basically, Intel hasn't bought much EUV equipment, and it doesn't look like they'll have the ability to change that fact until maybe after 2023. There are signs that they're working with IBM on a 2nm node (which will presumably also require EUV equipment) but no telling how that will work out for them.

Anyway, it looks like Intel's own 7nm node will reach 20-36 kwpm by 2023 depending on whom you believe. And that's really a low number. But then I've already beaten that to death in this thread haven't I?

Yes, I should have mentioned it, I had TSMC in mind.

Well okay. We'll see if Intel can pull it off.

 

[Racan]

What's that based on? We would have had competition if AMD hadn't massively fumbled their way into the latter half of the 2000s and hardly innovating in the 2010s first half due to lack of funds from their prior second half of the 2000s mistakes, plural implied.

The cheese blog reads like some teenage boy's adventure fantasy story.

Well it's a rumour... But after the V-cache tech reveal I don't see those IPC gains mentioned as totally implausible. Anyway sorry for going off topic, this isn't a Zen 4 thread.

 

[A///]

You referring to my post? Basically, Intel hasn't bought much EUV equipment, and it doesn't look like they'll have the ability to change that fact until maybe after 2023. There are signs that they're working with IBM on a 2nm node (which will presumably also require EUV equipment) but no telling how that will work out for them.

Anyway, it looks like Intel's own 7nm node will reach 20-36 kwpm by 2023 depending on whom you believe. And that's really a low number. But then I've already beaten that to death in this thread haven't I?

Nah, it was referencing the cheese post. I thought I had it quoted. 20-36 kwpm is pretty low, yes. Depending on die size and order, that's what you'd get from a major foundry customer that's sharing machine time with others on the same node.

 

[A///]

Well it's a rumour... But after the V-cache tech reveal I don't see those IPC gains mentioned as totally implausible. Anyway sorry for going off topic, this isn't a Zen 4 thread.

The date on that blog post? suggests those gains without the v-cache.

 

[DrMrLordX]

Nah, it was referencing the cheese post. I thought I had it quoted.

Woops nm

 

[eek2121]

My Life Is Dire has also been correct on a few, emphasis on 'few,' occasions but it doesn't mean anyone would give it credence now. I'm not familiar with either of the two people you've listed. Not trying to be rude to you here, Gideon. Best to tame expectations than rile ourselves up and feel like a weather-beaten saggy balloon in four or five months when Alderlake launches and possibly turns out to be a flop for one or more reasons.

Now, on the other hand, if Alderlake final turns out to be better and is this decade's Conroe, then I'd hope to high heaven or whatever deity or supernatural force y'all believe in if any that AMD can answer Intel in the coming year or two. It is infinitely better when these two companies go at it like an old married couple slapping each other in the head when we, the consumer, win with better hardware each year.



Not true. They did design a processor that doubled as a space heater and another processor that smashed its figurative head into a wall to communicate with its twin brother that was also hot and slow.

AMD will likely be able to counter this with vcache. They don’t need Zen 4.

Intel is said to be in talks to buy GloFo. Extremely curious development.

 

[mikk]

Still, if the rumors are accurate, this is the biggest achievement Intel has done since Conroe and an undersell of similar proportions to Zen 1:

Yes and this is what Raja Koduri basically told in an interview a year ago.

"Alderlake – it was only a small part of our Architecture Day event, but it’s the biggest architecture movement since the Core architecture in 2006. It’s a huge leap and will be very exciting.

We haven’t always delivered content creator-level performance, or at the very least you would have to choose between one platform and another, but with Alder Lake and hybrid technology, you can get both great gaming performance and content creation performance at the same time.

Intel’s Surprise Ryzen Killer: Alder Lake Hybrid Processors Offer Biggest Performance Leap In 14 Years

Intel has it has a killer CPU in the works. I chat to Intel's Raja M. Koduri about Alder Lake and why it's the biggest CPU for 14 years

www.forbes.com

 

[jpiniero]

Intel promising 20% more performance (Golden Cove vs Willow Cove) but rumors point to delivering closer to 35% in MT workloads (big-core to big core including the 300 Mhz uplift from 4.7Ghz to 5Ghz) and all in all near doubling of their R20 score when including small cores. And in no special AVX-512 workload, but a legacy AVX2 workload that has represented quite accurate average IPC changes in the past for both AMD and Intel.

Looking at the numbers, if the 11k score is true the outlier is the Gracemont cores and not Golden Cove. Gracemont being able to get Sunny Cove performance seems unrealistic unless it's a lot bigger than I think it should be.

 

[IntelUser2000]

I highly doubt AVX does anything for Cinebench, even the R20 version. There's an AT thread that talked about it and people did bunch of testing and it was of similar opinion. Maybe AVX helps speeding up 5% of code but that's it.

The Pentium Gold Cometlake doesn't have AVX, not even the first one.

Look at this test: https://www.computerbase.de/2019-03...iagramm-cinebench-r20-multi-thread-ergebnisse

Pentium Gold G5500 performs pretty much identical in MT performance to the i3 6100.

@eek2121 You are way overestimating the impact of AVX on Cinebench. Maybe the wider issue rate due to the dual cluster approach helps Cinebench over Skylake but AVX itself is of minimal impact. I'm not talking about AVX2 over AVX. I'm talking about AVX in general.

Go and look at the results yourself. Any Intel chip besides Tigerlake mobile chips the Pentium and the Celerons only support up to SSE4.2.

 

[Zucker2k]

Yes and this is what Raja Koduri basically told in an interview a year ago.

Intel’s Surprise Ryzen Killer: Alder Lake Hybrid Processors Offer Biggest Performance Leap In 14 Years

Intel has it has a killer CPU in the works. I chat to Intel's Raja M. Koduri about Alder Lake and why it's the biggest CPU for 14 years

www.forbes.com

You ought to realize any news from Intel in the past few years has gone through one ear and came out of the other ear of the particular audience you're trying to address. At best, you'd get the cliched response; "It's Intel we're talking about." It's still early days. Let's see how it goes when the retail silicon drops and random Joes can purchase and slap them into their rigs.

 

[IntelUser2000]

Looking at the numbers, if the 11k score is true the outlier is the Gracemont cores and not Golden Cove. Gracemont being able to get Sunny Cove performance seems unrealistic unless it's a lot bigger than I think it should be.

Are we of the opinion that Skylake only offers 3% over Broadwell and 6-8% over Haswell? I am still not sure how Skylake performs. While you can find many initial reviews that showed pitiful gains, you can also find videos that show some wild gains using high speed DDR4 over 4790K. What the hell is true?

Because if Skylake really is that poor, then yes 30% gain over Tremont plus 4-5% due to AVX2 will get us very close to Sunny Cove level. You are talking Sunny Cove being 2-4% faster.

Maybe Skylake did suck, but some of us viewed it with rose-colored glasses since they were stuck with it so long.

 

[IntelUser2000]

@Gideon

162x1.275x8+5000= 8262
115x0.95+8+3700= 3233

If 8 core Gracemont cores with 3.7GHz MT speeds can really get 3233 points in Cinebench R20(nevermind 4000 as some are suggesting) it means in just two generations their "Atom" core has effectively nullified every Core-based chip prior to Golden Cove. Cause what? It'll be quarter the core size?

The 6W mobile Tremont gets around 700 points. The 10W does much better, but still at 930 points. We're talking close to 40% perf/clock gains over Tremont. That's at 3233 points.

Look at what @Abwx just said:

At 3.7GHz a RKL is at about 4800pts, and that s with something like 32% SMT gain, without it and in 8T/8C mode it would be at 3700pts at said 3.7GHz...

Screw Rocketlake when you could have done the same with Gracemont instead!

I'm not saying Gracemont is that good. I'm saying I'm skeptical.

 

[Gideon]

You ought to realize any news from Intel in the past few years has gone through one ear and came out of the other ear of the particular audience you're trying to address. At best, you'd get the cliched response; "It's Intel we're talking about." It's still early days. Let's see how it goes when the retail silicon drops and random Joes can purchase and slap them into their rigs.

Dude it's Raja you're talking about. If you take him verbatim Polaris and Vega were also the second coming of Jesus. He likes to hype things up (not saying Alder Lake is bad, just not to base expectations off that guy)

 

[dullard]

I'm not saying Gracemont is that good. I'm saying I'm skeptical.

I don't know how good Gracemont is, and I take all rumors with a grain of salt. But the rumors are consistent enough that if any of them come true then Gracemont is quite a lot better than the typical Atom. I do think that the only reason that Intel would continue down this big.little path is if it is true that the little cores actually have some oomph behind them. Otherwise, if the little cores are dogs then putting out a large Alder Lake chip just months after Rocket Lake makes no sense.

For example, if the 2x multithread performance slide is true, then you can't get to that improvement level with slow Atom cores and no additional big cores. https://cdn.videocardz.com/1/2021/03/Intel-Alder-Lake-S-Specifications.jpg

 

[IntelUser2000]

I don't know how good Gracemont is, and I take all rumors with a grain of salt. But the rumors are consistent enough that if any of them come true then Gracemont is quite a lot better than the typical Atom.

Up until this rumor, most of us agreed on Gracemont having Skylake performance. That's no typical Atom, and the core size is likely going to continue to be tiny. That kind of performance is absolutely fantastic! That's no dog!

And for mobile 2x performance is no problem. Tigerlake is 4 cores. The 2+8 is likely going to be faster than current Tigerlake. When you go 4+8 or 6+8 I can totally believe it being twice as fast.

But for desktop being 2x as fast? Desktop is already super high clocked with Rocketlake and the maximum Alderlake config is 8+8. Remember it's a marketing document. Of course when they say it'll be "twice as fast" it'll be an up to figure. Like some-of-our-products will offer 2x the performance.

But now the hype has suddenly reached a fever pitch. People are believing in scores that need Gracemont to be Sunny Cove performance, or even Golden Cove!

 

[Mopetar]

To some degree I can buy them (Gracemont cores) being a lot better than anyone thinks. Atom has always been a bit of a dog so there's plenty of room to grow with a good design. If they were as good as Sunny/Golden Cove though it should make you wonder why it wouldn't just replace them outright.

 

[Accord99]

To some degree I can buy them (Gracemont cores) being a lot better than anyone thinks. Atom has always been a bit of a dog so there's plenty of room to grow with a good design. If they were as good as Sunny/Golden Cove though it should make you wonder why it wouldn't just replace them outright.

Lack of clock headroom and SMT may limit it's outright performance and competitiveness as a stand-alone core.

 

[Mopetar]

Lack of clock headroom and SMT may limit it's outright performance and competitiveness as a stand-alone core.

I don't really know how much of a difference there is in die space for their "big" and "little" cores but Intel's SMT is usually considered to get up to 30% more performance in the best case, but in some applications it can barely be double-digit. If adding an extra 2 cores isn't too much more die area it would offset the performance loss for an 8-core ship.

Obviously they can't do much about the clock headroom, but even if it wouldn't make a good halo product for desktop, it seems like it would make a great laptop chip. I guess it also lacks things like AVX-512, but that's incredibly niche as is.

 

[eek2121]

Up until this rumor, most of us agreed on Gracemont having Skylake performance. That's no typical Atom, and the core size is likely going to continue to be tiny. That kind of performance is absolutely fantastic! That's no dog!

And for mobile 2x performance is no problem. Tigerlake is 4 cores. The 2+8 is likely going to be faster than current Tigerlake. When you go 4+8 or 6+8 I can totally believe it being twice as fast.

But for desktop being 2x as fast? Desktop is already super high clocked with Rocketlake and the maximum Alderlake config is 8+8. Remember it's a marketing document. Of course when they say it'll be "twice as fast" it'll be an up to figure. Like some-of-our-products will offer 2x the performance.

But now the hype has suddenly reached a fever pitch. People are believing in scores that need Gracemont to be Sunny Cove performance, or even Golden Cove!

From what I can gather, Gracemont isn’t close to Sunny/Golden Cove except in applications that utilize AVX.

More importantly, people have underestimated both Golden Cove and Gracemont.

EDIT: specifically, the 11900k scored 69% of what the 5950X according to Anandtech. Golden cover is said to be 20% faster, but let’s say it is only 10% faster…

 

[Exist50]

Why would Gracemont be closer to Core specifically in AVX? If its capabilities were that good, it would probably support AVX512, so if anything, I expect AVX performance to be a weakness. And also think people should scale back expectations to Skylake IPC level. Kind of funny how fast that went from nearly unheard of to the low end of speculation.

 

[A///]

AVX2 performance should be better. Why is 512 being mentioned? It's still a fairly niche instruction set. I'm not 100% on board with extrapolated values based on on fuzzy math. I am more curious how inter core group switching will be handled. Someone mentioned latency on a comments board a few days ago and I haven't found anything discussing such matters.

 

[mikk]

Dude it's Raja you're talking about. If you take him verbatim Polaris and Vega were also the second coming of Jesus. He likes to hype things up (not saying Alder Lake is bad, just not to base expectations off that guy)

If the scores are real (which I believe) he turned out right. Not only it's the biggest architecture movement with two interesting new architectures, it's also the biggest performance leap since Conroe. The thing is he works for Intel now and that means for most people he must be automatically wrong.