Question Raptor Lake - Official Thread

[Hulk]

Since we already have the first Raptor Lake leak I'm thinking it should have it's own thread.
What do we know so far?
From Anandtech's Intel Process Roadmap articles from July:

Built on Intel 7 with upgraded FinFET
10-15% PPW (performance-per-watt)
Last non-tiled consumer CPU as Meteor Lake will be tiled

I'm guessing this will be a minor update to ADL with just a few microarchitecture changes to the cores. The larger change will be the new process refinement allowing 8+16 at the top of the stack.

Will it work with current z690 motherboards? If yes then that could be a major selling point for people to move to ADL rather than wait.

 

[Hulk]

Let's go down the rabbit hole a bit to try and figure out what the hell Intel is up to with Gracemont. They told us it was for the best multithreading performance in the least area. With that being a starting point let's not consider ST performance, that is what Golden Cove is for.

After much testing I have found that in Cinebench R23 1 Golden Cove w/HT earns 506 CB R23 points for every GHz of frequency. 1 Gracemont earns 242 CB R23 points.

Now let's turn those numbers into CB R23 points per GHz per mm^2.

And we get 73.5 CB R23 points per mm^2 for a Gracemont core at 1GHz and 48.06 CB R23 per mm^2 for a Golden Cove core at 1GHz. At equal frequency Gracemont does better with CB R23 by +53% over Golden Cove. And that is the reason for being for Gracemont.

But what if we make this comparison with Gracemont at 3.8GHz and Golden Cove at 5GHz? Then the Gracemont lead shrinks to 1.16%.

So, in this ONE MT performance benchmark, Cinebench R23, which one of our colleagues recently reminded me is "ridiculously multithreaded" in terms of efficiency and raw performance, even with GC at 5GHz and GM at 3.8GHz Gracemont wins by 16% if each occupy the same silicon area. Or put another way, if we have two CPU's each with the same area and the Golden Cove one run 5GHz and the Gracemont one runs 3.8GHz, with CB R23 Gracemont will be 16% faster. And of course, much more efficient due to it's lower clock speed. If we equalize clocks, efficiency will be much closer, perhaps GC would even be more efficient, but Gracemont would be 53% faster. Gracemont is there for MT performance while using little space on the die.

In summation, let's called Cinebench R23 "best case" MT performance.

So, best case MT performance at equal clock speeds with equal die area Gracemont is 53% better than Golden Cove.

With clock speed difference taken into account, best case MT performance with equal die area Gracemont is 16% better than Golden Cove.

 

[igor_kavinski]

I'm sure I've seen GM cores scaling somewhere but can't quite remember. Can anyone help out with that? It would be helpful to see how well GM scales from 2 to 4 to 6 to 8 cores in performance. So then we could predict somewhat how much of a performance uplift can be expected from 16 GM cores in 13900K.

 

[nicalandia]

Now let's turn those numbers into CB R23 points per GHz per mm^2.

What source are you using for mm2 on both Gracemont and Golden/Raptor Cove?


Here is the info we have so far
2.2 mm2 for Single Grace Mont With L2$ and 3.29 mm2 with L3$ and Ring Bus
7.04 mm2 for Single Golden/Raptor Cove with L2$ and 10.53 mm2 with L3$ and Ring Bus

 

[LightningZ71]

So, technically speaking, a 6 GC/16 GM core processor could have had equivalent ST and lightly threaded performance while also providing superior MT performance to a roughly similarly sized 12900K?

 

[Doug S]

Now let's turn those numbers into CB R23 points per GHz per mm^2.

And we get 73.5 CB R23 points per mm^2 for a Gracemont core at 1GHz and 48.06 CB R23 per mm^2 for a Golden Cove core at 1GHz. At equal frequency Gracemont does better with CB R23 by +53% over Golden Cove. And that is the reason for being for Gracemont.

But what if we make this comparison with Gracemont at 3.8GHz and Golden Cove at 5GHz? Then the Gracemont lead shrinks to 1.16%.

Why are you comparing in such a ridiculous category as "points per GHz per mm^2"?

What matters to Intel with the smaller cores is performance per watt, and presumably the more efficient cores will do much better there even with the ~25% frequency penalty.

Given that Cinebench is already overly dependent on frequency, adding a "per GHz" caveat to it is pretty redundant.

 

[Hulk]

What source are you using for mm2 on both Gracemont and Golden/Raptor Cove?


Here is the info we have so fa
2.2 mm2 for Grace Mont With L2$
7.04 mm2 for Golden/Raptor Cove with L2$

I used the post in this forum. Here's my Excel napkin math.
Same area numbers you quoted above.

16 Gracemont		8 Golden Cove
52.68​		84.23​
1 Gracemont		1 Golden Cove
3.2925​		10.52875​
Gracemont/mm		Golden Cove/mm
73.50​		48.06​
	1.53​
279.30​		240.29​
	1.16​

 

[Hulk]

Why are you comparing in such a ridiculous category as "points per GHz per mm^2"?

What matters to Intel with the smaller cores is performance per watt, and presumably the more efficient cores will do much better there even with the ~25% frequency penalty.

Given that Cinebench is already overly dependent on frequency, adding a "per GHz" caveat to it is pretty redundant.

I disagree. Intel told us Gracemont was there for area efficiency, not power efficiency.

 

[nicalandia]

I used the post in this forum. Here's my Excel napkin math.
Same area numbers you quoted above.

16 Gracemont		8 Golden Cove
52.68​		84.23​
1 Gracemont		1 Golden Cove
3.2925​		10.52875​
Gracemont/mm		Golden Cove/mm
73.50​		48.06​
	1.53​
279.30​		240.29​
	1.16​

Pretty good Math. I also agree that Gracemont was specifically for Multi-Threaded Throughput

 

[Hulk]

So, technically speaking, a 6 GC/16 GM core processor could have had equivalent ST and lightly threaded performance while also providing superior MT performance to a roughly similarly sized 12900K?

Yes. Perhaps in the future Intel will offer parts that suit specific workloads. 8+16, 6+20, who knows? I think 8+16 will be a beast.

 

[nicalandia]

So, technically speaking, a 6 GC/16 GM core processor could have had equivalent ST and lightly threaded performance while also providing superior MT performance to a roughly similarly sized 12900K?

Yes

Mock Up of 13900HK with 6 P Cores and 16 e Cores


Alder Lake S 8 + 8

 

[Exist50]

Also the block diagram will not change the size because these are the same Gracemont Cores used in Alder Lake, but the additional 2 MB of L2$ is just being enabled(it was disabled on Alder Lake)

Where do you see that they have extra cache sitting around that isn't enabled in any product?

 

[nicalandia]

Where do you see that they have extra cache sitting around that isn't enabled in any product?

They will not design two separate Gracemont Cores, they just lasered off half of it on the first iteration(Alder Lake)

They specifically stated "Up To" in their presentation.

 

[Exist50]

They will not design two separate Gracemont Cores, they just lasered off half of it on the first iteration(Alder Lake)

They specifically stated "Up To" in their presentation.

View attachment 57022

Supporting "up to" does not mean every implementation has that much in silicon. You do not need to redesign the core to add some more L2. And frankly, the idea that they're currently sandbagging by pointlessly lasering off half their GRT L2 is ridiculous. Intel is in no position to be sandbagging anything.

 

[nicalandia]

Supporting "up to" does not mean every implementation has that much in silicon. You do not need to redesign the core to add some more L2. And frankly, the idea that they're currently sandbagging by pointlessly lasering off half their GRT L2 is frankly ridiculous. Intel is in no position to be sandbagging anything.

We will see when the Real Die shot comes in.

 

[HurleyBird]

With clock speed difference taken into account, best case MT performance with equal die area Gracemont is 16% better than Golden Cove.

On one hand, Golden Cove is disproportionately well suited to Cinebench. I'm not sure about Gracemont, but to the extent it isn't as well suited, that will look worse for Gracement than the 16% number you've derived.

On the other hand, you're only considering area and performance. Little cores are mostly about perf/w. Incidentally, when you have power constraints better perf/w can easily translate to better absolute performance. When you need to stay inside a power budget, then moving work to more efficient cores can improve performance.

 

[eek2121]

Intel said that a single Gracemont core was exactly 1/4 the size of a single Golden Cove core and that made people think that a Gracemont Cluster(4 cores) would use the same die size area of a single P core and as you can see on the annotations, they dont. The 16 core cluster use more area(52.68 vs 42.11)

Nobody ever claimed the areas were exact. However, 4 Gracemont is very close to 1 GC core…

I disagree. Intel told us Gracemont was there for area efficiency, not power efficiency.

I have been trying to tell everyone this for quite a while, not just about Intel “hybrid”, but about these big/medium/little designs, period. Not a single company, including Apple, uses these types of designs for power savings. It all comes down to cost and nothing else. Of course everyone here continues to be in denial…

 

[coercitiv]

Little cores are mostly about perf/w. Incidentally, when you have power constraints better perf/w can easily translate to better absolute performance.

I'm running out of energy to repeat this, little cores are first about perf/area. When compared in equal numbers, E-cores can be matched by P cores in terms of perf/watt at anywhere between 2 and 3Ghz depending on the workload.

Spoiler: Core power usage for transcoding & compression

The advantage they bring in perf/watt can only be seen in very light loads or alternatively when normalizing for ISO area. Intel engineers made this abundantly clear at Hot Chips last year.

 

[Exist50]

I have been trying to tell everyone this for quite a while, not just about Intel “hybrid”, but about these big/medium/little designs, period. Not a single company, including Apple, uses these types of designs for power savings. It all comes down to cost and nothing else. Of course everyone here continues to be in denial…

Apple definitely skews their small core design much more towards power efficiency vs area. I think they're pretty much unique in that regard. ARM and Intel are weighted down by business segments that care about the area efficiency.

 

[igor_kavinski]

"Higher computational density". I like the sound of that. So P-cores would be "Higher Chances of Melting" cores.

 

[coercitiv]

Apple definitely skews their small core design much more towards power efficiency vs area. I think they're pretty much unique in that regard.

I think we can agree that Apple is willing to sacrifice area (or cost) on all their IP as long as they get a large enough gain in efficiency. With the recent addition of M1 and derivatives, that includes packaging too.

They're also uniquely able to push this cost to the consumer, the recent MacBook line pricing is living proof of that.

 

[IntelUser2000]

That would take pretty strong self-control, for an Alder Lake owner not to upgrade to Raptor Lake at some point. The increased cache and additional GM cores alone would make it worth it, even without any architectural tweaks.

Really? You think a year's worth of changes are worth spending your hard earned money for it? Go ahead, it's your money but if you want to argue it has to do with things like "self control" or that's a "need" you totally lost me, because it's neither of those at all, except for the few that explicitely need the performance. And when I say need I mean work not gaming. And you can game with LOT less hardware too.

Also to that guy saying Intel had to laser off L2 cache on GRT cluster for Alderlake is being ridiculous. So you are saying they have 4MB cache dies, but they are saving it for next year! Hahah!

 

[HurleyBird]

The advantage they bring in perf/watt can only be seen in very light loads or alternatively when normalizing for ISO area. Intel engineers made this abundantly clear at Hot Chips last year.

Looking at the graphs, I stand in large part corrected. The advantage in perf/w is almost entirely derived from higher compute density. At ISO performance (at least in a well threaded case), you have more compute resources at your disposal and so you don't need to push any individual resource too high up the efficiency curve.

Of course, rephrased it this way still challenges Hulk's idea that the Gracemont die area might be better utilized on more Golden Cove cores instead.

 

[igor_kavinski]

Really? You think a year's worth of changes are worth spending your hard earned money for it? Go ahead, it's your money but if you want to argue it has to do with things like "self control" or that's a "need" you totally lost me, because it's neither of those at all, except for the few that explicitely need the performance. And when I say need I mean work not gaming. And you can game with LOT less hardware too.

If there are people who can silence the voice in the back of their head telling them to throw out ADL in favor of Raptor Lake when that comes out and works in the same mobo, well, either I'm the only one with voices in my head or people have a LOT of self-control. True, nobody really needs all that performance. Haswell and heck even Ivy Bridge can still run Windows 11 and the majority of software. But most of us upgrade to save time and Raptor Lake may end up saving enough seconds to justify an upgrade for a lot of ADL owners.

Also, I have a strong feeling that Raptor Lake may not be the typical subsequent generational leap in performance. It will coincide with Zen 4 and if Intel is wise enough to consider it as a serious threat, then it logically follows that they have put the best of their best at ensuring that they maintain their general performance lead. Gelsinger remarking that AMD is in the rear view mirror may actually be a sign of Raptor Lake being more than a "tick" following up the "tock" of ADL, because he has likely seen the Raptor Lake benchmarks and all their performance projections for Zen 4 (or even leaks from industry insiders) are suggesting to them that AMD may not be able to thoroughly embarrass them again like they did with Zen 3.

 

[Mopetar]

So, best case MT performance at equal clock speeds with equal die area Gracemont is 53% better than Golden Cove.

With clock speed difference taken into account, best case MT performance with equal die area Gracemont is 16% better than Golden Cove.

One thing to consider is how much power is going to be used in those different scenarios. In a world where power doesn't matter, you could just use more Golden Cove cores, but in reality Intel is already at the wall as is. Golden Cove can reclaim some of the performance deficit by running at higher frequencies, albeit at the cost of a lot of power.

I think the full picture only becomes clear when we look at performance, area, and power all at the same time. In TPUs testing a 12900K at 125W limits for PL1 and PL2 scored 18138 points and 27780 points when at 241W for both. That's a ~50% increase in performance for almost double the power.

 

[Hulk]

One thing to consider is how much power is going to be used in those different scenarios. In a world where power doesn't matter, you could just use more Golden Cove cores, but in reality Intel is already at the wall as is. Golden Cove can reclaim some of the performance deficit by running at higher frequencies, albeit at the cost of a lot of power.

I think the full picture only becomes clear when we look at performance, area, and power all at the same time. In TPUs testing a 12900K at 125W limits for PL1 and PL2 scored 18138 points and 27780 points when at 241W for both. That's a ~50% increase in performance for almost double the power.

Area and Power are dependent. More area provides more compute so you can lower clocks and get better efficiency. Gracemont is more power efficient than Golden Cove because it is more area efficient.. therefore frequency can be lower for equal performance. Lower frequency means less power consumption.

We are circling back to the same point. Gracemont is there for computational efficiency in terms of physical space (Intel says this). Golden Cove is for ST performance.

Golden Cove cores are fantastic for efficiency. If size of the CPU didn't matter you could simply put 24 GC's on the die, clock them at 3.8GHz and call it a day. It would be fantastic in terms of performance and efficiency but terribly expensive.

You see GC is expensive in terms of performance (more area) and GM is cheap (less area). That is the bottom line as to why they are there.