Discussion Intel current and future Lakes & Rapids thread

[Hulk]

No, Anandtech found that using SPEC - which contains a large variety of workloads - showed that on average there was a drop in IPC compared to Sunny Cove. They didn't just use a single workload for their analysis.

@ uzzi38,

Today is a good day. Thank you for setting me straight on this. I was under the assumption that for the most part the larger L2/L3 cache of Willow Cove would lead to better overall IPC performance vs. Sunny Cove. As you informed me I was dead wrong.

I went ahead and calculated every Anandtech bench from the Tiger Lake review for both Sunny and Willow. Since these were single core tests I used the maximum boost frequency and equalized results for clock speed. For the application benchmarks, which are multicore I used 3.0GHz for Willow and 2.6GHz for Sunny. Ian stated that after 20 seconds of running a multicore bench Ice Lake (Sunny) would thermal down to 2.6GHz. It took a couple of hours to account for variation in clocks and such but I have the numbers. I included the other processors for the single core results since we have a pretty good idea of clock speeds for them.

If anyone would like to see my Excel spreadsheet just PM me and I'll e-mail it to you.

Sunny Cove shows 5.08% better IPC in combined SPEC2006. Zen 3 is a beast.

Summary	SPECint2006	SPECfp2006(C/C++)	SPEC int+fp
5050	AMD Ryzen 9 5950x	0.0%	0.0%	0.0%
3900	Intel i7-1065G7	-20.9%	-10.1%	-17.3%
4800	Intel i7-1185G7	-22.8%	-24.7%	-23.5%
4700	AMD Ryzen 9 3950x	-29.6%	-20.7%	-26.6%
5300	Intel i9-10900k	-24.4%	-32.4%	-27.0%
4200	AMD Ryzen 7 4800u	-62.1%	-67.1%	-63.8%

For the application benchmark testing Sunny Cove is showing 0.7% better IPC than Willow Cove.
Now I understand for smaller applications that fit well in the Sunny Cove's smaller but faster L2 (like SPEC int) then Sunny has an advantage. But there are some apps that do require the large L2 even at the "expense" of longer latency.

Again, uzzi38 thanks for setting me straight on this! Took me a few hours to prove it to myself but hard work never hurt anyone (too much)!

Total IPC percentage difference for each group of tests was added and then divided by the number of benchmarks to arrive at these results and the ones above. So because there are more Science and Simulation benches than the other that category is weighted higher in Total result.

Summary	Office and Web Average	Science and Simulation Average	Rendering and Encoding Average	Total
2600	Intel i7-1065G7 (15W)	0.0%	0.0%	0.0%	0.00%
3000	Intel i7-1185G7 (15W)	-4.6%	4.3%	-1.0%	0.66%

 

[inf64]

Hulk, you can find the normalized performance per clock figures for both spec2006 and spec2017(rate 1) in Anadtech article here : https://www.anandtech.com/show/1621...e-review-5950x-5900x-5800x-and-5700x-tested/9
They even included Apple A13 (spec2006 results) which displayed stellar performance per clock (albeit running at much lower clocks which is a design choice).

 

[Bam360]

Mobile processors have worse IPC compared to Desktop processors, at least in some of the Spec workloads, this is not only because of cache differences, but also because LPDDR4X has worse timings compared to standard DDR. So this has to be taken into account when comparing Tiger lake IPC to Zen3 Desktop.

 

[Hulk]

Hulk, you can find the normalized performance per clock figures for both spec2006 and spec2017(rate 1) in Anadtech article here : https://www.anandtech.com/show/1621...e-review-5950x-5900x-5800x-and-5700x-tested/9
They even included Apple A13 (spec2006 results) which displayed stellar performance per clock (albeit running at much lower clocks which is a design choice).

Yes I noticed that thanks. I wanted to see dig into the numbers a little deeper to satisfy my curiosity.

In order to replicate Ian's numbers I had to put a clock of 5050MHz on the 5950x (as per what he said in the article) and 4625MHz for the 3950x.

 

[shady28]

Not sure why you did all that work, AT already did it for you.

This shows +7.2% Zen 3 IPC vs TGL on SPECfp, and +11.7% on SPECint.

It also shows an advantage of ICL vs TGL in float, while TGL has a slight advantage in int. AT






None of this really matters much to raw performance.

To that point, if you adjusted these raw rate numbers into a napkin math "IPC" for the 3Ghz M1, you'd find that it is about 60% higher "IPC" than Zen 3 in float and about 28% higher in int.

IPC is a useless metric for deriving performance as you have to ignore the clock speeds. It doesn't even tell us much about future potential, because a chip like the M1 that is very wide will never be able to clock up to the stratosphere like x86 chips.

I mean if you really want to look at IPC, Nuvia is expected to have significantly higher IPC than even M1 / A14.

 

[tamz_msc]

Quality i9 11900K QS CPUs reaching 700 in CPU-Z 1T, 1700 in CB R23 1T and 660 in CB R20 1T.

https://twitter.com/x/status/1342762159583850498

 

[JoeRambo]

RKL performance right where expected - ~15% performance gain versus my 5.1Ghz locked SKL sounds nice, hopefully it is possible to improve the scores by overclocking mem, uncore, fixing clock.

 

[shady28]

Quality i9 11900K QS CPUs reaching 700 in CPU-Z 1T, 1700 in CB R23 1T and 660 in CB R20 1T.

https://twitter.com/x/status/1342762159583850498

Looked at the original forum site, looks like that is still on a B560 motherboard and it has a crappy air cooler with bad thermal paste?

It appears that is an 11900K with a 100Mhz overclock from 5.3Ghz -> 5.4Ghz, but it's really hard to tell, need a translation of the translation.

Also there is some indication that the B560 may no longer be locked down like the X60 chipsets of the past.

There's also a snip of an avx512 workload. 3.8Ghz all core with 161W power draw. This illustrates the crappy cooler being used, hitting 94C here. Comments indicate that this is actually lower power and temp than a 10900K under the same work load though :

 

[Bam360]

Looked at the original forum site, looks like that is still on a B560 motherboard and it has a crappy air cooler with bad thermal paste?

It appears that is an 11900K with a 100Mhz overclock from 5.3Ghz -> 5.4Ghz, but it's really hard to tell, need a translation of the translation.

Also there is some indication that the B560 may no longer be locked down like the X60 chipsets of the past.

There's also a snip of an avx512 workload. 3.8Ghz all core with 161W power draw. This illustrates the crappy cooler being used, hitting 94C here. Comments indicate that this is actually lower power and temp than a 10900K under the same work load though :

View attachment 36445

3,8GHz using 1.241V, yeah this is the issue here, at this clock speed you would hover around 1V or even less using something like 8700K or 9900K or 10900K.

 

[Zucker2k]

3,8GHz using 1.241V, yeah this is the issue here, at this clock speed you would hover around 1V or even less using something like 8700K or 9900K or 10900K.

Running AVX-512 code?

 

[shady28]

Running AVX-512 code?

Yeah exactly. I got the gist from the original source thread that nothing could be concluded from the AVX test. There's just not enough data presented, only one data point. We'd need clock speed, temp, and power draw graphs over time which didn't appear to be there.

 

[Hulk]

Not sure why you did all that work, AT already did it for you.

To be honest I didn't scroll down the page of the Anandtech review to those charts. Yes I wasted two hours.
I'm stupid I know.

Using that data I'm getting 5950x vs Sunny Cove as follows. I weighted based on number of score in each category for "combined."

5950x vs 1065 - SPEC2017	5950x	1065
SPEC2017int - 10 scores	1.514	1.323	-14.4%
SPEC2017fp - 12 scores	2.413	2.335	-3.3%
Combined	3.675	3.438	-6.9%

5950x vs 1065 - SPEC2006	5950x	1065
SPEC2017int - 12 scores	13.57	11.47	-18.3%
SPEC2017fp - 6 scores	18.62	17.05	-9.2%
Combined	22.880	19.995	-14.4%

 

[Bam360]

Running AVX-512 code?

Well, I ran prime95 on my i7 8700 at 3.8GHz with both AVX on and off and the auto voltage was about the same. The power consumption was about 30-35% higher with AVX on though, not because of voltage, but because of SIMD units being naturally more power hungry.
What I'm saying is, at 1,241V you are going to get way higher power consumption, AVX or not.

 

[Zucker2k]

Well, I ran prime95 on my i7 8700 at 3.8GHz with both AVX on and off and the auto voltage was about the same. The power consumption was about 30-35% higher with AVX on though, not because of voltage, but because of SIMD units being naturally more power hungry.
What I'm saying is, at 1,241V you are going to get way higher power consumption, AVX or not.

Let's wait for final silicon, shall we? What matters is PL1 and PL2, and how long and aggressive boosting is going to be like. These engineering samples could be running generous power levels for various reasons.

 

[Hulk]

Let's wait for final silicon, shall we? What matters is PL1 and PL2, and how long and aggressive boosting is going to be like. These engineering samples could be running generous power levels for various reasons.

I agree. Plus we already know the performance will be like Ice Lake parts with clocks very similar to Comet Lake. But who knows maybe Intel will surprise us with something unknown and cool.

 

[shady28]

I agree. Plus we already know the performance will be like Ice Lake parts with clocks very similar to Comet Lake. But who knows maybe Intel will surprise us with something unknown and cool.

I wouldn't count on it, but a 5.2 or 5.3Ghz Ice Lake with PCIe 4.0 and an IMC that can handle DDR4-4200 with aplomb will still be a killer setup for some time to come.

I'm actually more worried about pricing than performance at this point. I've been putting off a 10700/10700K for this, and those things are down to $279 and $319 at Best Buy no less.

Now, I'm liable to wind up staring down a $500 11700K, then watching the 10700K spike back up to $389 like it was right after launch.

 

[Hulk]

I wouldn't count on it, but a 5.2 or 5.3Ghz Ice Lake with PCIe 4.0 and an IMC that can handle DDR4-4200 with aplomb will still be a killer setup for some time to come.

I'm actually more worried about pricing than performance at this point. I've been putting off a 10700/10700K for this, and those things are down to $279 and $319 at Best Buy no less.

Now, I'm liable to wind up staring down a $500 11700K, then watching the 10700K spike back up to $389 like it was right after launch.

That's a really great prices on the CPU's. Once you unlock the power setting on the 10700 isn't it like 100MHz less than the k model all core turbo?

 

[shady28]

That's a really great prices on the CPU's. Once you unlock the power setting on the 10700 isn't it like 100MHz less than the k model all core turbo?

You can power unlock a 10700, setting bclk to 102.5 (roughly), which will give you a bit over all core 4.7Ghz turbo. A stock 10700K will do 4.8Ghz for about 7 seconds before slowing down. Power unlocked otherwise 'stock' 10700K will of course beat a power unlocked 10700, as it will maintain its 4.8Ghz (provided you can keep it cool).

OFC, from what I've read pretty much all 10700K's can be set to 4.9Ghz all core, something like 80% can do 5Ghz all core, while about 2/3 can do 5.1Ghz all core.

The non-K chips do limit you a lot in playing with the multipliers, and that also affects what you can do with uncore (cache). I like to tinker, so my next chip will be a K chip.

 

[Hulk]

So we have a pretty good idea of how Rocket Lake will perform, how many cores it will have, and the clocks. I think the big question now is how is Intel going to price it?

This is the first time probably ever that the top of the stack part that they will be releasing, 11900k, will be smashed by two AMD parts, and most likely beaten pretty soundly by the 5800x. They must long for the days when they could release top of the stack "Extreme" parts with a $999 price.

Since the 5800x is $450, and assuming it is available at that price in quantity by March, that kind of puts an upper limit on the 11900k price, right?

 

[jj109]

So we have a pretty good idea of how Rocket Lake will perform, how many cores it will have, and the clocks. I think the big question now is how is Intel going to price it?

This is the first time probably ever that the top of the stack part that they will be releasing, 11900k, will be smashed by two AMD parts, and most likely beaten pretty soundly by the 5800x. They must long for the days when they could release top of the stack "Extreme" parts with a $999 price.

Since the 5800x is $450, and assuming it is available at that price in quantity by March, that kind of puts an upper limit on the 11900k price, right?

I don't know where you get your confidence from.

The Ice Lake Benchmark Preview: Inside Intel's 10nm

www.anandtech.com

Skylake 8C desktop has 17% higher IPC over Skylake 4C laptop from double the cache and a desktop memory subsystem. This should clue you into the nature of the benchmarks you are trying to extrapolate from.

Looks like 11900K QS numbers above your post are a few % higher than the 5800X on non-memory sensitive benchmarks, too.

 

[shady28]

Looks like 11900K QS numbers above your post are a few % higher than the 5800X on non-memory sensitive benchmarks, too.

Those numbers are about 6% higher than avg 5800X on CB R23 ST and 7.2% higher on geekbench 1T. They are also higher than a 5950X.

That's also with a B560 using a 6-phase VRM though. People always ignore motherboards on test setups.

If I had to bet, I'd guess that with a good Z590 motherboard, that 11900K will get at least a few percent more, given that there's a 5% difference between high end Z490 performance to start with, undoubtedly more vs an average B460 with a 6-phase VRM.





Zen 3 CB R23 1T scores :

 

[coercitiv]

New GB 5 entry for 11700K
1807 ST at 5Ghz as reported by GB
10673 MT

posted on Twitter by APISAK

 

[exquisitechar]

New GB 5 entry for 11700K
1807 ST at 5Ghz as reported by GB
10673 MT

posted on Twitter by APISAK

Haha, I was just about to post this. Matches with what I expected. Int score/clock seems a bit better compared to Zen 3, float score worse, and, of course, it has a healthy lead in crypto. I'm most interested in seeing prices now.

 

[coercitiv]

I'm most interested in seeing prices now.

And availability.

 

[cortexa99]

Haha, I was just about to post this. Matches with what I expected. Int score/clock seems a bit better compared to Zen 3, float score worse, and, of course, it has a healthy lead in crypto. I'm most interested in seeing prices now.

It seem that Crypto just add AVX512 support and that's why RKL could take the lead in this item.