Discussion Intel current and future Lakes & Rapids thread

[dullard]

If I had to guess I would say the official MSRP is going to be $399 for the 11700K. Gotta make up for the decline of the dollar, right?

If we exclude the nearly non-existent after thought of the 5775C (especially since Intel lowered the price right after that with the 6700k), then the most that Intel had raised the top non- extreme edition i7 price is $15. That was from the 8700k to the 9700k. Your $399 number would be higher than normal for their price increases. In thinking the 11700k will be closer to $389.

 

[dmens]

19% faster in what though?

Targeted synthetics. Looks like the backporting included ripping out bits and pieces of core resources to meet frequency tagets, keep area down, or both. New generation indeed.

 

[Exist50]

Targeted synthetics. Looks like the backporting included ripping out bits and pieces of core resources to meet frequency tagets, keep area down, or both. New generation indeed.

It's a straight backport, as far as anyone has been able to tell. Why shouldn't it have the same IPC?

Please find another forum to troll.

 

[dmens]

It's a straight backport, as far as anyone has been able to tell. Why shouldn't it have the same IPC?

Please find another forum to troll.

Key phrase being, "as far as anyone has been able to tell". As in, you have no clue.

It amuses me that you know so little and yet have the chutzpah to call others trolls.

 

[SAAA]

Targeted synthetics. Looks like the backporting included ripping out bits and pieces of core resources to meet frequency tagets, keep area down, or both. New generation indeed.

Key phrase being, "as far as anyone has been able to tell". As in, you have no clue.

Actually 18% was an average



Assuming Rocket lake has the same core architecture as Icelake (that given most leaks, known cache sizes/layout and a few other seems plausible) it retains most of that, maybe better given desktop memory over mobile systems that will help on some tests.

Not that moving a node has ever lead to much variance in IPC if we look at previous examples: Sandy-Ivy, or Haswell-Broadwell, or even Skylake-Cannonlake comparisons are all within 3-5%.
Those were jumps where they actually tried to slightly change the architecture, adding whatever improvements they could before the new one… a straight port would have probably netted 0%.

 

[dmens]

Actually 18% was an average

Assuming Rocket lake has the same core architecture as Icelake (that given most leaks, known cache sizes/layout and a few other seems plausible) it retains most of that, maybe better given desktop memory over mobile systems that will help on some tests.

Not that moving a node has ever lead to much variance in IPC if we look at previous examples: Sandy-Ivy, or Haswell-Broadwell, or even Skylake-Cannonlake comparisons are all within 3-5%.
Those were jumps where they actually tried to slightly change the architecture, adding whatever improvements they could before the new one… a straight port would have probably netted 0%.

Yeah, an average of what? Benchmarketers can spin IPC all day long by messing with the test list.

If there is a regression in Rocketlake performance vs its 10nm parent with equivalent memory, then it is very likely the backporting had logical changes, the most reasonable cause being to ease convergence. And the easiest thing to change on a CPU core is the buffer size parameters. You do the math.

 

[inf64]

If it had a mean of ~19% IPC uplift intel would had said so in no uncertain terms. What they did instead is use a phrase "up to 19%". If you look at the orginal Sunny Cove unveil they used a different phrase "geo-mean" and up-to went all the way to 40% range. It is clear as day that average will not be 19% or they would have said so. I personally expect an average to be ~5-6% behind a "true" Sunny Cove (which sits at ~19% vs Skylake). In summary, around ~12ish %.

 

[amrnuke]

Well while the AMD fans focus on the "leak" that tells them what they want to hear, here is a reminder.

In this comparison the 5800X and the 11700K appear to be clock locked at just under 5Ghz according to the .gb5

That is among the fastest 5800X scores I can find, most 5800X score a solid 10% lower. And it is losing to the pre-prod 11700K.

View attachment 37989




5800X GB5 file :

https://browser.geekbench.com/v5/cpu/5960362.gb5

11700K GB5 File :

https://browser.geekbench.com/v5/cpu/5571984.gb5

1) When looking at the gains the 11700K makes over the 10700K, I think it's fun to put it in context of gains the 5800X made over the 3800XT.
2) We can't draw many conclusions because we don't know about memory...
3) GB5 needs to fix crypto.

GB5 results for each
11700K (from you, link)
5800X (from you, link)
10700K (random higher-end selection, link)
3800XT (random higher-end selection, link)
(and my 5600X run, for fun, link)

-----
First, generational improvements on new core designs.

As GB5 do, I took the geomean of the tests in each category. I also corrected each geomean result for clock speed to give somewhat of a hint at uarch improvements. Here is what I found:
Single-thread tests
CL -> RL = +13.9% integer, +16.8% FP, +213.8% crypto
Z2 -> Z3 = +13.5% integer, +13.8% FP, +60.5% crypto
Multi-thread tests
CL -> RL = +8.5% integer, +7.4% FP, -6.5% crypto
Z2 -> Z3 = +5.4% integer, +7.4% FP, -16.1% crypto

This is remarkable. When compared to what AMD did with Zen 3, this is just as good. (Yes, overall, if you include crypto, Intel created an ST improvement of 27% compared to AMD's improvement of 18%, but I refused to include crypto, will detail below why it's just purely broken when comparing different CPUs.)

-----
Second, memory plays a huge role in many subsegments of the test, FP and crypto mostly. Speed is important. So are timings. You can fiddle around with memory alone and see variations of hundreds of points in MT just with relatively minor timing / speed changes. Not knowing what memory this test chip is using (or cooling, for that matter) makes it very difficult to make any ascertainments about the validity of the results.

-----
Third, the large swings in effects of crypto scores on overall scores (pulling RL and Zen3 ST scores up artificially by just including support willy-nilly for new instructions that simultaneously drag down those very chips' multi-threaded scores), is a huge issue, IMO. It's much worse now with VAES512 (RL) and VAES256 (Zen3) being included for these latest generations compared to previous ones.

While some of the negative effect of VAES256 on Zen3 multithreaded scores is mitigated by fast memory and good cooling, in no world should my 5600X, which on INT and FP tests is 10.9% slower than the above-referenced 5800X in MT loads, be 45.1% faster in MT crypto. The results just don't make much sense with crypto.

Similarly, with the 11700K vs 10700K, obviously its ST score is boosted by RL supporting VAES512, which CL doesn't support. But that's a double-edge sword, because by leveraging it, I'm guessing it increases power draw and heat generation substantially, thus gimping the multithreaded results. The 11700K is 8.1% faster than the 10700K in MT INT/FP workloads, but 6.5% slower in MT crypto.

This is actually baffling to me the more I think about it. GB5 have intentionally introduced asymmetry with adding VAES256 and VAES512 instructions, which produces vastly different results that have nothing to do with actual real-world processing power/speed. Further, the effect this has is that GB5's encryption results aren't replicable or consistent with anyone else's cryptography benchmarking results (Phoronix, Anandtech, Techspot). No one is running VAES512 encryption on a single thread, yet it is a hindrance when run on multiple threads, and that's a huge barrier to its applicability to the real world. Which begs the question as to why these instructions are even included in this benchmark - they're inefficient and slow the chips down in real use (again, no one is doing encryption on a single thread), and preclude us from accurately comparing chips.


Really have to wonder what they're thinking. Including instruction sets that are not widely adopted, kill real-world performance, and aren't included on all modern chips seems like a really stupid way to run things. "How will it compare to the newest devices on the market?" Well, I don't know, because you've obscured the results behind some obtuse crypto result!

 

[Exist50]

Key phrase being, "as far as anyone has been able to tell". As in, you have no clue.

It amuses me that you know so little and yet have the chutzpah to call others trolls.

You being in denial is not the same thing as me being ignorant. You just refuse to accept that a backport with some BE optimization can gain 1GHz clock speed. Hell, you even contradicted yourself, first claiming they removed some vague thing, and then claiming no one knows. If they were to remove something, it would be AVX-512, and yet it's still there. As are all the caches.

Look dude, we get you hate Intel. But no one wants to hear the constant whining, or your attempt to pretend to have some "insider knowledge" that you clearly lack.

 

[dmens]

You being in denial is not the same thing as me being ignorant. You just refuse to accept that a backport with some BE optimization can gain 1GHz clock speed. Hell, you even contradicted yourself, first claiming they removed some vague thing, and then claiming no one knows. If they were to remove something, it would be AVX-512, and yet it's still there. As are all the caches.

Look dude, we get you hate Intel. But no one wants to hear the constant whining, or your attempt to pretend to have some "insider knowledge" that you clearly lack.

Heh, you are so ignorant you treated a reasoned guess at how such a performance regression is possible as a "vague thing". Do you even understand the line of reasoning?

And gain 1ghz? At what corner? Is the backported process slower or faster at that corner? Do you have any clue what you are talking about? If you go from 3ghz to 4ghz with say 10% for skew/setup/guardband, that is taking the convergence period from 300ps to 225ps. That means a 25% reduction in delay, on a backported process. If it is the same architecture, and they did not re-pipeline or re-design anything, that is a massive difference. Such is possible two ways: first, the first design at 3ghz was grossly unoptimized (possible but unlikely given the amount of time that 10nm design has been kicking around), or secondly, the backport 14nm process is so much faster at that corner, that they ought to have just abandoned 10nm. Take your pick.

 

[Bouowmx]

Second, memory plays a huge role in many subsegments of the test, FP and crypto mostly. Speed is important. So are timings. You can fiddle around with memory alone and see variations of hundreds of points in MT just with relatively minor timing / speed changes. Not knowing what memory this test chip is using (or cooling, for that matter) makes it very difficult to make any ascertainments about the validity of the results.

Writes obituary anyway...
There, I can cherry pick another result (one much closer to the Geekbench aggregate)

ASUS System Product Name - Geekbench

Benchmark results for an ASUS System Product Name with an Intel Core i7-10700K processor.

browser.geekbench.com

 

[Exist50]

Heh, you are so ignorant you treated a reasoned guess at how such a performance regression is possible as a "vague thing". Do you even understand the line of reasoning?

And gain 1ghz? At what corner? Is the backported process slower or faster at that corner? Do you have any clue what you are talking about? If you go from 3ghz to 4ghz with say 10% for skew/setup/guardband, that is taking the convergence period from 300ps to 225ps. That means a 25% reduction in delay, on a backported process. If it is the same architecture, and they did not re-pipeline or re-design anything, that is a massive difference. Such is possible two ways: first, the first design at 3ghz was grossly unoptimized (possible but unlikely given the amount of time that 10nm design has been kicking around), or secondly, the backport 14nm process is so much faster at that corner, that they ought to have just abandoned 10nm. Take your pick.

+1GHz at Vmax on each, which is notably much higher with 14nm. Something like 0.2V worth, iirc. And that's not even touching on ULT vs XLT, design changes, FIVR, etc. It's not worth going into further detail with someone who doesn't even understand or accept the basics.

And yes, Ice Lake's version of 10nm was bad. I didn't even think that was in question. A similar gain was seen with 10SF, and will surely not be repeated with 10ESF. Intel even said that Willow Cove was basically just Sunny Cove on 10SF with the new L2.

 

[dmens]

+1GHz at Vmax on each, which is notably much higher with 14nm. Something like 0.2V worth, iirc. And that's not even touching on ULT vs XLT, design changes, FIVR, etc. It's not worth going into further detail with someone who doesn't even understand or accept the basics.

And yes, Ice Lake's version of 10nm was bad. I didn't even think that was in question. A similar gain was seen with 10SF, and will surely not be repeated with 10ESF. Intel even said that Willow Cove was basically just Sunny Cove on 10SF with the new L2.

Why aren’t you comparing similar voltages? Why aren’t you comparing Tigerlake? You really think 200mV is going to get 25% improvement when the voltage is that high already? What are you smoking? And what about FIVR? The issue with FIVR is response time and not steady state power, so how is that relevant to max core clocks? What is your point?

You literally just said backport then you toss in design changes. Then you talk about cell swaps, as if low leakage to nominal cells even makes a significant difference on delays at the voltages you are talking about. It doesn’t.

Do you even have an EE degree? Hope you don’t for your own sake, because you wasted your money.

 

[shady28]

1) When looking at the gains the 11700K makes over the 10700K, I think it's fun to put it in context of gains the 5800X made over the 3800XT.
2) We can't draw many conclusions because we don't know about memory...
3) GB5 needs to fix crypto.

GB5 results for each
11700K (from you, link)
5800X (from you, link)
10700K (random higher-end selection, link)
3800XT (random higher-end selection, link)
(and my 5600X run, for fun, link)

-----
First, generational improvements on new core designs.

As GB5 do, I took the geomean of the tests in each category. I also corrected each geomean result for clock speed to give somewhat of a hint at uarch improvements. Here is what I found:
Single-thread tests
CL -> RL = +13.9% integer, +16.8% FP, +213.8% crypto
Z2 -> Z3 = +13.5% integer, +13.8% FP, +60.5% crypto
Multi-thread tests
CL -> RL = +8.5% integer, +7.4% FP, -6.5% crypto
Z2 -> Z3 = +5.4% integer, +7.4% FP, -16.1% crypto

This is remarkable. When compared to what AMD did with Zen 3, this is just as good. (Yes, overall, if you include crypto, Intel created an ST improvement of 27% compared to AMD's improvement of 18%, but I refused to include crypto, will detail below why it's just purely broken when comparing different CPUs.)

-----
Second, memory plays a huge role in many subsegments of the test, FP and crypto mostly. Speed is important. So are timings. You can fiddle around with memory alone and see variations of hundreds of points in MT just with relatively minor timing / speed changes. Not knowing what memory this test chip is using (or cooling, for that matter) makes it very difficult to make any ascertainments about the validity of the results.

-----
Third, the large swings in effects of crypto scores on overall scores (pulling RL and Zen3 ST scores up artificially by just including support willy-nilly for new instructions that simultaneously drag down those very chips' multi-threaded scores), is a huge issue, IMO. It's much worse now with VAES512 (RL) and VAES256 (Zen3) being included for these latest generations compared to previous ones.

While some of the negative effect of VAES256 on Zen3 multithreaded scores is mitigated by fast memory and good cooling, in no world should my 5600X, which on INT and FP tests is 10.9% slower than the above-referenced 5800X in MT loads, be 45.1% faster in MT crypto. The results just don't make much sense with crypto.

Similarly, with the 11700K vs 10700K, obviously its ST score is boosted by RL supporting VAES512, which CL doesn't support. But that's a double-edge sword, because by leveraging it, I'm guessing it increases power draw and heat generation substantially, thus gimping the multithreaded results. The 11700K is 8.1% faster than the 10700K in MT INT/FP workloads, but 6.5% slower in MT crypto.

This is actually baffling to me the more I think about it. GB5 have intentionally introduced asymmetry with adding VAES256 and VAES512 instructions, which produces vastly different results that have nothing to do with actual real-world processing power/speed. Further, the effect this has is that GB5's encryption results aren't replicable or consistent with anyone else's cryptography benchmarking results (Phoronix, Anandtech, Techspot). No one is running VAES512 encryption on a single thread, yet it is a hindrance when run on multiple threads, and that's a huge barrier to its applicability to the real world. Which begs the question as to why these instructions are even included in this benchmark - they're inefficient and slow the chips down in real use (again, no one is doing encryption on a single thread), and preclude us from accurately comparing chips.


Really have to wonder what they're thinking. Including instruction sets that are not widely adopted, kill real-world performance, and aren't included on all modern chips seems like a really stupid way to run things. "How will it compare to the newest devices on the market?" Well, I don't know, because you've obscured the results behind some obtuse crypto result!

You know, instead of posting all that pontification stuff you could have just done a direct comparison of the two systems I linked to on Geekbench's site.

You might learn something by looking at the data itself instead of regurgitating AMD fan talking points.

That learning being, these CPUs are better at different things, but they're not being carried by any one thing.

For example, Intel is faster at single thread :

HTM5 decode
SQL
PDF rendering
Text rendering
Camera functions
Machine learning
Speech recognition

Now what was it you were saying about Crypto, which is weighted at 5% of the overall score on Geekbench?

In multi-thread, Comet Lake does even better, taking narrow single thread losses and turning them to wins in things like navigation, HDR, and horizon detection. This implies that Comet Lake is more efficient at scaling with threads than Zen 3 is.

And it's doing it on a Z490. Wouldn't it do better on a Z590? I would think so.

Not that any of this in-the-weeds stuff makes much difference, as my point is not that the 11700K is better than a 5800X. My main point is that the 11700K seems a match for a 5800X.

I posted that, in context, to a bunch of posts about a recent "leak" that implied the 11900K was slower than the 5800X. That's extremely unlikely when the 11700K seems a match for the 5800X.

 

[DrMrLordX]

It's a straight backport, as far as anyone has been able to tell. Why shouldn't it have the same IPC?

It hasn't borne out in some leaks including results from multiple applications. Some of the earlier GB5 leaks also pointed to maybe a 10% improvement in IPC over Skylake-family cores (Comet Lake, in particular). It would be nice to have launch data on these chips, but until Intel chooses to finally produce review samples, we won't get that data.

 

[Exist50]

Why aren’t you comparing similar voltages? Why aren’t you comparing Tigerlake? You really think 200mV is going to get 25% improvement when the voltage is that high already? What are you smoking? And what about FIVR? The issue with FIVR is response time and not steady state power, so how is that relevant to max core clocks? What is your point?

You literally just said backport then you toss in design changes. Then you talk about cell swaps, as if low leakage to nominal cells even makes a significant difference on delays at the voltages you are talking about. It doesn’t.

Do you even have an EE degree? Serious question.

I'm giving actual reasons why Sunny Cove (or any other core) on 14nm clocks higher than on Ice Lake's version of 10nm. You just don't want to accept those reasons, even ones as direct as more voltage. I'll leave it at that, rather than wasting even more time convincing someone who refuses to listen.

 

[Thunder 57]

Well while the AMD fans focus on the "leak" that tells them what they want to hear, here is a reminder.

In this comparison the 5800X and the 11700K appear to be clock locked at just under 5Ghz according to the .gb5

That is among the fastest 5800X scores I can find, most 5800X score a solid 10% lower. And it is losing to the pre-prod 11700K.

View attachment 37989




5800X GB5 file :

https://browser.geekbench.com/v5/cpu/5960362.gb5

11700K GB5 File :

https://browser.geekbench.com/v5/cpu/5571984.gb5

You know, instead of posting all that pontification stuff you could have just done a direct comparison of the two systems I linked to on Geekbench's site.

You might learn something by looking at the data itself instead of regurgitating AMD fan talking points.

That learning being, these CPUs are better at different things, but they're not being carried by any one thing.

For example, Intel is faster at single thread :

HTM5 decode
SQL
PDF rendering
Text rendering
Camera functions
Machine learning
Speech recognition

Now what was it you were saying about Crypto, which is weighted at 5% of the overall score on Geekbench?

In multi-thread, Comet Lake does even better, taking narrow single thread losses and turning them to wins in things like navigation, HDR, and horizon detection. This implies that Comet Lake is more efficient at scaling with threads than Zen 3 is.

And it's doing it on a Z490. Wouldn't it do better on a Z590? I would think so.

Not that any of this in-the-weeds stuff makes much difference, as my point is not that the 11700K is better than a 5800X. My main point is that the 11700K seems a match for a 5800X.

I posted that, in context, to a bunch of posts about a recent "leak" that implied the 11900K was slower than the 5800X. That's extremely unlikely when the 11700K seems a match for the 5800X.

I'm so tired of you calling people "AMD fans". I know I can't be the only one. No one will take you seriously if you keep doing crap like that. So the 11700k likely matches the 5800X, at over two times the power. Wake me up when Intel gets its crap in order.

 

[dmens]

I'm giving actual reasons why Sunny Cove (or any other core) on 14nm clocks higher than on Ice Lake's version of 10nm. You just don't want to accept those reasons, even ones as direct as more voltage. I'll leave it at that, rather than wasting even more time convincing someone who refuses to listen.

You are being deliberately misleading by comparing an older 10nm process and product when there is a better reference point readily available. Also, the comparison of interest is with 10th gen, not Icelake, since RKL replaces Cometlake. And lastly, all your "actual reasons" are actual nonsense. Cell swaps for 25% delay drop at max volts? You are actually serious? Good lord.

 

[Exist50]

It hasn't borne out in some leaks including results from multiple applications. Some of the earlier GB5 leaks also pointed to maybe a 10% improvement in IPC over Skylake-family cores (Comet Lake, in particular). It would be nice to have launch data on these chips, but until Intel chooses to finally produce review samples, we won't get that data.

I recall similar discussions when Ice Lake itself was released. And that was without the less than ideal scaling that comes at higher clocks. But wait for wikichip et al. if you want to know for sure what architecture it is. I don't mind waiting 2-3 months to be proven right, though I find the entire discussion quite odd after Intel literally said it was Sunny Cove architecture.

 

[DrMrLordX]

though I find the entire discussion quite odd after Intel literally said it was Sunny Cove architecture.

Yet little of the available performance data bears that out.

 

[amrnuke]

You know, instead of posting all that pontification stuff you could have just done a direct comparison of the two systems I linked to on Geekbench's site.

You might learn something by looking at the data itself instead of regurgitating AMD fan talking points.

That learning being, these CPUs are better at different things, but they're not being carried by any one thing.

For example, Intel is faster at single thread :

HTM5 decode
SQL
PDF rendering
Text rendering
Camera functions
Machine learning
Speech recognition

Now what was it you were saying about Crypto, which is weighted at 5% of the overall score on Geekbench?

In multi-thread, Comet Lake does even better, taking narrow single thread losses and turning them to wins in things like navigation, HDR, and horizon detection. This implies that Comet Lake is more efficient at scaling with threads than Zen 3 is.

And it's doing it on a Z490. Wouldn't it do better on a Z590? I would think so.

Not that any of this in-the-weeds stuff makes much difference, as my point is not that the 11700K is better than a 5800X. My main point is that the 11700K seems a match for a 5800X.

I posted that, in context, to a bunch of posts about a recent "leak" that implied the 11900K was slower than the 5800X. That's extremely unlikely when the 11700K seems a match for the 5800X.

I'm not interested in doing a direct comparison - we know nothing about the surrounding system of the 11700K, which makes comparison quite difficult. Most of my post was 1) applauding Intel for making what seems like more progress with RL than AMD with Zen 3 and 2) picking apart the faults of GB5 and 3) trying to sort out WHERE the 11700K gets its benefit from.

Yes, crypto is weighted at 5%. But it carries a much more outsized effect on the test, far higher than the 5% weight it is given. What are you trying to say about crypto and its influence on the total score?

As for your main point - yes! They do seem to be closely matched based on this result! Certainly more closely matched than the 10700K and the 3800XT.

 

[lobz]

I'm giving actual reasons why Sunny Cove (or any other core) on 14nm clocks higher than on Ice Lake's version of 10nm. You just don't want to accept those reasons, even ones as direct as more voltage. I'll leave it at that, rather than wasting even more time convincing someone who refuses to listen.

You can call people ignorant all day if you want, but it's not a straight backport, it has different cache configuration. Just wanted to say that before you inevitably make yourself look like an arrogant fool, should you continue to condescend people with lectures that you got wrong.

 

[Zepp]

I'm so tired of you calling people "AMD fans". I know I can't be the only one. No one will take you seriously if you keep doing crap like that. So the 11700k likely matches the 5800X, at over two times the power. Wake me up when Intel gets its crap in order.

we really just need people to stop entertaining these kind of posters. I put that one on ignore months ago but still have to sift through other user's long back and forth arguments with them in these threads

 

[Exist50]

You can call people ignorant all day if you want, but it's not a straight backport, it has different cache configuration. Just wanted to say that before you inevitably make yourself look like an arrogant fool, should you continue to condescend people with lectures that you got wrong.

It does not have a different cache configuration. What on earth gave you that idea?

 

[dullard]

19% faster in what though?

Intel has it listed on their website:

CES 2021 - 1 | Performance Index

Up to 19% IPC performance improvement (gen over gen)- Source: Intel estimates as of January 2021. Based on measurements on Intel Internal reference platforms running SPEC CPU 2017 1-copy rate on 11th Gen Intel® Core™ i9-11900K vs 10th Gen Intel® Core™ i9-10900K (running each at the same fixed frequency).

This up to 19% refers to "SPEC CPU 2017 1-copy" and only "SPEC CPU 2017 1-copy". Therefore the IPC claim comes from a combination of integer processing, floating point processing, cache, and memory speed.