Discussion Intel current and future Lakes & Rapids thread

[uzzi38]

It probably won't change things either way. Cometlake is already a lost cause in very parallel applications.

But in other applications it may turn out even better. Lots of multithreaded applications where core scaling is poor. Also the new uarch will bring gains in the areas Skylake is weak in.

Games are going to love the new uarch as they always do. If you see GN tests, they scale poorly with clocks, and worse with cores. Just 20% gain there will wreck most chips even at 4.5GHz.

IPC vs core clocks doesn't change much for games. They both have the same effect more or less. Also, a 20% IPC gain at 4.5GHz is still roughly equal in performance to 5.4GHz on the original uArch, so I'm not sure what your point is there? That's higher ST performance than Comet Lake.

Rocket Lake would need to clock higher than that and still provide that IPC upgrade of a notable improvement in games over Comet Lake.

Memory latency is something that drastically improves performance on both AMD and Intel systems right now. If you want to improve gaming performance significantly, lower latency access to memory is the best way to do it.

 

[IntelUser2000]

@uzzi38 That's not the case. Perf/clock(the more accurate term) not only improves pretty much as is, but also addresses weaknesses of the previous architecture. You can clearly see that from the Zen 2 review.

Clock scaling is far poorer in games than applications. You can typically get 80% in SpecCPU or Cinebench. You're lucky to get 50-60% in games, and that's only when you are running on settings and tests that show the difference.

AMD Ryzen 5 1600 AF CPU Review & Benchmarks: Best CPU Under $100 in 2020 | GamersNexus

stub This isn’t a revisit of the old AMD Ryzen 5 1600 – it’s a review of the new variant, named the AMD Ryzen 5 1600 “AF” by the community, dubbed as such for its SKU change from AE to AF. The AMD R5 1600 AF is a brand new CPU with an old, old name from 2017. It’s mostly an R5 2600, in that it’s...

www.gamersnexus.net

Overclocking it from 3.6GHz boost to 4.2GHz gets you 5-7% gains resulting in 30-40% scaling. This is typical. I don't know why anyone believes otherwise. Games are somewhat GPU bound even in the 100 fps+ range. You'll never get the scaling like applications because majority of compute is run by the GPU. Before Turbo wasn't as complex, I used to sit down for hours and analyze data.

Now this assumes a 14nm port of Willow/Sunny Cove cores are going to be identical to the 10nm version. It might not. But if it does its a much needed change.

Skylake is a core that's going to be in kindergarten soon. It's going to suck.

 

[jpiniero]

Benchmarks - OpenBenchmarking.org

openbenchmarking.org

Tiger Lake 2.3/4.2 benchmark results. Note it is no faster than what Phoronix got on the 1065G7.

 

[mikk]

The sample is very old, it's a stepping 0. The exact same bios version is known from some old Geekbench entries. Not really interesting therefore.

 

[yuri69]

24c Ice Lake - Family 6 Model 106 Stepping 4 - 1.2-2.9GHz, median 2.8GHz but with weird fluctuations - 4048/41962

24c EPYC 7402P - median 3.3GHz with stable frequency - 4498/42155

Ice Lake IPC is really solid. Let's see how is the power consumption.

 

[Markfw]

24c Ice Lake - Family 6 Model 106 Stepping 4 - 1.2-2.9GHz, median 2.8GHz but with weird fluctuations - 4048/41962

24c EPYC 7402P - median 3.3GHz with stable frequency - 4498/42155

Ice Lake IPC is really solid. Let's see how is the power consumption.

Thats no benchmark, where did you get that from ?

 

[jpiniero]

You have to look at the tests themselves. The Epyc does win most of the tests, although there are some where it is close. The HTML 5 DOM and SQL List MT score got screwed up somehow.

 

[Markfw]

You have to look at the tests themselves. The Epyc does win most of the tests, although there are some where it is close. The HTML 5 DOM and SQL List MT score got screwed up somehow.

OK, where is a link to the tests themselves ? I don't see it.

 

[jpiniero]

OK, where is a link to the tests themselves ? I don't see it.

empty S5645 vs GIGABYTE MZ32-AR0-00 - Geekbench

browser.geekbench.com

 

[yuri69]

Here is a "GUI view" available.

 

[LightningZ71]

It seems quite apparent that the larger L1 dCache and double the L2 cache are big uplifts in certain workloads. The difference in memory latency is nontrivial and the multicore measured bandwidth is eye opening. This should be quite interesting to watch...

 

[Markfw]

empty S5645 vs GIGABYTE MZ32-AR0-00 - Geekbench

browser.geekbench.com

Well considering they(the Ice Lake) are not out yet, and neither is the new Milan, I would say that puts them at best case, one generation behind AMD, as Milan and Ice Lake may come out at the same time. Also, how many cores will Ice lake have as a max ?

Not impressive IMO.

 

[IntelUser2000]

The difference in memory latency is nontrivial and the multicore measured bandwidth is eye opening. This should be quite interesting to watch...

Where are you seeing this? It doesn't stand out compared to Rome. Also according to the comparisons Rome offers 50% more bandwidth for a single core.

The differences are 24% per clock if we assume the results are accurate.

Intel has no problem against AMD in uarch and perf/clock. But they can't get enough cores to save themselves. It'll be 38 cores versus 64.

 

[jpiniero]

It'll be 38 cores versus 64.

24... 24 is the HCC, 38 is the XCC. I'm sure they will announce it but it will be MIA due to the 10 nm yields. Cooper was supposed to handle that market.

 

[Gideon]

Well considering my stock 3700X with 8c/16t does 5840/36695 doesn't seem that impressive.

Ice lake pretty much only performs well in AES an SGEMM (both AVX-512 tests). I'm aware that my 3700x runs at almost double the frequency most of the time, but then it should be about equal to 16 core not 24 core Ice lake.

 

[Gideon]

I really wish they had done geekbench 5.0 and Linux benchmarks though. Geekbench 4 scales horribly with windows, just compare EPYC 7742 results between Windows and Linux, run buy the same reviewer (from Servethehome):

AMD Corporation DAYTONA_X vs AMD Corporation DAYTONA_X - Geekbench

browser.geekbench.com

MT score 41576 (windows) vs 187446 (linux) same chip ...

 

[yuri69]

Yea, testing servers chips on Windows is a joke but this is what we have now.

AMD will keep having the upper hand till Intel really adopts a form of multi-die as a mainstream approach. Gluing two ultra-expensive 38c dies together won't help 'em given the CLX-AP failure. Too bad Sapphire Rapids configurations are still murky.

I'm looking forward to see perf/W of ICL chips compared to Rome and/or ARMv8.

 

[ondma]

Well considering my stock 3700X with 8c/16t does 5840/36695 doesn't seem that impressive.

Ice lake pretty much only performs well in AES an SGEMM (both AVX-512 tests). I'm aware that my 3700x runs at almost double the frequency most of the time, but then it should be about equal to 16 core not 24 core Ice lake.

I dont think that your extrapolation is valid (8 cores vs 24). As you said, it looks like a scaling issue. Your 3700x multi score is about 6.3 times the single core, or almost 80% of the expected value (100% scaling with cores). Both icelake and AMD show less than 50% scaling. So if you use your scaling logic, *neither* AMD or icelake is impressive.

 

[DrMrLordX]

Also, how many cores will Ice lake have as a max ?

Supposedly 38c for IceLake-SP

24... 24 is the HCC, 38 is the XCC. I'm sure they will announce it but it will be MIA due to the 10 nm yields. Cooper was supposed to handle that market.

Cooper is still there if you want a 4P board. We don't know availability on the XCC die yet. It's so delayed that Intel might pull it off anyway. Or maybe not.

 

[jpiniero]

Cooper is still there if you want a 4P board.

It's not going to be sold to the general public, and Facebook might be the only customer. Cascade Lake Refresh is what the public gets.

 

[DrMrLordX]

Yay, availability lower than Cascade Lake-AP?

 

[Ajay]

It's not going to be sold to the general public, and Facebook might be the only customer. Cascade Lake Refresh is what the public gets.

And for investors, 'see, we delivered ICL-SP as promised!'. Meh. 7nm or bust for Intel.

 

[uzzi38]

24c Ice Lake - Family 6 Model 106 Stepping 4 - 1.2-2.9GHz, median 2.8GHz but with weird fluctuations - 4048/41962

24c EPYC 7402P - median 3.3GHz with stable frequency - 4498/42155

Ice Lake IPC is really solid. Let's see how is the power consumption.

Geekbench does not report all core sustained clocks. Clock speed reporting is done right at the very beginning before multi-core workloads are run.

A while ago I ran this test on my laptop with 1T and 2T multiplier capped at 38x and everything else capped at 35x using Throttlestop.

Not once was 3.5GHz reported.

I'm going to try something else, but I have a hunch it's like 3DMark in that it reports clocks before ANY tests are run, forget MT tests.

EDIT: Trust me to not link the test.

PC Specialist LTD N8xEJEK - Geekbench

Benchmark results for a PC Specialist LTD N8xEJEK with an Intel Core i5-8300H processor.

browser.geekbench.com

 

[lobz]

Yea, testing servers chips on Windows is a joke but this is what we have now.

AMD will keep having the upper hand till Intel really adopts a form of multi-die as a mainstream approach. Gluing two ultra-expensive 38c dies together won't help 'em given the CLX-AP failure. Too bad Sapphire Rapids configurations are still murky.

I'm looking forward to see perf/W of ICL chips compared to Rome and/or ARMv8.

The real joke here is GeekBench 4

 

[IntelUser2000]

Geekbench does not report all core sustained clocks. Clock speed reporting is done right at the very beginning before multi-core workloads are run.

It's a fair point if you are correct but you can see the breakdown of the clock speeds. That's why median clocks are reported.

Both icelake and AMD show less than 50% scaling. So if you use your scaling logic, *neither* AMD or icelake is impressive.

I think what's happening is it doesn't scale well above 16 threads.

From what I observed:
-4 cores offer ~3x improvement
-SMT adds ~20%

In that case its saying even his(@Gideon) system starts to fall off in scaling since it only offers 6.3x with 8x cores.