Discussion Intel current and future Lakes & Rapids thread

[itsmydamnation]

@itsmydamnation Thanks for the test that's great.

Those system power numbers are in line with Notebookcheck systems that are set at 15W. Of course its running on one core so its to be expected.

did some 8T runs, Extreme performance seems to pump voltages for not much gain, its almost like intel cooling voltage logic with a higher wattage would be a good 3rd power option.

wall plugged	time (sec)	power mode	wall measure (w)	hwmon
				Package	Core	1T core	clock
idle		intel cooling	15​	2.5​	0.4​
prime small FFT 1T	15​	intel cooling	37​	19​	18.2​	16.2​	4.2ghz
prime small FFT 1T	60​	intel cooling	37​	19​	18.4​	16.2​	4.2ghz
prime small FFT 1T	180​	intel cooling	30​	12.9​	11.4​	10​	4.0ghz
prime small FFT 1T	780​	intel cooling	30​	12.98​	11.4​	10​	4.0ghz
idle		extreme performance	15​	2.5​	0.3​
prime small FFT 1T	15​	extreme performance	38​	24​	19​	17.7​	4.2ghz
prime small FFT 1T	60​	extreme performance	38​	19​	18.7​	16.2​	4.2ghz
prime small FFT 1T	420​	extreme performance	38​	18.8​	18.7​	16​	4.2ghz
prime small FFT 1T	900​	extreme performance	38​	18.8​	17.7​	17.8​	4.15ghz
prime small FFT 1T	1200​	extreme performance	39​	19​	19​	17.7​	4.15ghz
idle		intel cooling	15​	2.5​	0.4​
prime small FFT 8T	15​	intel cooling	64​	37​	36​	4.7​	3.0ghz
prime small FFT 8T	60​	intel cooling	36​	12​	9.8​	1.29​	2.0ghz
prime small FFT 8T	120​	intel cooling	29​	12​	9.6​	1.2​	1.9ghz
prime small FFT 8T	240​	intel cooling	29​	12​	9.6​	1.2​	1.9ghz
idle		extreme performance	15​	2.5​	0.4​
prime small FFT 8T	15​	extreme performance	65.3​	37​	36​	6.7​	3.0ghz
prime small FFT 8T	60​	extreme performance	65.3​	37​	36​	6.7​	3.0ghz
prime small FFT 8T	120​	extreme performance	65.3​	37​	36​	6.7​	3.0ghz
prime small FFT 8T	240​	extreme performance	50​	27​	24​	3.2​	2.7ghz
prime small FFT 8T	300​	extreme performance	48​	25​	24​	3.2​	2.7ghz
prime small FFT 8T	600​	extreme performance	48​	25​	24​	3.2​	2.7ghz
idle		intel cooling	15​	2.5​	0.4​
Cinebench R20 1T	15​	intel cooling	29​	12​	10​	9​	4.2ghz
Cinebench R20 1T	60​	intel cooling	30.8​	12​	10​	9​	4.2ghz
Cinebench R20 1T	180​	intel cooling	29.4​	11.5​	10.9​	9.7​	4.2ghz
Cinebench R20 1T	900​	intel cooling	28.8​	11.5​	10.9​	9​	4.2ghz

 

[tamz_msc]

It only looks to have lower perf/clock because its clocked nearly 25% higher.

Andrei doesn't seem to think that its the usual clock-frequency vs memory latency thing which drags IPC down with higher clock-speeds.

It's not that straightforward, the TGL system had faster memory and appears to have a much stronger memory subsystem with more bandwidth and lower latency.

The issue here is on the caches which are clocked with the CPUs, and they're simply slower.

 

[TheGiant]

And what are the 2 clocks ? Ice Lake and Tiger Lake that you are using before normalization ?

Icelake 3.9GHz and Tiger 4.8GHz

 

[Markfw]

Icelake 3.9GHz and Tiger 4.8GHz

So what that means is that Tigerlake clocks higher, but needs to, since its slower at the same speed ?

 

[IntelUser2000]

Andrei doesn't seem to think that its the usual clock-frequency vs memory latency thing which drags IPC down with higher clock-speeds.

I doubt it. It makes sense for SpecFP, which is very memory subsystem sensitive and typically shows only 60% scaling.

On SpecInt, its really not. If you go to the site and do comparisons based on clocks it comes out to be ~85%. Of course I don't doubt some subtests are slower on Tigerlake as the cache subsystem changes quite a bit. But overall its still few % better.

Unless he actually does the testing by clocking them identically I won't believe this. If they score the same, Tigerlake has 2.8% advantage, its simple as that. You can also see that it scores few % better by using 28W mode instead of 15W, because Tigerlake goes slightly over that at max ST frequency. Icelake is already at 25W with the Surface Laptop 3.

Also, I'm not sure where people are getting that WC cores are actually slower.
55.28 @ 4.8GHz vs 44.77 @ 3.9GHz in SpecCPU 2006. Even if you assume it does scale perfectly linearly, its identical at the same clock. In SpecCPU2017, its about a percent faster.

This is why clock for clock comparisons are done on desktops.

 

[Thunder 57]

So what that means is that Tigerlake clocks higher, but needs to, since its slower at the same speed ?

I think it's a wash. Really depends on what software one is running. Intel focused on higher clocks rather than IPC with TGL. There was a slide that mention three options I think and they said they went for GHz, but I can't find it ATM.

 

[Hitman928]

@Hitman928 There's something definitely going on with the GPU power figure, but the Slim 7 is not at 15W. It uses more than Intel systems that have 25W settings, and significantly higher than AMD systems at 15W.

I would lean toward agreeing with you but then I can't rationalize their sensor numbers showing 15W operation. Maybe HWinfo is just wrong or looking at @itsmydamnation's numbers, maybe they changed power profiles between tests, or maybe their power profiles just work differently than his, I don't know. Hopefully we get a yoga slim TGL model for comparison to see how it is treated in the same lineup in regards to power/performance.

 

[IntelUser2000]

I would lean toward agreeing with you but then I can't rationalize their sensor numbers showing 15W operation.

I get you. So I looked at other tests too. Sometimes, it would show it being a positive number when its running a CPU-only benchmark. Sometimes its really close to the package number, like being 1W difference. Sometimes the difference is greater.

The places reporting 38W for PL1 figure also makes sense looking at @itsmydamnation's power figures. They must be running at Extreme Performance mode and the long term figure is really 25W, but just like some Intel systems, they are using PL1 similar to how PL2 is.

The difference between using PL1 for a short term versus just using PL2 seems to be the desire to be flexible. PL2 is limited by tau, while PL1 is only limited by thermals. So it can be used under AC(as with the XPS) and changed if in a Tablet mode. At least for Intel systems.

 

[mikk]

Another youtube test:

There is a graphics glitch in Doom Eternal, overall there are several driver issues in different tests and games they have to sort out.

 

[TheF34RChannel]

Alder Lake questions:

1. Will it only be big.LITTLE cores and no regular cores like we have today?

2. Will a gaming or prosumer system benefit from the design and if yes; how?

3. PCI-e 5? Really? Hard to believe.

4. Will it support DDR5?

5. Will we ever get back to normal CPU and mobo prices? 10900K and an Asus hero board are obscenely priced here (more than double the price of what I paid for the current parts).

I won't get a new system until DDR5 support. Anything else seems senseless to me for my own personal needs.

 

[Exist50]

1. Will it only be big.LITTLE cores and no regular cores like we have today?

Sharkbay's leak suggested a 6+0 S die, but the rest had Atom cores.

 

[TheF34RChannel]

Sharkbay's leak suggested a 6+0 S die, but the rest had Atom cores.

A single SKU would be highly disappointing - personally I don't care about little cores. I'd be happy if this were a seperate line altogether but it seems like we're all being forced into a singular, same direction - and I don't like it.

 

[Exist50]

A single SKU would be highly disappointing - personally I don't care about little cores. I'd be happy if this were a seperate line altogether but it seems like we're all being forced into a singular, same direction - and I don't like it.

I'd imagine there'll be an 8+0 SKU as a derivative of the 8+8 S die, but the direction is clearly hybrid everywhere.

 

[TheF34RChannel]

I'd imagine there'll be an 8+0 SKU as a derivative of the 8+8 S die, but the direction is clearly hybrid everywhere.

Very apparently so :/ would be great if you're in the market for that but otherwise I like me some power!).

Wouldn't mind 10-12 cores plus double the treads all clocking to above 5.0GHz on their own before us tinkering. Nope, not a fan of the direction it's headed; what can we do?

 

[A///]

I'd imagine with a big.little approach you could split the clocking of the cores.

 

[DrMrLordX]

A single SKU would be highly disappointing - personally I don't care about little cores. I'd be happy if this were a seperate line altogether but it seems like we're all being forced into a singular, same direction - and I don't like it.

I wouldn't worry about it too much. Lakefield is Intel's pipecleaner for sorting out problems with heterogeneous core configurations on Windows in particular. The scheduler should be able to figure out how and when to prioritize threads by the time Alder Lake-S actually reaches the market (which I'm thinking will be 2022 but we'll see).

The Gracemont cores will be there to pick up background threads and increase MT throughput in 'prosumer'-type applications where raw MT throughput counts. For games and such, the Golden Cove cores will be there to essentially deliver the same performance that you've come to expect from an Intel 8c CPU, albeit with higher IPC than CoffeeLake/Comet Lake, Rocket Lake, or Willow Cove/TigerLake-H. Not sure what the clocks will be like. Assuming Intel sticks with 10SF, it'll be more power-limited than anything else. But it's a ways out so we'll see.

 

[eek2121]

A single SKU would be highly disappointing - personally I don't care about little cores. I'd be happy if this were a seperate line altogether but it seems like we're all being forced into a singular, same direction - and I don't like it.

It sounds like the Gracemont cores perform pretty well. An 8+8 chip would be an interesting one to keep an eye on.

I am contemplating ordering that Yoga laptop to replace my current laptop that has a 6xxx chip. If I decide to pull the trigger I will be happy to provide whatever information is requested as long as it doesn’t involve overclocking.

 

[Bouowmx]

Is Rocket Lake still coming 2020 Q4 or 2021 Q1 with Sunny/Willow Cove 5 GHz on 14 nm? Haven't heard about Rocket Lake in a bit.

 

[Shivansps]

Ok after watching a lot of reviews and read about the issues i can sit down and say im shocked that intel was able to beat a Vega 8 IGP in the first implementation of Xe Gen 12 graphics with a driver that still has graphical errors and is not fully DX12/Vulkan compilant. Im not sure if Raja had anything to do with that but it is already good and it has a lot of room to improve with drivers. It is AMD fault for keep using Vega igps for far too long but still, Intel has something here.

 

[Thala]

Ok after watching a lot of reviews and read about the issues i can sit down and say im shocked that intel was able to beat a Vega 8 IGP in the first implementation of Xe Gen 12 graphics with a driver that still has graphical errors and is not fully DX12/Vulkan compilant.

Making a driver more compliant does not make it faster - the contrary could very well be the case.

 

[Shivansps]

Making a driver more compliant does not make it faster - the contrary could very well be the case.

If the driver is not feature complete it means it still has to go trough the optimization phase. It is likely that very little or no optimization has been done so far.

 

[nerp]

Is Rocket Lake still coming 2020 Q4 or 2021 Q1 with Sunny/Willow Cove 5 GHz on 14 nm? Haven't heard about Rocket Lake in a bit.

2021, Q1 or Q2. Yes on arch, no certainty on freqs (depends on skus) 14nm.

Ok after watching a lot of reviews and read about the issues i can sit down and say im shocked that intel was able to beat a Vega 8 IGP in the first implementation of Xe Gen 12 graphics with a driver that still has graphical errors and is not fully DX12/Vulkan compilant. Im not sure if Raja had anything to do with that but it is already good and it has a lot of room to improve with drivers. It is AMD fault for keep using Vega igps for far too long but still, Intel has something here.

You get it. TL is a beast and Intel has a good product for their do-over. AMD cannot make one mistake or they're behind again. I will get flamed for this, but AMD is 1-2 years away from being left in the dust forever based on what we see with TL. I caution people to set reasonable expectations for Zen 3.

till has to go trough the optimization phase. It is likely that very little or no optimization has been done so far.

Correct. Won't be massive change in what we see now, but it will be something.

 

[DrMrLordX]

I will get flamed for this, but AMD is 1-2 years away from being left in the dust forever based on what we see with TL.

Meh. It's not that good. Intel can't/won't even yield more than 4c parts on 10SF. TigerLake-H should have been in the lineup day one, and it's MIA. "Coming soon".

 

[LightningZ71]

I really don't see either AMD or Intel leaving each other in the dust anytime soon. This isn't like it was eight to ten years ago where Intel's foundries were hitting on all cylinders and Ad was stuck short funded and fighting with being multiple nodes behind. It seems like most of the IPC advances that we're seeing these days involve trade-offs, like cache restructuring that helps some loads, but hurts a few, or pipelinelength tweaks that help clocks but hurt branches. You get give of transistors thrown at things like machine leartgat is currently barely used, or super wide vector units tgare usable by maybe 5% of your customers, or so many cores that some essentially stay idle all the time.

The competition is going to be as much on system level optimization as it is on core improvements. While Intel is pushing big.LITTLE, AMD is pushing in other directions. Once Intel gets enough OS level buy in, AMD can look at big.LITTLE if they need to.

I just don't see either company performing a paradigm shift anytime soon.

 

[nerp]

Meh. It's not that good. Intel can't/won't even yield more than 4c parts on 10SF. TigerLake-H should have been in the lineup day one, and it's MIA. "Coming soon".

Not going to deny Intel is digging out of a hole. A lot of "woulda shoulda coulda" going on at Intel. But 8c TL will come, the future lakes will come, etc. FOr sure, question is when. But execution is key, lots of factors need to go right. But I expect Intel to address that more than I expect AMD to not make a single mistake or have any setbacks here on out. That is where I place my bet.