Intel Cannonlake, Ice Lake, Tiger Lake & Sapphire rapid thread

Page 160 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

mikk

Platinum Member
May 15, 2012
2,705
590
136

IntelUser2000

Elite Member
Oct 14, 2003
6,831
1,392
136
As equal to what? Who's to decide "how benchmarks are done"?
You run comparison systems as equal as possible, down to the driver versions and the software installed. Ideally both will be run on systems with fresh OS install.

When you are in the business of selling systems they do what they want, and despite somewhat misleading tactic of comparing against 5-year old systems, the truth is PC sales are mostly replacement sales so it serves their purpose.

But as a reviewer and running benchmarks different standards apply. Yes its unspoken but that's the golden standard. Here I'm disagreeing with the decision to run them at default behavior instead of running it fixed.

The rate benchmarks in their characterisation are almost identical to the speed versions, with runtime being by far the biggest difference, and that imagick is missing from the rate suite.
Ok, thanks for the explanation.

I'm curious why the 2017 Speed tests exist if its actually that different with one test not existing at all and if manufacturers don't use it.
 

IntelUser2000

Elite Member
Oct 14, 2003
6,831
1,392
136
That's just because fewer EUs stress mem BW a lot less - well... theoretically it's not scaling, but in the end in practice it is, so you're right anyway :)
Hmm. The 64EU G7 Iris results are shown as 30-35% faster compared to 48EU G4 Iris, however there's a caveat: The only G4 Iris result is passively cooled, while the G7s are actively cooled and likely set at 25W too.

Roughly speaking a balanced architecture's performance characteristics are split 1/3rd each between fillrate, shader throughput, and memory bandwidth.

I've seen another benchmark out there showing 20% difference between G4 and G7 illustrating the 1/3rd rule.

In Intel GPU architectures, the unslice doesn't change and that'll also serve to limit scaling from being linear. There's also that lower end GPUs get lower fps meaning there's less impact due to the CPU. Of course the memory bandwidth is the biggest.

Not bad if this is a base clock. Also the iGPU is fully enabled in this model:
If 2.7GHz is really the base clock then its top notch on par with 28W version of the Coffeelake mobile CPUs.

So on a slightly different topic. A manufacturer has said their Icelake system is set for 28W. At that point I wonder why the 28W version existed at all. Not only Intel allows choosing TDP in finer granularity between official TDP ratings, but you can go above it as well.

Furthermore modern laptops don't have a set TDP. For 2-in-1s it might be at 12W when in Tablet mode, 15W in laptop mode, but setting higher performance settings could boost that to 25W. This is a good thing for buyers, but makes comparisons difficult as reviewers are unaware of it.
 
Last edited:

uzzi38

Senior member
Oct 16, 2019
815
962
96

Not bad if this is a base clock. Also the iGPU is fully enabled in this model:

subslice total: 6
slice0: 6 subslices
EU total: 96
EU per subslice: 16

And Tigerlake is indeed a 10++ product, Intel confirmed it. The "new transistor optimization" makes sense now.
-U on TGL is 28W.

Still a nice uplift though, and extra 400mhz on the base clock at 28W.
 

lobz

Golden Member
Feb 10, 2017
1,200
1,164
106
Last edited:

liahos1

Senior member
Aug 28, 2013
571
44
91
Oh my God, so it's true? Skylake is better than Zen+ ??????????? No way...

Joking aside, I would not be very happy if I were Intel, looking at how the Iris Plus still can't beat the 14nm Vega11 in gaming.
that's icelake bruh
 

krumme

Diamond Member
Oct 9, 2009
5,872
1,463
136
that's icelake bruh
Yeaa but who can blame him. Same cpu perf as top end skl arch. Where is the tdp?
The 64eu and 2x3733 ram makes them on par for gaming vs the vega11 with some cheap ddr4 2400.
AT can paint this in the usual Intel friendly way. This expensive solution is just more a step backwards and shows why the regular skl 14nm is needed. Makes better sense as a business solution as it can do the same just cheaper. Amd will dump those 12nm apus all over the market for cheap consumer lines. My guess is it will far outsell 10nm Intel products and any 7nm fancy new amd apu as they are forced to move gf capacity.
 

krumme

Diamond Member
Oct 9, 2009
5,872
1,463
136
its 100 dollars more than the ryzen sku lol. Tigerlake clocks will improve as will IPC and GPU performance. you guys are delirious with hatred
Yeaa i saw that nonsense pricing. Not happening in the real world when consumers buy in large numbers. Marketing pricing.

Ipc will stay the same. You can argue that if they solve the fmax problem in perhaps 3 years time, that ipc actually is a tiny bit lower then. Its not relevant as there is a new arch then and a hopefully working euv 7nm node.

This product is like a tuned tractor driving 100mph, and is build from Formel 1 parts.
It adresses markets that is better served by either a standard old 14nm skl with or without a discrete gpu or a cheap 12nm amd apu.
 

ajc9988

Senior member
Apr 1, 2015
250
155
116
its 100 dollars more than the ryzen sku lol. Tigerlake clocks will improve as will IPC and GPU performance. you guys are delirious with hatred
On what basis do you think that tiger lake will improve clocks and by how much. Intel's 10nm and 10nm+ are having issues of brittle cobalt, among other things, while being plagued with leakage issues leading to higher temps preventing higher frequency. Please support for hypothesis because Intel will be four years late on 10nm in sizeable volume, is putting out 1/3d the normal amount of mobile skus on the process, killed off 10nm, and had yields at 12% like a year ago and will not give a percentage on yields, instead just repeating that it is improving.

As to IPC, Intel did not focus on ST for Willow cove. So this will not be the IPC jump of sunny cove. Further, it will go against Zen 2 on mobile and won't be on desktop at all, where it would have to go against Zen 3 with a rumored 10-12% IPC increase on integer and 17% IPC on used workloads.

Rocket lake is supposedly a Willow cove back port to 14nm+++, but won't arrive until 2021, meaning it would compete against Zen 4 with another IPC improvement, both having new memory subsystems for DDR5 giving an increase, along with I/O improvements otherwise, and Zen 4 will get a new socket which should allow for much better support than forcing conformity into the old socket.

So, even though you could be correct on the hate of others, you still haven't supported your claim. I'll go back to see what your post before this one was to be sure though.

Edit: and to clarify on my comments on frequency for Tiger Lake, base frequency increases do not necessarily correlate with boost frequency increases. A person pointed to the higher leaked probable base clock on a tiger lake benchmark. This would point to correcting some leakage issues with the process, which is good. But it does not mean the boost frequency is significantly changed. If that is the case, then Intel still may have issues in competing with a Zen 2 mobile chip, although I will conceded it will likely win. Just the margins of that win are the question.

In desktop and server, I have my reservations about Intel's 10nm capabilities.
 
Last edited:

inf64

Platinum Member
Mar 11, 2011
2,969
1,443
136
its 100 dollars more than the ryzen sku lol. Tigerlake clocks will improve as will IPC and GPU performance. you guys are delirious with hatred
If you think that intel has ANY chance competing with Zen3/Zen4 lineup with their 14nm/10nm SunnyCove derivatives you are the one who is delirious. It will be Zen2 Vs Skylake rerun as Zen 3 is a huge IPC uplift with Zen4 not that far behind. Intel is the one playing the catchup game this time while being on inferior process node(s). The game has changed since 2017, it's a whole other ball game now.
 

jpiniero

Diamond Member
Oct 1, 2010
7,801
1,115
126
If you think that intel has ANY chance competing with Zen3/Zen4 lineup with their 14nm/10nm SunnyCove derivatives you are the one who is delirious. It will be Zen2 Vs Skylake rerun as Zen 3 is a huge IPC uplift with Zen4 not that far behind. Intel is the one playing the catchup game this time while being on inferior process node(s). The game has changed since 2017, it's a whole other ball game now.
Gaming wise it'll be OK. Pure MT, yeah... you won't see more than 10 cores on mainstream for awhile.
 

A///

Member
Feb 24, 2017
181
57
71
Gaming wise it'll be OK. Pure MT, yeah... you won't see more than 10 cores on mainstream for awhile.
I think Zen3 will close the gap in gaming, but it won't extend past it. I'm as close to reality as possible. MT will improve yet again leaving Intel more in the dust, ST gaming will only match Intel, even their 10th gen products. Raw hardware is great but at the end of the day a lot of games were optimized for Intel. Ideally, people should get equal optimization but that won't ever happen.
 

uzzi38

Senior member
Oct 16, 2019
815
962
96
Er, I never thought I'd be saying this, but @liahos1 is right you know... Tiger Lake will be better in terms of ST and iGPU performance than Renoir, and will clock notably better than Ice Lake as well... the biggest problem with 10nm is the yields, not the clock-ability anyway. ICL-U was a little on the rushed side, so clocks worse than it should.

AMD will be in the lead for MT performance and maybe power efficiency with Renoir (7nm+LPDDR4X will rein in idle power draw very significantly), but let's not kid ourselves and pretend Renoir is a full on spanking.

Also, anyone that thinks Zen 3 will be the same as Willow Cove IPC-wise is kidding themselves. It'll be Sunny Cove level, maybe a little higher depending on workload, but I highly doubt it will flat out beat Willow.

That's just for mobile though. Intel are still in a rough patch for desktop mind you.
 

mikk

Platinum Member
May 15, 2012
2,705
590
136
On what basis do you think that tiger lake will improve clocks and by how much.

There are new transistor optimizations for Tigerlake says Intel, in fact it's 10nm+ vs 10nm++, and some of the earlier leaks were promising, this is the basis.

As to IPC, Intel did not focus on ST for Willow cove.
Willow Cove could improve the IPC by 5-10% though. Willow Cove is not a typical refresh generation, the L2 and L3 cache sizes are a lot bigger.
 

uzzi38

Senior member
Oct 16, 2019
815
962
96
Eh, so far Willow Cove is a cache redesign over Sunny Cove. It will take more than a cache redesign to gain significant IPC from Sunny Cove.
I'm doubling down on what Mikk said above you - Willow is like 5-10%. Zen 3 and Sunny Cove should be within 3% of one another (with Zen 3 in the lead on average), and Willow sitting just slightly ahead of both.

EDIT: I'm referring to IPC only. Not accounting for clocks.
 

OriAr

Member
Feb 1, 2019
63
35
51
The one GB we have of TGL suggests Willow Cove has better IPC than Sunny Cove by 9%, although it ran at an unusually low clock, so we are probably looking at something like 7% increase in IPC from Sunny to Willow, not a trivial increase by any means
 

uzzi38

Senior member
Oct 16, 2019
815
962
96
The one GB we have of TGL suggests Willow Cove has better IPC than Sunny Cove by 9%, although it ran at an unusually low clock, so we are probably looking at something like 7% increase in IPC from Sunny to Willow, not a trivial increase by any means
It was actually a little over 10%, but yes, running the two at 500mhz is not gonna end up giving you an accurate IPC figure, so it's a complete waste of time.

But I mean, before it I was saying <5%. Now I think it's 5-10%.
 

OriAr

Member
Feb 1, 2019
63
35
51
I think 7% on average will be a fair estimate for the IPC increase between Willow and Sunny Cove
 

ajc9988

Senior member
Apr 1, 2015
250
155
116
There are new transistor optimizations for Tigerlake says Intel, in fact it's 10nm+ vs 10nm++, and some of the earlier leaks were promising, this is the basis.



Willow Cove could improve the IPC by 5-10% though. Willow Cove is not a typical refresh generation, the L2 and L3 cache sizes are a lot bigger.
Yes, and look at the cannonlake 10nm low boost clocks, power consumption, etc. Look at the yields on that being so low. Then look at 10nm+ having a Max single core boost of 3.9GHz and much lower frequencies after the boost. Now, to be fair, the base clock was abysmal, and as I mentioned there is a ton of leakage.

Intel has been working on 10nm for longer than it was supposed to be implemented in 2016. Last year in 2018, yields were around 12%.

So, I ask again, other than Intel saying it is a refinement, which I will accept, how does that mean ANYTHING at this point considering what we have seen of Intel's 10nm to date and what proof is there of any significant change in frequency?

I even pointed to the leak with the higher clock. This is usually base clock. But base is not boost and you can improve performance on the more efficient end of the curve while not really moving the needle much as you approach the voltage knee for higher frequencies. This is why my edit explained those leaks only necessarily suggest they have worked out some of the leakage issues at the base frequencies to improve efficiency, which may not correlate with effecting the boost frequency in any significant way.

As to Willow IPC, when it makes it into desktop silicon, that will be in 2021. That is Zen 3 plus Zen 4 IPC gains. To put this into perspective, that means AMD has had the following IPC increases: 53%, 3-4%, 15-17%, expected 10-17% (redgamingtech rumor of 10-12% IPC on integer for Zen 3 and 17% average for mixed workloads). Zen 4 is a smaller process node allowing them to go wider, has a new memory subsystem and I/O changes, new AM5 socket, etc. So, there is a very real possibility to see 14nm++++ Willow lake going against a slower 5nm Zen4, but with enough of an IPC advantage to eek out the win on performance.

Intel's sunny cove was a huge IPC jump. Willow on 10nm++ will be on mobile only next year, and as explained, even though they suggest an increase in base clock, there is little to suggest changes in boost clock, due to voltage and heat considerations of what has been seen of 10nm to date.

So I applaud their architecture skills and designs. But Intel's word on process is, well, lacking these days. There is very little to believe from them at the moment. This is why Intel was vaporware memed equivalent at the hot chips conference when they tried at an IBM presentation to suggest they had a similar project they were working on and everyone basically laughed.

Edit: to be clear, Intel on mainstream and server seems to be behind due to process moving forward and back porting will help, but don't know if enough. On mobile, because AMD lags on releasing their current tech until around the Q1 of the year after release, AMD looks like it will remain behind on mobile performance moving forward, at least so long as Intel can deliver on 7nm in 2021-22 and doesn't see further issues delaying that significantly, while also needing to get yields up further on 10nm.
 
Last edited:

mikk

Platinum Member
May 15, 2012
2,705
590
136
Yes, and look at the cannonlake 10nm low boost clocks, power consumption, etc. Look at the yields on that being so low.

Then you should realize how much better 10nm+ was compared to the original 10nm, I mean this is a good example. The first 10nm product wasn't even fully enabled, it was a partially enabled dualcore wihout iGPU and a turbo speed of up to 3.2 Ghz and with 10nm+ they can go up to 3.9 Ghz with better yields. So what is unclear to you? That Intel is able to improve the 10nm process with new iterations of it?
 

ajc9988

Senior member
Apr 1, 2015
250
155
116
Then you should realize how much better 10nm+ was compared to the original 10nm, I mean this is a good example. The first 10nm product wasn't even fully enabled, it was a partially enabled dualcore wihout iGPU and a turbo speed of up to 3.2 Ghz and with 10nm+ they can go up to 3.9 Ghz with better yields. So what is unclear to you? That Intel is able to improve the 10nm process with new iterations of it?
They won't give yield numbers because it is likely still in the high teens to low 20s. Anand tech noted the number of SKUs is 1/3d their normal stack. There is low volume available of 10nm+ based ice lake chips from ODMs and OEMs.

If we examine 14nm for clues, this may shed light. For broadwell c type chips, we saw closer to 4.4GHz OC as normal. For skylake, we got 4.8. For Kaby, 5.0 was more common with 14nm+. Then with 14nm++ and 14nm+++, we arrive at 5.1-5.3GHz for normal cooling with a mid to powerful AIO or needing an open loop.

Here, you just claimed improved yields and a 600MHz increase. The prior nodes improvements got 900MHz and once you get towards the voltage knee I mentioned, the average overclock improvement and end of the boost spectrum slowed. Intel used to leave a lot of frequency on the table, something that has ended since the competitiveness of Zen, et seq.

So what is unclear is the level they can achieve above what they have. My current hypothesis is for desktop chips, considering that ice lake was 900MHz lower than cometlake boost of 4.8, which is 500MHz less than single core boost on desktop, that we will see for desktop frequencies in the mid-4GHz range, similar to AMDs. If that is correct, then we are entering an IPC race.

So, yes, I see your point. Yet thinking they magically will significantly boost the frequency for the boost end seems awfully wishful.

Edit: and to be clear, new architecture + new iteration on process from Broad well to Kaby was about 600MHz, similar to what Intel got from cannon to ice.
 
Last edited:

ASK THE COMMUNITY