Intel Cannonlake, Ice Lake, Tiger Lake & Sapphire Rapid Thread

Page 102 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

coercitiv

Diamond Member
Jan 24, 2014
3,245
605
136
Articles are also saying it needs work from the software side. So perhaps something at least like a patch is required.
At hardware level they already optimized responsiveness for short time windows, I reckon the data needed to optimize for user specific workloads & behavior is mostly to be found at software level. Maybe it's a hybrid approach, learning is done at OS level and the resulting model is running at hardware level for increased responsiveness?
 
Oct 14, 2003
6,084
218
126
The digital portion of the CNVi WiFi uses Intel's 14nm process, while the analog portion of it uses UMC's 28nm process.

It was few years ago when EETimes had an article that said Intel was the largest licensee of ARM chips. Probably the largest portion of that is due to their chipsets.

Intel uses TSMC for other product lines as well. Quite extensively.

At hardware level they already optimized responsiveness for short time windows, I reckon the data needed to optimize for user specific workloads & behavior is mostly to be found at software level.
I've been thinking the software thing has to do with taking advantage of the deep learning part?
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
Yup. We'll have to see what the shipping quantities are on these IceLake U/Y products and try to compare that to some kind of fudged wafer allocation data that in the end we'll all make up anyway. Intel isn't just going to tell us this stuff, at least not anytime soon.
There's plenty they can cut. The i3 has HALF the cores, HALF the GPU EU and HALF the L3 cache.

And there's certainly more they could cut, the GPU could be cut entirely, they could cut the Thunderbolt.
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
Also 9W for Y is really high and much higher than I expected.
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
Seems that the 4.1 max clock on Icelake is for the 28 W models. The 15W Comet from that leak had a max of 4.9 on the Quad Core and 4.6 on the Six, although it probably won't actually hit 4.6+ unless it was TDP'd up.
 

coercitiv

Diamond Member
Jan 24, 2014
3,245
605
136
Seems that the 4.1 max clock on Icelake is for the 28 W models.
Max tubo boost shouldn't be affected between 15W and 28W, it's 1C boost anyway.

The 15W Comet from that leak had a max of 4.9 on the Quad Core and 4.6 on the Six, although it probably won't actually hit 4.6+ unless it was TDP'd up.
They already have 4.8Ghz max turbo on 15W Whiskey Lake i7 8665U, probably when Thermal Velocity conditions are met.
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
Max tubo boost shouldn't be affected between 15W and 28W, it's 1C boost anyway.
You'd be surprised. NBC tested 28 different 8565U notebooks and had a pretty wide spread of scores in Cinebench R15 ST from 160 to the top score which is 191. The median is 173 though, which is roughly what the desktop i3 7320 gets (4.1 Ghz).
 

sxr7171

Diamond Member
Jun 21, 2002
5,066
5
91
So they announced 10th gen icelake today. Apparently they are shipping but laptops with it won’t ship until the holiday season. What does that mean?
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
So they announced 10th gen icelake today. Apparently they are shipping but laptops with it won’t ship until the holiday season. What does that mean?
Intel said they are shipping chips to OEMs but it's not even a paper launch. The Dell looks like it will be available in July but only in extremely limited quanitities.
 
Jul 24, 2000
108
2
81
Lots of thunderbolt ports and graphic units along with lower clocks. Wonder if this is due to bad yields and the ability to salvage chips.
 

.vodka

Golden Member
Dec 5, 2014
1,071
133
136


Performance results are based on testing as of dates shown in configuration and may not reflect all publicly available security updates. See configuration disclosure for details. No product or component can be absolutely secure. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products.
That disclaimer... How are we supposed to trust these numbers? I mean, we have that Geekbench leak showing ~+15% increase, but still.

Is there another slide around that has the setup and benchmark details?
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
Also of note, the Dell that has Icelake now includes fans, as it needs to since it is shipping with 15 W CPUs. The current model ships with Amber Lake and is fanless.
 
Apr 27, 2000
11,857
1,048
126
That Dell model is in need of a teardown. Here's hoping AT gets one through back channels.
 

TheGiant

Senior member
Jun 12, 2017
410
69
86
Also of note, the Dell that has Icelake now includes fans, as it needs to since it is shipping with 15 W CPUs. The current model ships with Amber Lake and is fanless.
I suspect I am out of luck these days
fan on the x570
now fan with Icelake...

looks like icelake has some hotspots that cant be solved right now, like with my skylake i5 6300u
 
Apr 1, 2015
193
12
116




That disclaimer... How are we supposed to trust these numbers? I mean, we have that Geekbench leak showing ~+15% increase, but still.

Is there another slide around that has the setup and benchmark details?
I did a bit of a write up on this earlier today. I am reposting it with mild editing (not shown):


Around Kaby to Coffee, due to changes in the cache sizes, there was an appreciable (more than variance) increase in IPC, even if limited. Cache changes will effect IPC, are hard to separate from IPC, and should be noted.

Moreover, I noticed that the Intel graph DOES NOT HAVE an X-axis label on their infographic. THAT is a problem. Take a look for yourself:
1559086667466.png
1559086696288.png

So, SPEC 2016 & 2017 were used, Sysmark 2014 SE, WEBXPRT, and CB15 were used in the IPC slide (which doesn't separate them out nor have any label for the X-axis to understand what the increase is relative to in the slide). For the Core Performance slide, they used SPECInt 2006. All of these were internally compiled, which Intel has been caught before using flags in compilers that would NOT be allowed in a commercially released product (not saying that occurred here, just saying it has occurred before).

Have you seen what sort of scores can be reached on first gen Ryzen 1950X when using Linux and the -O3 flag? Night and day versus regular compiling without platform optimizations. Intel specifically claims that they were compiled for the specific platform and that using it on any other platform could materially change the results. That sounds like hyper-optimization.

But, aside from those things, look at the benchmarks used. Are any of those your "go to" on benchmarks to estimate performance? CB15 gets the closest, yet then you run into the same criticism that is hoisted upon AMD.

Let's dive into the impact on mitigations. Start with Sysmark 2014 SE by BAPco.


https://www.anandtech.com/show/1365...ith-spectre-and-meltdown-hardware-mitigations

As you can see, there is a 6% drop in performance from before being patched to after. Now the article shows lower numbers with the overall categories, and it doesn't give us the full breakdown by test, but it should be remembered that the impact is seen more on elevated requests for I/O processing, which is not going to impact consumer programs as much, and which the bulk of Anandtech's testing takes place on, versus what is seen in professional workloads, which those benches more closely address. But the point is that, if the tests used have an average similar to that 6% drop due to mitigations, then they really only have a 12% IPC change, which is a very different claim (although even that may put AMD and Intel's IPC on parity again, meaning the fastest CPU wins, which currently is Intel).

When you look at the SPEC 2016 and 2017, then a custom SPECInt 2006, and it seems like some gaming is afoot. From the footnotes, Intel specifically claims it may or may not have included all software mitigations at the time of testing. That means, specifically, some of the mitigations ARE in place. Now, there is a question on when the tests were conducted. The public release of patches for Zombieload, etc., didn't arrive until May 11th. That means those may or may not be included. We don't know what all microcode was applied to the platform, what was turned on in OS mitigations, etc. Hell, we don't even know for certain whether it is comparing HT on and off, although I will assume for Intel's sake it was on. We don't know if the Portsmash fix was incorporated either.

So, depending on the state of the vulnerability mitigations, much of that IPC increase truly could be mitigations and clawing back lost performance. We simply do not know.

What we can say for sure is the 40% raw performance number comes from comparing a skylake-u sample to an icelake-U sample, which comes with more advanced boosting and higher clock speeds which is mixed with the IPC gains and with the software mitigation sandbags. Plus, that was ONLY showed with SPECInt 2006, which is curious why they wouldn't do a multi-benchmark analysis for overall performance, instead relying on an old integer benchmark with new CPU optimizations in compiling.

As for the claims on the laptop iGPU, those are NOT accurate. They used 3733, then they stuck AMD's APU with DDR4 2400, which means they are using 55% less bandwidth for the comparison.

They also OVERCLOCKED the Ice Lake-U CPU to 25W, meaning the performance shown is not what you get, then tried to justify the 2400MHz as what you will commercially buy on the AMD machine, meaning that they are using dual standards for the testing of the device (not surprising with Ryan Shrout). If you wanted to make the fight about the performance of the actual tested component, you have to standardize the other variables. Without that, you are not properly comparing ANYTHING.

But, with that said, let's dive into changes they made to the uArch that could actually achieve a sizeable IPC uplift, which would make the claim real.

1559087090017.png
1559087112652.png
https://wccftech.com/intel-10nm-ice-lake-18-pc-ipc/

Intel has increased the L1 data cache by 50%, they doubled the L2 cache to 512KB, increased microop cache, the Out of Order Execution window has been expanded, etc. Now, for the cache, this can lower cache and memory latencies further, which can boost IPC. The OoO window can allow for the data to not be retired as quickly with the speculative execution mitigations in place, which would save on the stale data problem, helping to bring that back up in performance.

They also increased the width of allocation, expanded the execution ports by 25%, doubled the L1 store bandwidth, added an extra AGU (33% more), and doubled the store data. If you look, much of this can help with the mitigations applied for security vulnerabilities when combined with above features.

But, putting security aside for a moment, all of the above can improve the IPC performance of a chip on certain workloads. It isn't a question of if they improved the IPC, rather it is a question of by how much they increased it.

Intel never sat still while not using the uArch Sunny Cove. They have been refining it. So I'm waiting for better third party reviews to find out the actual amount.
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
As for the claims on the laptop iGPU, those are NOT accurate. They used 3733, then they stuck AMD's APU with DDR4 2400, which means they are using 55% less bandwidth for the comparison.
IIRC, you can't overclock Ryzen mobile's memory controller beyond stock, which on the 3700U is 2400. Definately unfair but the settings were pretty low, I don't know the extra bandwidth would make that much of a difference. Either way I do agree that I would expect Vega Mobile to be decently faster than the 64 EU models.
 
Apr 1, 2015
193
12
116
IIRC, you can't overclock Ryzen mobile's memory controller beyond stock, which on the 3700U is 2400. Definately unfair but the settings were pretty low, I don't know the extra bandwidth would make that much of a difference. Either way I do agree that I would expect Vega Mobile to be decently faster than the 64 EU models.
Yet you think overclocking the CPU on the Intel by 66% of TDP is *fair*?

My point is, even though you may not be able to go higher on the AMD, you can go lower on the memory bandwidth for the Intel to make it apples to apples on bandwidth.

Otherwise, let's show the stock performance of both, Intel's chip with the faster memory but 15W TDP versus the AMD APU "U" series at stock with 25W TDP (configurable up to 45W TDP).

It's either you standardize the variables for a test or you don't. It is that simple. Anything other than that is showing something else other than a comparison.

Edit: Had to change the percentage because 10W compared to the 15W on the intel CPU is 66%, not 40%, which is 10W out of 25W
 

jpiniero

Diamond Member
Oct 1, 2010
6,411
260
126
Yet you think overclocking the CPU on the Intel by 40% of TDP is *fair*?
Intel typically allows 15 W U parts to be cTDP'd up to 25 if the OEM chooses, so yeah. I imagine that will be the case with Icelake as well.
 
Apr 1, 2015
193
12
116
Intel typically allows 15 W U parts to be cTDP'd up to 25 if the OEM chooses, so yeah. I imagine that will be the case with Icelake as well.
But at that point, OEMs would also look to the 7nm ARM chips that can do almost the same performance as Intel U series at 7W.

Granted, this was to show the iGP performance, but when you restrict memory bandwidth on most GPUs, the performance drops quite a bit.
 
Oct 14, 2003
6,084
218
126
The 1065 is only running at 3.48GHz. The Integer portion is a fair bit lower than the 8665U, however, the 8665U runs at 4.65GHz.

16% perf/clock advantage.

They also OVERCLOCKED the Ice Lake-U CPU to 25W, meaning the performance shown is not what you get, then tried to justify the 2400MHz as what you will commercially buy on the AMD machine,
What? They said its at 25W because the AMD systems use 25W. 25W is actually a quite common practice for "15W" chips.

DDR4-2400 is a common max configuration on Ryzen laptops. Ryzen doesn't support LPDDR4/x. The Ryzen 2nd gen mobile released in March. It'll be competing with Icelake U for a significant majority of its life.

The Vega iGPU on 2nd Gen Ryzen mobile will be faster, but not by much. Maybe 5-10%.
 
Last edited:

beginner99

Diamond Member
Jun 2, 2009
4,109
176
126
This geekbench doesn't inspire confidence :(

Intel Core i7-1065G7 vs Intel Core i5-8265U

Both has the same turbo clocks.
Yeah. I would assume if the process just clocks poorly, then it wouldn't matter at these clocks and yet icelake one has the lower base clock (or is that because of AVX-512 and a non-issue unless you run AVX-512?)
 


ASK THE COMMUNITY

TRENDING THREADS