Discussion Intel current and future Lakes & Rapids thread

[Gideon]

This geekbench doesn't inspire confidence

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

Both has the same turbo clocks.

Great comparison! Yeah, the 10% clock-speed drop eats away a lot of the benefits (let's not fo.rget GeekBench does have AVX-512, so Ice-Lake should benefit more than in benches without that). Ice lake will still be a noticeable update for the decent IGP, possibly for power-draw, but unless they get the clocks up, it doesn't seem that much faster in CPU-limited workloads.

 

[DrMrLordX]

Does GB4 try to run AVX512 and non-AVX512 code concurrently? Unless Intel has done something to mitigate the massive clockspeed drop on all cores when running AVX512 workloads, trying to do just that could actually lower scores in some ways.

 

[Nothingness]

Yea but that Dell in the Geekbench result does not.

I know that for ARM platforms Geekbench does not report correct clocks. Is it better on Windows x86?

 

[krumme]

10nm was and still is an unmitigated disaster.
It needs another spin. At least. 7nm probably arrives before this is fixed.

 

[krumme]

https://www.anandtech.com/show/14436/intel-10th-gen-10nm-ice-lake-cpus

 

[ajc9988]

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.



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%.

But, the CPU they used was specifically the 15W variant, which does make it an overclock, even if it is allowed under spec and common. But, since it is within spec for allowed function, it is somewhat fair.

I argue, though, that cTDP being allowed does not remove it being an overclock. I can set my cTDP on my 1950X from 180W to 250W. Would you agree that 250W on my 1950X is an overclock? That change alone would allow my system to boost higher for longer due to a higher ceiling.

Now, I'm not meaning this combatively, what is your conceptualization of overclocking? The reason I ask this is due to features on both AMD and Intel CPUs, the concept of what overclocking is has become murky. For example, do you consider multi-core enhancement where the MB mfr set all cores to the single core boost an overclock? Even if the MB mfr sets that as the default? Or do you consider AMD's Precision Boost Overdrive or XFR overclocking, even though those are built into the CPU? Do you consider exceeding the TDP and power ratings of a chip overclocking? Is it overclocking if you have a cTDP allowing for it to be changed?

It's an interesting question with the changes of features within chips these days and I'd like to get your opinion on it. Some have drawn the line at the chips behavior versus settings in the bios, or by the MB mfr or system integrator, that manipulate the basic functions of the CPU (in other words, where PBO and XFR are a function of the chip, much like Intel's stock boost settings, but manipulating the boost on Intel's chips manually, like setting Boost 3.0, is overclocking). It's kind of a meta question here.

 

[Ajay]

But, the CPU they used was specifically the 15W variant, which does make it an overclock, even if it is allowed under spec and common. But, since it is within spec for allowed function, it is somewhat fair.

Really, it doesn't matter. What's an overclock, or not, is set by Intel. In this case, Intel's own specs say - no, it isn't an overclock. Done. Not point in making a case against it.

 

[TheGiant]

First GB4 of I7 1065 G7 in XPS 2-1

https://browser.geekbench.com/v4/cpu/13303489

And Gen11

https://browser.geekbench.com/v4/compute/4094642

that IPC looks nice..
even the clock if can be sustained I can manage that...
I really want to see how often that fan starts and how long can the icelake chip hold its 3,5GHz with what power a temps
so far that icelake looks like major technical upgrade

with 4GHz, that icelake should have geekbench single core close to 9900K@5Ghz (around 6K result)
not bad at all, but we don't eat from geekbench
any info on the reviews? or when can we buy that dell xps ?

 

[ajc9988]

Really, it doesn't matter. What's an overclock, or not, is set by Intel. In this case, Intel's own specs say - no, it isn't an overclock. Done. Not point in making a case against it.

That is a cop out.

Awhile ago, Gamers Nexus did a video on Intel not following their own spec in regards to boost speeds, etc. Intel has put out conflicting specs so that you can comply with one and fail the other.

My overall point, then, is that what amounts to an overclock versus stock performance needs examined so that clarity is given, overall, on what is occurring.

It isn't up to the company to say if it is an overclock, other than for warranty purposes.

 

[Ajay]

That is a cop out.

Awhile ago, Gamers Nexus did a video on Intel not following their own spec in regards to boost speeds, etc. Intel has put out conflicting specs so that you can comply with one and fail the other.

My overall point, then, is that what amounts to an overclock versus stock performance needs examined so that clarity is given, overall, on what is occurring.

It isn't up to the company to say if it is an overclock, other than for warranty purposes.

Okay. You win. I just care that a device performs as I expect after reviewing benchmarks, battery life, cooling/power requirements, etc.

 

[ajc9988]

Okay. You win. I just care that a device performs as I expect after reviewing benchmarks, battery life, cooling/power requirements, etc.

But that is my very point.

With cTDP changes among vendors, you could have a wide variety of performance with the same exact hardware, which can confuse consumers.

That was the point of the Gamers Nexus videos, where they addressed vendors having multi-core enhancement on by default, ignoring spec on boost duration, changing and cheating the BCLK bus timings, etc.

So, this isn't about "winning," it is about trying to figure out what the performance is for a given product.

When you looked at motherboard reviews, or even CPU reviews, you would see the variance in performance between reviewers, outside of the expected variance of the product. This often then caused reviewers to contact other reviewers to try to figure out why this variance was there, that way they could control for what was happening. That makes life hard for them, but also makes life hard for consumers when trying to pick out hardware.

When I ask the question as to what counts as an overclock, what I am asking for is to figure out what the performance is for the product in a controlled situation.

With cTDP and with boosting not based on duration, or that duration being modified from spec, instead relying on the ability for cooling, etc., there is a possibility for more variance and requires more attention on the part of the consumer, rather than just seeing the hardware ID, to understand what they are getting.

Does that make more sense why these questions are important?

 

[jpiniero]

https://newsroom.intel.com/wp-content/uploads/sites/11/2019/05/10th-Gen-Intel-Core-Product-Brief.pdf

PDF confirms that the PCIe 3.0 lanes are on the PCH, and appears that the connection between the two is still 3.0 x 4. So any dGPUs would only be 3.0 x 4 max. Oh and Y for some reason doesn't support it despite having a high TDP for Y.

Also notes that using Thunderbolt might require increasing the TDP to be able to maintain base clock in a long term load.

 

[IntelUser2000]

This geekbench doesn't inspire confidence

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

Both has the same turbo clocks.

Actually, it doesn't. The 8265U system is running the benchmarks at 3.85GHz. The 1065G7 system runs at 3.45GHz. That's still a 12-13% advantage. Again, I'm only looking at the Single-threaded Integer portion, because that's what defines uarch and others are easier to improve.

I know that for ARM platforms Geekbench does not report correct clocks. Is it better on Windows x86?

How off from reality is the ARM version? I've been comparing using the .gb4 at the end and it seems pretty decent. You could always run it on your PC while you have HWInfo open for example for double checking.

PDF confirms that the PCIe 3.0 lanes are on the PCH, and appears that the connection between the two is still 3.0 x 4. So any dGPUs would only be 3.0 x 4 max. Oh and Y for some reason doesn't support it despite having a high TDP for Y.

The current Whiskey Lake also only supports up to x4 lanes.

The TDP increase for the Y is said to be due to the doubling of the core, but it seems that they are just planning to shift the stack. There were no small tablets with Y, and most systems were set at 7W. Some, like the Dell were set at 9W. I guess they could have put x4 support for Y too, but something's off with that datasheet so I want to see the ARK page.

Lakefield seems to be filling the low power end, and its actually compact enough to put in 10-inch or smaller devices.

 

[Gideon]

Actually, it doesn't. The 8265U system is running the benchmarks at 3.85GHz.

That's the whole point. The listed Single Threaded turbo is the same, yet 8265U manages to clock higher (and even in multi-threaded workloads, where process shrink should provide the biggest uplift in sustained clocks). Yes, Ice-Lake looks very solid IPC-wise, but the clocks on 10nm are definitely holding it back, for now.

EDIT:
And the biggest outliers are AES (upgraded fixed-function hardware that won't benefit most workloads) and SFFT, which uses AVX-512. Curiously GEMM also uses AVX-512, yet is slower.

 

[mikk]

https://newsroom.intel.com/wp-content/uploads/sites/11/2019/05/10th-Gen-Intel-Core-Product-Brief.pdf

PDF confirms that the PCIe 3.0 lanes are on the PCH, and appears that the connection between the two is still 3.0 x 4. So any dGPUs would only be 3.0 x 4 max. Oh and Y for some reason doesn't support it despite having a high TDP for Y.

Also notes that using Thunderbolt might require increasing the TDP to be able to maintain base clock in a long term load.

Tigerlake will get PCIe4 support anyway.

The TDP increase for the Y is said to be due to the doubling of the core, but it seems that they are just planning to shift the stack. There were no small tablets with Y, and most systems were set at 7W. Some, like the Dell were set at 9W. I guess they could have put x4 support for Y too, but something's off with that datasheet so I want to see the ARK page.

Also keep in mind there is more stuff integrated shifted from the board into the CPU package like Thunderbolt or FIVR, from the pdf:

Board Savings3 due to IP Integration FIVR (both CPU and PCH ), Type-C sub system , HDMI2.0/HDCP2.2, Wi-Fi* (CNVi MAC) etc.

An 15W Icelake device should be lower in overall power consumption than a 15W Skylake based notebook.

 

[mikk]

That's the whole point. The listed Single Threaded turbo is the same, yet 8265U manages to clock higher (and even in multi-threaded workloads, where process shrink should provide the biggest uplift in sustained clocks). Yes, Ice-Lake looks very solid IPC-wise, but the clocks on 10nm are definitely holding it back, for now.

EDIT:
And the biggest outliers are AES (upgraded fixed-function hardware that won't benefit most workloads) and SFFT, which uses AVX-512. Curiously GEMM also uses AVX-512, yet is slower.

1C Turbo didn't work in this device, it only clocked with about 3.5 Ghz which is the 4C Turbo, this ICL SKU can go up to 3.9 Ghz, so with a working 1C Turbo it should easily outpace the 8265U ST score in Geekbench.

 

[Abwx]

EDIT:
And the biggest outliers are AES (upgraded fixed-function hardware that won't benefit most workloads) and SFFT, which uses AVX-512. Curiously GEMM also uses AVX-512, yet is slower.

Nice catch, i was about to point it, in the ST test those two subscore would account for 3% IPC improvement (actually rather throughput) if all other scores were even at same frequency.

Neverless there seems to be a solid 9-10% in the LZMA subscore, wich is a good indicator for Integer IPC improvement unless they rely on new instructions.

 

[JoeRambo]

The elefant in the room is 105ns memory latency. Full 50% worse than i5-8265U. So maybe uncore was running fixed <800mhz all the time, maybe memory was CL33, maybe IMC is not being setup properly. Latency has big impact on quite of few of those scores.

 

[Abwx]

Hence why ST scores are more relevant, FI LZMA is greatly impacted by latency but in ST there s enough room to not alterate the IPC, the 9-10% i pointed are in the ST subscore assuming frequencies are accurate.

 

[JoeRambo]

Hence why ST scores are more relevant, FI LZMA is greatly impacted by latency but in ST there s enough room to not alterate the IPC, the 9-10% i pointed are in the ST subscore assuming frequencies are accurate.

I don't agree with ST scores being more relevant, every memory access taking 50% more is going to hurt every test that is not trivial and optimized to fit in caches. 105ns is not good place for memory latency to be.

 

[Abwx]

I don't agree with ST scores being more relevant, every memory access taking 50% more is going to hurt every test that is not trivial and optimized to fit in caches. 105ns is not good place for memory latency to be.

The new uarch has significantly more cache, wich help hide some latency, besides you ll notice that in LZMA the ST to MT scaling is the same within 1%, wich is not only indicative of equivalent ratio for ST frequency/MT freq but also that both those low clocked CPUs do not have enough throughput, for this kind of load, to saturate the bandwith of relatively slow DDR4.

 

[JoeRambo]

The new uarch has significantly more cache, wich help hide some latency, besides you ll notice that in LZMA the ST to MT scaling is the same within 1%, wich is not only indicative of equivalent ratio for ST frequency/MT freq but also that both those low clocked CPUs do not have enough throughput, for this kind of load, to saturate the bandwith of relatively slow DDR4.

I expect this CPU to shine in tests like HTML5 Parse / HTML5 DOM / LLVM and it does not. And these do involve memory accesses that are hard to cache due to random nature of pointer chasing. And what about poster child of memory accesses - SGEMM. Single thread improves, but multi thread fails to scale? Common, i know You are smarter than this.

 

[tamz_msc]

For what it's worth Cannon Lake had worse memory latency than Skylake. Could it be the same for Ice Lake?

 

[JoeRambo]

For what it's worth Cannon Lake had worse memory latency than Skylake. Could it be the same for Ice Lake?

CNL had some horribad memory to start with, wasn't it running some laptop 2133 stuff? Still 105ns is horrible when previuos gen had 70?

EDIT: i looked up the Anandtech CNL article and it had same 50% latency increase over last gen. So yeah, probably 10nm is biting Intel hard here.

 

[Abwx]

I expect this CPU to shine in tests like HTML5 Parse / HTML5 DOM / LLVM and it does not. And these do involve memory accesses that are hard to cache due to random nature of pointer chasing. And what about poster child of memory accesses - SGEMM. Single thread improves, but multi thread fails to scale? Common, i know You are smarter than this.

The tests you re pointing are not that arithmetic computing heavy but instead are making big usage of instructions related to data manipulation, computed flows are bigger and ther will be more RAM accesses required.

For the LZMA test there is effective heavy arithmetic computation that will slow down the flow of data required from and sent back to the RAM, indeed you can see that the ST computed data flow is 7MB/s for LZMA and about 22MB/s for HTML5 Parse, in MT the computed flows are 30 and 76MB/s respectively...