[Ashraf] 10nm "Lakefield" SoC with Intel big + little cores

[IntelUser2000]

This is CPU TDP versus Soc package TDP isn't it? Aside from this, an 8 core won't lose in Cinebench MT against 4 cores unless it has a really big IPC and clock deficit which is not the case.

The 4 core Tigerlake Y should get pretty close there too.

You can see the HWInfo screenshot. It's CPU package power.

I can say the Ryzen 4000 Mobile CPU is really efficient for CPU workloads. The fact that the 8 core can do that without throttling excessively is amazing since at 10W uncore power starts dominating, and it also means the frequency doesn't drop drastically as voltage drops. Yet here we are. Ryzen 4000 isn't even meant for 10W operations.

Page 129 - Speculation: Ryzen 4000 series/Zen 3

Page 129 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

320 in Cinebench R15 MT with just a case fan a meter away and without a heatsink(nevermind a fan). That's a 6C 6T BTW.



I know process is a big factor there because the TSMC 7nm is quite efficient. Intel can tout density all day long but based on Icelake the power efficiency is seriously behind.

 

[IntelUser2000]

400 Cinebench might sound amazing, but for most people it'll be something that performs worse than Whiskeylake, a chip that's 2 years old on a 5 year old CPU and GPU uarch.

At $699 for a device perhaps that's a good product. For a $999 one I think it should get 500 points as thermal throttling will get sustained performance lower.

The problem is not even the Tremont portion is doing better than Goldmont Plus. Why is it not 30% faster and getting 330 points? Why is the ST Tremont only 10% better? It should be getting 105-110 points not 88.

And they want $2500 Base configuration price for the X1 Fold!

 

[mikk]

The problem is not even the Tremont portion is doing better than Goldmont Plus. Why is it not 30% faster and getting 330 points? Why is the ST Tremont only 10% better? It should be getting 105-110 points not 88.

It should get 100-105 points with the maximum Tremont boost in Lakefield. Tremont runs only at 2.4 Ghz in Cinebench R15 ST while it could theoretically boost up to 2.8 Ghz, this is a theoretical 16.7% improvement over 88 points.

Does the Lakefield TDP include the RAM? We need power consumption measurements, even though the most likely scenario is that 10nm is just poor, both Tremont and Sunny Cove are not looking good.

 

[IntelUser2000]

Does the Lakefield TDP include the RAM? We need power consumption measurements, even though the most likely scenario is that 10nm is just poor, both Tremont and Sunny Cove are not looking good.

If TDP is datasheet TDP yes it should include RAM, since it needs to cool it. The N5000 runs at 6W though so its not much lower.

Yes to the power consumption numbers, but only through battery drain numbers in HWInfo. AC power adapter measurements are bad.

N5000 doesn't need to clock to 2.7GHz to get its 77-80 point score.

 

[mikk]

If TDP is datasheet TDP yes it should include RAM, since it needs to cool it. The N5000 runs at 6W though so its not much lower.

Don't forget that the PL2 is much higher (15W vs 9.5W).

N5000 doesn't need to clock to 2.7GHz to get its 77-80 point score.

It needs about 2.6 Ghz. Have a look to the Goldmont Plus based N4100, it has a maximum boost of 2.4 Ghz and the best devices can do roughly 70 points in Cinebench 15 ST, it means Tremont has a ~25% IPC advantage there.

 

[IntelUser2000]

It needs about 2.6 Ghz. Have a look to the Goldmont Plus based N4100, it has a maximum boost of 2.4 Ghz and the best devices can do roughly 70 points in Cinebench 15 ST, it means Tremont has a ~25% IPC advantage there.

25% is not stellar for a core that gets updated every 2 years(Hopefully Gracemont is end 2021, not mid to end 2022). At least with N5000 it was 30% faster per clock and it clocked higher. The end product was 50% faster in Cinebench.

If you assume 2.6GHz for N5000 and 80, you only get 19%. Problem is not perf/clock but the fact that it can't clock enough to save itself. Kinda like how Sunny Cove clocked lower but a faster uarch.

 

[jpiniero]

Don't forget that the PL2 is much higher (15W vs 9.5W).

Almost all of them are fanless, and all of them don't throttle at least in R15. Can't be much more than 6. That may be part of the problem with the Galaxy S Lakefield, 9 is really pushing it fanless.

 

[mikk]

25% is not stellar for a core that gets updated every 2 years(Hopefully Gracemont is end 2021, not mid to end 2022). At least with N5000 it was 30% faster per clock and it clocked higher. The end product was 50% faster in Cinebench.

If you assume 2.6GHz for N5000 and 80, you only get 19%. Problem is not perf/clock but the fact that it can't clock enough to save itself. Kinda like how Sunny Cove clocked lower but a faster uarch.

80 points is the very best N5000 score at Notebookcheck out of 11 devices. i would expect this is a run with its maximum boost of 2.7 Ghz. The Atom IPC improvements every generation are fine, the IPC is more than twice as good as Airmont in Cinebench R15 which came out in 2015. Sure, the starting base is much lower but the Atom progress is much more impressive than what Core uarch delivered since 2015. The big issue is the stagnation in clock speeds since Airmont. While Core based models went higher and higher every generation (until Icelake), Atom stagnated in the 2.x Ghz range. This might change in the future hopefully. Intels architecture roadmap back in December 2018 highlighted ST perf +frequency for Gracemont, same for the "Next Mont".

 

[DrMrLordX]

You are comparing an overclocked chip vs an overclocked chip.

Yup. That was the big comparison of the day, and it also highlighted clockspeed regressions Intel suffered moving from Sandy to Haswell. It is an enthusiast's perspective, but this forum is an enthusiast forum, so . . .

 

[Thala]

It works as Intel has stated it should:

https://www.extremetech.com/wp-content/uploads/2020/06/Tremont-Performance-Curves-1.jpg

SunnyCove is meant for burst type ST-loads only. If they give power budget to SunnyCove in MT-loads it will result lower MT performance as seen in published graphs.

SunnyCove performance/watt is extremely poor, don't compare it to Zen2 as SunnyCove loses in efficiency to even Zen1 Picasso. Skylake has much better performance/watt at low clock speeds than SunnyCove.

Precisely. The graph pretty much shows, that if you are at the power limit, putting more power in SunnyCove and reducing power to Tremonts will actually reduce performance.
This also means that the Windows scheduler makes the right decision.

ps. It actually depends where the actual power limit is relative to the graph - with the graph alone it is not decidable.

 

[IntelUser2000]

While Core based models went higher and higher every generation (until Icelake), Atom stagnated in the 2.x Ghz range.

That's a GOOD thing. It prevents thermal and power runway that comes with running at 4+ GHz frequencies.

ARM way is the right way to go, where frequency stays the same and the performance gain comes entirely using architecture. Then it won't have to clock at 5GHz. We know how bad it is looking at 10900K versus Zen 2 comparison.

 

[Roland00Address]

I want to see a comparison of the Core i5-10210Y to Lakefield (quad core / eight threads, 7 watt tdp but also with a ctdp up of 9w and a ctdp down of 5.5w.)

Of course these benchmarks are talking load, or max load performance and part of the goal of Lakefield is idle and standby power, but I am just curious. Come on Anandtech get the hardware directly from intel and break out those volt-meters to satisfy my curiosity 😂

 

[IntelUser2000]

From the reply in the Notebookcheck test:

I just received my Samsung Galaxy Book S today in Canada. Got all the Windows updates and quickly ran a few benchmarks. Geekbench 5 is about the same as that shown here, but I got better results with Cinebench R20.

Single Core - 237 pts
Multi Core - 553 pts

NBC result

Single Core - 182
Multi Core - 479

30% better in ST and 15% better in MT.

@Roland00Address 10210Y gets 132 in Cinebench R15 ST and 403 in MT.

 

[Roland00Address]

@Roland00Address 10210Y gets 132 in Cinebench R15 ST and 403 in MT.

I originally wrote a longer post about why I wanted to see the 10210Y but I truncated it for it felt rambly.

But to expand on my thoughts ... I want to see it at various tdps, I am curious due to seeing the benchmarks for a OneGx1 (aka think an asus 701 but even smaller, mixed with a nintendo switch, and going for a $1k.)

OneGx1 benchmarks (7 inch mini laptop with Intel Core i5-10210Y) - Liliputing

OneGx1 benchmarks (7 inch mini laptop with Intel Core i5-10210Y)

liliputing.com

[None of these chips are Lakefield, just small laptops.]


I want to more about what is the "real" tdp of the mute mode vs non mute mode. Supposedly mute mode is 2 to 5w of power consumption but I want to seem a more detailed information and not just handwaving saying it is 2w to 5w.

Same silicon just different performance modes 112 vs 132 (ST) and 248 vs 403 (MT.) Also the Dell XPS 9300 with the core i7, if I recall that is a cTDP up at 25w. Also that is a 10nm part and not 14nm while the 10210Y is 14.

WOW. I just realized the 10210Y is a 14nm part not a 10nm. 😬

 

[IntelUser2000]

I want to more about what is the "real" tdp of the mute mode vs non mute mode. Supposedly mute mode is 2 to 5w of power consumption but I want to seem a more detailed information and not just handwaving saying it is 2w to 5w.

It's probably 5W then. At 2W the uncore would start dominating and it would have very little room for the core to work.

You still need to take into account extra power for the RAM on the 10210Y.

 

[coercitiv]

Anandtech has an article up about Lakefield:

Knowing that Lakefield was going to have to take the lowest common denominator from the two core designs, Intel physically removed the very bulky AVX-512 unit from the Sunny Cove core.

With the AVX-512 unit, we would have expected the single Sunny Cove core to be on par on size with all the Tremont cores put together.

Intel has made the 1+4 design to act as a 0+4 design that sometimes has access to a higher performance mode. Whereas smartphone chips are designed for all eight cores to power on for sustained periods, Lakefield is built only for 0+4 sustained workloads. And that might ultimately be its downfall. This leads onto a deep discussion about Lakefield’s performance, and what we should expect from it.

Benchmarks vs. Intel Amber Lake i7-8500Y in 5W mode

• +12% single threaded performance, measured by SPEC2006 (3.0 GHz vs 4.2 GHz)
• +70% graphics performance, 3DMark11 comparing HD615 (24 EUs, Gen 9.5 at 1.05 GHz, 2x4 GB LPDDR3-1866) vs HD (64 EUs, Gen11 at 500 MHz, 2x4 GB LPDDR4X-4267)
• +24% power efficiency, score per Watt on WebXPRT 3
• +100% AI workloads on graphics, ResNet50 batch 128 on OpenVINO, comparing

Only 12% ST performance advantage with a node advantage that should shine in low power scenarios.

 

[Zucker2k]

Isn't that

Anandtech has an article up about Lakefield:



View attachment 24731



Benchmarks vs. Intel Amber Lake i7-8500Y in 5W mode

Only 12% ST performance advantage with a node advantage that should shine in low power scenarios.

Isn't that at a 40% frequency disadvantage though? If so, an unrestricted willow cove core should pack a punch.

 

[coercitiv]

Isn't that at a 40% frequency disadvantage though? If so, an unrestricted willow cove core should pack a punch.

It can't be, with ~18% IPC advantage and ~30% clock disadvantage the SNC core would actually be 15% slower than Skylake. Amber Lake operates in 5W TDP mode so I assume it couldn't sustain max boost for the duration of whatever workload Intel put them trough.

As for Willow Cove, sure, we're bound to start seeing performance numbers soon enough.

 

[Zucker2k]

Amber Lake operates in 5W TDP mode so I assume it couldn't sustain max boost for the duration of whatever workload Intel put them trough.

I thought so too, but what's that 3GHz vs 4.2GHz in parenthesis? Surely, the sunny cove core is the one running at 3GHz?

 

[coercitiv]

I thought so too, but what's that 3GHz vs 4.2GHz in parenthesis? Surely, the sunny cove core is the one running at 3GHz?

3Ghz vs. 4.2Ghz is max boost clocks SNC vs. Amber Lake. Unfortunately it's hard to analyze this further without access to those Intel slides Ian Cutress was talking about.

 

[Roland00Address]

It's probably 5W then. At 2W the uncore would start dominating and it would have very little room for the core to work.

You still need to take into account extra power for the RAM on the 10210Y.

*Nods*

That is why I want to see Anandtech or someone other technical hardware site do things like breaking out the voltmeters and so on and give me more detailed information. I want to see this for Lakefield and other **similar** silicon (not specifically this silicon) is likely some kind of future and even if this first version is bad I want to see how it compares to silicon from another **parallel** tradition that is coming out in 2020 aka the same time more or less.

I want a deep dive 😁 , I crave hardware news and other stuff since the world is *gestures* for the last few months.

[Laughs for I see the anandtech article and will not be able to read it in a deep way for a few hours.]

 

[IntelUser2000]

The Core chips can't operate above 3GHz in a 7W power envelope even in ST which is why Amberlake is significantly behind U processors. Not even TDP upped to 9W it can do that. Same with Icelake-Y, as its barely faster than Amberlake. 15W+ is needed for a single Skylake core to operate at 4.2GHz.

Unfortunately it's hard to analyze this further without access to those Intel slides Ian Cutress was talking about.

The slide that compares to 8500Y has been public for nearly 2 weeks now. Just go to Intel's site.

12% gain over 8500Y would be fairly decent if it can manage it elsewhere. That would make it 140-145 Cinebench R15 and 330 R20 for 1T. That would slightly beat the Icelake-Y chips.


The article is very optimistic about the product. At 180 R20 score, its quickly going to be on a firesale. Even if it gets 14 hours of battery life(15 hours for ARM version) its a $999 device. It'll be a very niche sale where a handful of people that would have bought the ARM version goes for it instead. That's a guaranteed losing proposition.

At 230 R20, its better, but the audience still won't be large.

 

[coercitiv]

The slide that compares to 8500Y has been public for nearly 2 weeks now. Just go to Intel's site.

I have yet to find that slide.

 

[IntelUser2000]

I have yet to find that slide.

When it was announced they had a press release. You just go to their newsroom.

Newsroom Home

Intel Newsroom home page

newsroom.intel.com

Intel.com-->Newsroom-->Scroll down and click View all News-->Click on View more-->June 10, 2020 article named
Intel Hybrid Processors: Uncompromised PC Experiences for Innovative Form Factors Like Foldables, Dual Screens

 

[coercitiv]

When it was announced they had a press release. You just go to their newsroom.

Yeah, I kept looking through their June 19th Press Kit.

Something good come out of that too, as I hadn't seen this either:


This reminded me about one "old" article published by @Andrei. on real world CPU utilization in smartphones. Here's the Chrome browser test results and article quotes:

Spoiler: Big CLuster Run-Queue Depth

Spoiler: Total Run-Queue Depth

The big cores also see a fair amount of load. Similarly to the S-Browser we have 1 very large thread that puts a consistent load on 1 CPU. But curiously enough we also see some significant activity on up to 2 other big cores. Again, in terms of burst loads we see up to 3 big CPUs being used concurrently.
The total run-queue depths for the system looks very different for Chrome. We see a consistent use of 4-5 cores and a large burst of up to 8 threads. This is a very surprising finding and impact on the way we perceive the core count usage of Chrome.