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

[vigilant007]

As i already mentioned elsewhere, Windows on big.LITTLE ARM cores like on the Surface Pro X works really well. You see the small core utilized either on very low workloads like background processes or when the application requests 8 threads with full utilization, then all 8 cores are loaded 100%.
Of course the application can still make stupid decisions, like distributing the work equally to all 8 threads - but this is rarely the case in application where MT performance counts. For example Blender distributes one tile at a time to each core. Once a core is finished with a tile it gets the next tile assigned. You can literally see which tiles are assigned to the small cores

ps. I do still have the issue, that under Linux the CPU does not increase the clock when all cores are requested with full utilization on my Surface Pro X. As workaround i start a program on the Windows side which pulls the clock-speed up while running the Linux app under virtualization. Even then i am beating Lakefield in Blender by miles even with a relatively ancient Cortex A76 system - all while using less power from battery.

Thank! That is perfect!

Out of curiosity, how do you like the Surface X for day to day use? I am tempted to pick one up this week.


Sent from my iPad using Tapatalk Pro

 

[DrMrLordX]

I don't think it's fair to slag Tremont as being inferior. Goldmont+ -> Tremont seems like a substantial jump in performance at isoclocks. The problem is that Tremont got caught up in the node mess. We still haven't seen proper replacements for Goldmont/Goldmont+ consumer devices featuring Tremont, nor have we seen any movement in Intel's automotive lineup.

 

[coercitiv]

hardwareLUXX is taking a closer look at Lakefield, inter-core latencies are interesting. Core 4 is Sunny Cove.

 

[Thala]

Thank! That is perfect!

Out of curiosity, how do you like the Surface X for day to day use? I am tempted to pick one up this week.

I'd say its the perfect device for me. Fast and snappy combined with gorgeous screen and form factor plus excellent battery duration. Most software i am running is native. In some cases you might need to looks for alternatives. For example instead of Photoshop why not running GIMP under native WSL2? Be aware that most 32bit programs do work under emulation, some just do not and crash. GPU performance is excellent. For example Diablo 3 is running at around 30 fps at 2880x1920 resolution - thats 5.5 mio pixel. Other games like Dust: An Elysian Tail even achieving stable 60fps under same resolution with all gfx options on. Ori and the Blind Forrest - stable 60fps lower resolution though.

Some other limitation like the missing OpenGL are looking to get better. Microsoft announced lately, that they are working with Collabora to work on an OpenGL driver based on the Mesa project which has a D3D12 backend. Now since the effort is open source i cloned the sources in addition to the quake 3 sources compiled everything and here is the result:



Accelerated OpenGL on my Surface Pro X

ps: sorry for off-topic - but it is somehow related to lakefield.

 

[vigilant007]

I'd say its the perfect device for me. Fast and snappy combined with gorgeous screen and form factor plus excellent battery duration. Most software i am running is native. In some cases you might need to looks for alternatives. For example instead of Photoshop why not running GIMP under native WSL2? Be aware that most 32bit programs do work under emulation, some just do not and crash. GPU performance is excellent. For example Diablo 3 is running at around 30 fps at 2880x1920 resolution - thats 5.5 mio pixel. Other games like Dust: An Elysian Tail even achieving stable 60fps under same resolution with all gfx options on. Ori and the Blind Forrest - stable 60fps lower resolution though.

Some other limitation like the missing OpenGL are looking to get better. Microsoft announced lately, that they are working with Collabora to work on an OpenGL driver based on the Mesa project which has a D3D12 backend. Now since the effort is open source i cloned the sources in addition to the quake 3 sources compiled everything and here is the result:

View attachment 26983

Accelerated OpenGL on my Surface Pro X

ps: sorry for off-topic - but it is somehow related to lakefield.

It’s not going to be a primary device of mine. I mostly use Macs for work, but I occasionally need to do a few things using Microsoft Office that ONLY has the feature on Windows. I’m looking at it as a fun experiment.

I see the move to ARM for every platform incredibly interesting because for a good while Intel just hasn’t been interesting. Qualcomm makes brilliant hardware, and I’m glad Microsoft is using that hardware to hold Intels feet to the fire.

As a tinkerer at heart, I want to see how far I can get with it by being creative. I know I could wait a few months but seeing as it’s not my primary device, I feel compelled to take advantage of the discounting that is currently being offered.

It would also be nice to have a Windows tablet with native LTE connectivity, as when I’m traveling I’m normally working from the side of the road.

 

[coercitiv]

hardwareluXXX published their review and there's also this bit of info from their editor on Twitter.

https://twitter.com/x/status/1288485882111172609

Increasing the TDP from 5W to 7W may not seem like much but it's actually close to 50% more power budget for the cores.

 

[ondma]

hardwareluXXX published their review and there's also this bit of info from their editor on Twitter.

https://twitter.com/x/status/1288485882111172609

Increasing the TDP from 5W to 7W may not seem like much but it's actually close to 50% more power budget for the cores.

Performance per watt is exceptional at least at 5w. If it scales up proportionately to 7w with the same efficiency, it could be competitive.

 

[Thala]

Performance per watt is exceptional at least at 5w. If it scales up proportionately to 7w with the same efficiency, it could be competitive.

Not sure what you consider exceptional, but Galaxy Book S with Cortex 8CX has 50-100% higher performance while using much less battery power - this plays in completely different league. A turd stays a turd - even if you give it 2W more power.

 

[ondma]

Not sure what you consider exceptional, but Galaxy Book S with Cortex 8CX has 50-100% higher performance while using much less battery power - this plays in completely different league.

I was looking at the "Multi-threaded performance per watt" graph at the bottom of page 2 in the hardware luxx article linked a couple of posts above.
link

 

[Hitman928]

I was looking at the "Multi-threaded performance per watt" graph at the bottom of page 2 in the hardware luxx article linked a couple of posts above.
link

There seems to be something wrong with a lot of their calculations for perf/w. If I am understanding their translation correctly, they took the multi-threaded Cinebench and divided by the CPUs PL1 power limit. But if you do that on the 4900HS (for example), you get 4142/35 = 118.3 pts/w. Then take the i5-L16G7 and you get 607/5 = 121.4 pts/w. That's only a 2.6% improvement over the 4900HS.

Am I missing something in the translation here?

 

[Thala]

I was looking at the "Multi-threaded performance per watt" graph at the bottom of page 2 in the hardware luxx article linked a couple of posts above.
link

Saw this graph and have no idea how this was derived. The claim was Cinebench score devided by PL1. Lets do this:

i5-L16G7: 607pk/5W = 121pkt/W
i5-1021OU: 1203pkt/15W = 80.2 pkt/W

Very different than what the graph shows. In addition increasing the TDP from 5W to 7W will not improve efficiency as the frequency does not increase linearly with power.
Besides in the light of the meager battery duration i trust that the device stays within 5W SoC power not the slightest bit.

PS. Hitman928 beat be

 

[mikk]

There seems to be something wrong with a lot of their calculations for perf/w. If I am understanding their translation correctly, they took the multi-threaded Cinebench and divided by the CPUs PL1 power limit. But if you do that on the 4900HS (for example), you get 4142/35 = 118.3 pts/w. Then take the i5-L16G7 and you get 607/5 = 121.4 pts/w. That's only a 2.6% improvement over the 4900HS.

Am I missing something in the translation here?

This is default TDP though, real power usage differs. 4900HS uses more than 35W at this score: https://www.hardwareluxx.de/index.p...muss-sich-abermals-warm-anziehen.html?start=2

First run 65W, second run 50W. At 35W it scores 3300 points.

 

[Thala]

This is default TDP though, real power usage differs. 4900HS uses more than 35W at this score: https://www.hardwareluxx.de/index.p...muss-sich-abermals-warm-anziehen.html?start=2

First run 65W, second run 50W. At 35W it scores 3300 points.

Does not explain the graph - neither the 4900HS numbers nor the Lakefield numbers. According to your link the frequency comes down to 3.1-3.2GHz, which is where the score stabilize at around 3700 points at 35W. Quick calculation shows: 3700/35 = 107pkt/W.
Even better at 25W: 3300pkt/25W = 132pkt/w ....easily beating Lakefield.

 

[mikk]

Does not explain the graph - neither the 4900HS numbers nor the Lakefield numbers. According to your link the frequency comes down to 3.1-3.2GHz, which is where the score stabilize at around 3700 points at 35W. Quick calculation shows: 3700/35 = 107pkt/W.

Yes I know, I'm just saying they are using the default TDP which is flawed, they need real power usages for efficiency tests.

 

[Hitman928]

Yes I know, I'm just saying they are using the default TDP which is flawed, they need real power usages for efficiency tests.

Yeah, there's no way to make that graph work, they need to check their numbers.

 

[Hitman928]

According to this video, setting a 4700u to cTDP down of 10W and disabling turbo gives score of 1670 points in Cinebench R20 which results in 167 pts/w. Seems like the 10 - 15 W range is the sweet spot in terms of efficiency for Renoir. I wonder how much performance they'd have to sacrifice to get it down to the 5 - 7 W range. Will also be interesting to see how it compares to a 4C tigerlake with cTDP down.

 

[HurleyBird]

Yeah, there's no way to make that graph work, they need to check their numbers.

Or at least expand on their methodology. Maybe it's total power consumption of the entire laptop (which itself would introduce a slew of confounds). At a glance, it looks pretty nonsensical.

 

[CluelessOne]

One thing I don't really understand. Why going big little style? Let's say for hypothetical 1 big 1 little core compared to a dual core of the same architecture CPU with 1 core with transistors optimized for let's say 1.6 GHz max frequency and 1 core with transistors that can run up to 4 GHz? Is the power savings gained from in order architecture that much compared to out of order architecture?

 

[DrMrLordX]

One thing I don't really understand. Why going big little style? Let's say for hypothetical 1 big 1 little core compared to a dual core of the same architecture CPU with 1 core with transistors optimized for let's say 1.6 GHz max frequency and 1 core with transistors that can run up to 4 GHz? Is the power savings gained from in order architecture that much compared to out of order architecture?

Being able to power down the big core saves on power versus relying on a wider core that you can bring into a low power state via sophisticated pstates and power gating. Going with the multicore config does seem to consume more die area though.

 

[NTMBK]

According to this video, setting a 4700u to cTDP down of 10W and disabling turbo gives score of 1670 points in Cinebench R20 which results in 167 pts/w. Seems like the 10 - 15 W range is the sweet spot in terms of efficiency for Renoir. I wonder how much performance they'd have to sacrifice to get it down to the 5 - 7 W range. Will also be interesting to see how it compares to a 4C tigerlake with cTDP down.

How much I/O is integrated into Lakefield that is not in Renoir, but rather in a separate southbridge chip? You're not necessarily comparing apples to apples. The southbridge isn't generally very power hungry, but when you get to TDPs this low it starts to matter.

 

[itsmydamnation]

How much I/O is integrated into Lakefield that is not in Renoir, but rather in a separate southbridge chip? You're not necessarily comparing apples to apples. The southbridge isn't generally very power hungry, but when you get to TDPs this low it starts to matter.

None?
The entire right hand side of the soc is IO

 

[mikk]

How much I/O is integrated into Lakefield that is not in Renoir, but rather in a separate southbridge chip? You're not necessarily comparing apples to apples. The southbridge isn't generally very power hungry, but when you get to TDPs this low it starts to matter.

Is it the package or core power usage there? I'm asking because AMDs TDP is a core only TDP, so you can have a TDP of 25W but the Soc can draw 28-29W overall. On the other side Intels TDP includes the entire Soc, it's a package TDP. I would be surprised if it's different on Lakefield.

 

[Hitman928]

Is it the package or core power usage there? I'm asking because AMDs TDP is a core only TDP, so you can have a TDP of 25W but the Soc can draw 28-29W overall. On the other side Intels TDP includes the entire Soc, it's a package TDP. I would be surprised if it's different on Lakefield.

I don't think AMD's TDP is core only, it's SOC TDP. Having a core only TDP doesn't really make sense.

 

[mikk]

I don't think AMD's TDP is core only, it's SOC TDP. Having a core only TDP doesn't really make sense.

It's a core TDP, you can see it in the tests. The CPU power usage is exactly 15W or 25W under full load, package power is like 3-4W more than this.

 

[Hitman928]

It's a core TDP, you can see it in the tests. The CPU power usage is exactly 15W or 25W under full load, package power is like 3-4W more than this.

According to Techspot (Hardware Unboxed), the 4900HS (35 W TDP) sits at exactly 35 W package power after turbo duration is exceeded.

If TDP was core only, how would system builders/OEMs know how much cooling to apply to the SOC?

AMD Ryzen 9 4900HS Review

Mobile computing has become the next big target at AMD with its new series of Ryzen 4000 APUs. Today we have the first retail Ryzen 9 4900HS...

www.techspot.com