Discussion Mediatek SoC thread

[FlameTail]

According to Kurnal, Dimensity 9400 uses HP libraries for the Cortex X925 prime core.

Chip	Prime core	Library	Clock Speed
Dimensity 9400	Cortex X925	N3E HP	3.6 GHz
A18 Pro	A18-P	N3E HD	4.04 GHz
M4	M4-P	N3E HP	4.5 GHz
8 Elite	Oryon-L	N3E HP	4.32 GHz

Stock ARM cores still have a notable clock speed gap with custom ARM cores. Presumably due to some microarchitectural limitation?

PS: There's also a rumour that Google Tensor G7 will use their own custom cores, which has 16 MB shared L2 and a clock speed of 5 GHz.

Rather strange isn't it, that all these custom cores from Apple, Qualcomm and Google have many similarities: big shared L2, high clock speed, etc...

 

[gdansk]

Rather strange isn't it, that all these custom cores from Apple, Qualcomm and Google have many similarities: big shared L2, high clock speed, etc...

There may be a two or three different ways to build a high performance core. But there's only one way known to build a high performance low power core and Apple showed how.

 

[Meteor Late]

According to Kurnal, Dimensity 9400 uses HP libraries for the Cortex X925 prime core.

Chip	Prime core	Library	Clock Speed
Dimensity 9400	Cortex X925	N3E HP	3.6 GHz
A18 Pro	A18-P	N3E HD	4.04 GHz
M4	M4-P	N3E HP	4.5 GHz
8 Elite	Oryon-L	N3E HP	4.32 GHz

Stock ARM cores still have a notable clock speed gap with custom ARM cores. Presumably due to some microarchitectural limitation?

PS: There's also a rumour that Google Tensor G7 will use their own custom cores, which has 16 MB shared L2 and a clock speed of 5 GHz.

Rather strange isn't it, that all these custom cores from Apple, Qualcomm and Google have many similarities: big shared L2, high clock speed, etc...

A18 having HD and M4 having HP is weird to me, as previously there was like a 10% bump in frequency between A and M variants, just like now. Something is not right here.

 

[FlameTail]

A18 having HD and M4 having HP is weird to me, as previously there was like a 10% bump in frequency between A and M variants, just like now. Something is not right here.

Ah, that's where you might be wrong.

M4 N3E HP 4.5 GHz	A18 Pro N3E HD 4.04 GHz
M3 N3B 4.05 GHz	A17 Pro N3B 3.78 GHz
M2 N5P 3.5 GHz	A15 Bionic N5P 3.24 GHz
M1 N5 3.2 GHz	A14 Bionic N5 3.1 GHz

With the exception of M2/A15, it has been about a ~5% difference.

 

[FlameTail]

An update on that;

I have sources for M4, D9400 and 8 Elite using HP libraries for the P-core.

But the statement that A18 Pro uses HD libraries was an assumption by me, on the basis that the M4 vs A18 frequency gap is larger than the M3 vs A17 frequency gap.

It seems I was wrong to make that assumption, so I beg your pardon, good people. I took the dieshots and measured the P-core areas in A18 Pro and M4. The area is nearly identical (M4's is about 1% larger, but that could very well be measurement error). So it seems the P-cores of M4 and A18 Pro use the same libraries.

Also another thing to note is that TSMC 3nm has FinFlex, which blurs the line between HP and HD libraries. In TSMC 5nm, HP was 3-fin and HD was 2-fin. But FinFlex introduces new combinations such as 3-2 fin and 2-1 fin.

 

[Meteor Late]

An update on that;

I have sources for M4, D9400 and 8 Elite using HP libraries for the P-core.

But the statement that A18 Pro uses HD libraries was an assumption by me, on the basis that the M4 vs A18 frequency gap is larger than the M3 vs A17 frequency gap.

It seems I was wrong to make that assumption, so I beg your pardon, good people. I took the dieshots and measured the P-core areas in A18 Pro and M4. The area is nearly identical (M4's is about 1% larger, but that could very well be measurement error). So it seems the P-cores of M4 and A18 Pro use the same libraries.

Also another thing to note is that TSMC 3nm has FinFlex, which blurs the line between HP and HD libraries. In TSMC 5nm, HP was 3-fin and HD was 2-fin. But FinFlex introduces new combinations such as 3-2 fin and 2-1 fin.

Yeah, I've checked and I've seen 3 GHz, not 3.1, on A14 vs 3.2 GHz on M1, that's like 7% bump, same with M3. It wouldn't make sense to see such a small bump going from HD to HP. You don't go from 7% before to just 11% now using HP vs HD.

 

[DZero]

Still no info regarding the Mediatek 8400. Something happened since Redmi won't have an KXXE series anymore.

 

[DZero]

And HERE WE GO!
New info about the Dimensity 8400!

Complete core details of MediaTek Dimensity 8400 exposed ahead of December 23 launch - Gizmochina

The Dimensity 8400 features 1 Cortex-A725 core clocked at 3.25GHz, 3 Cortex-A725 cores running at 3.0GHz, and 4 Cortex-A725 cores at 2.1GHz

www.gizmochina.com

CPU:
1 x A725 - 3.25 Ghz
3 x A725 - 3.00 Ghz
4 x A725 - 2.10 Ghz

GPU:
Inmortalis MC7 - 1.3Ghz

Antutu estimated: 1.8m

 

[FlameTail]

That will have serious multi-core grunt.

Regarding the GPU, it has to be Mali G720 MC7.

It can't be Immortalis G720 MC7 because Immortalis has to have a minimum of 10 GPU cores.

Immortalis-G720

Arm's latest flagship GPU, designed to provide ultimate gaming experiences on next-generation premium smartphones.

developer.arm.com

 

[LightningZ71]

I wonder how cut down the 4 compact cores are. We know that there are options on cache and functional units for the 7xx cores in addition to the choice of doing what AMD does and reduce cell padding and extra transistors to support higher frequencies. Those things could be quite small.

 

[DZero]

With the Dimensity 8400 is confirmed that will be full Octa Big Core, I think how the Dimensity 7400 will be.
Considering that the Configs of the D7300 and D7350 goes like that, expect the D7400 going like this, fusing the best of the 2 former processors.

	D7300	D7350	D7400 (presumed)
Big Core	4x A78 @ 2.5Ghz	2x A715 @ 3.0Ghz	4 x A720 @ 2.8Ghz
Small Core	4x A55 @ 2.0Ghz	6x A510 @ 2.0Ghz	4 x A520 @ 2.0Ghz
GPU Core count	Mali G615 MC2	Mali G610 MP4	Mali G620 MP4

 

[FlameTail]

Dimensity 8400 is here!

1x Arm Cortex-A725, 1MB L2
3x Arm Cortex-A725, 512MB L2
4x Arm Cortex-A725, 256KB L2
6MB L3

Arm Mali-G720 MC7

5MB SLC

MediaTek | MediaTek Dimensity 8400

www.mediatek.com

 

[DZero]

Waiting for the reviews and preparing the trail for the Dimensity 7400.

EDIT: They duplicated the L2 cache per core! That's insane! These are from the D8300 and D8200

 

[adroc_thurston]

Dimensity 8400 is here!

1x Arm Cortex-A725, 1MB L2
3x Arm Cortex-A725, 512MB L2
4x Arm Cortex-A725, 256KB L2
6MB L3

Arm Mali-G720 MC7

5MB SLC

MediaTek | MediaTek Dimensity 8400

www.mediatek.com

Man if only this had a single Cortex-X4 (X5 is too chungusbloat for midrange).

 

[Curious_Inquirer]

Here’s the 8400 vs the 8 Gen 3. Looks like it’s in the range of the previous gen flagship of Qualcomm according to Mediatek.

https://twitter.com/x/status/1871104148709806279

 

[gdansk]

1T still crap?

 

[DZero]

1T still crap?

Indeed, for emulation will be poor...

 

[Meteor Late]

Yeah these chips are mainly for marketing purposes unfortunately, sounds good on paper until you realize that, on the vast majority of workloads, MT performance and efficiency are irrelevant because you are not going to be using all the 8 cores.
What if, instead of 8 Cortex A725 cores, one uses instead 1 X4 core + 1 A725 core + 4 A725 cores instead? are you telling me this chip won't be much better for the vast majority of the workloads a consumer will run on his smartphone?
Why are we having this fixation with 8 cores in the midrange? can't thing about anything other than marketing to claim high Geekbench MT and Antutu scores.

 

[gdansk]

Why are we having this fixation with 8 cores in the midrange?

It is a good question. If MTK aims to deliver the best user experience a X4 and two types of A725 would likely be better. The actual product suggests that best user experience in the given area is not their priority.

But someone here has claimed Android has a "8 core" preference/expectation. I'm not convinced, how much would the scheduler have to be messed up to ruin Linux like that? Well maybe it is possible.

 

[DZero]

Yeah these chips are mainly for marketing purposes unfortunately, sounds good on paper until you realize that, on the vast majority of workloads, MT performance and efficiency are irrelevant because you are not going to be using all the 8 cores.
What if, instead of 8 Cortex A725 cores, one uses instead 1 X4 core + 1 A725 core + 4 A725 cores instead? are you telling me this chip won't be much better for the vast majority of the workloads a consumer will run on his smartphone?
Why are we having this fixation with 8 cores in the midrange? can't thing about anything other than marketing to claim high Geekbench MT and Antutu scores.

Funny story in gaming and some heavy apps do use the 8 cores. But you are right in that matter.

The issue is that Android does HATE the In Order executions. Even more, In order executions are trash even for light loads now and is rendering totally useless in the long run. And using in order cores are a no go anymore, even Huawei is leaving it behind. And only is used on pieces of trash using the Helio G35/G50 or the Unisoc at 28 nm.

 

[DZero]

Meanwhile...
Geekerwan reviewed it...

Interesting enough, the Single core is nearing the performance of the X2 Core, while Multi core wise is near the SD 8 Gen 3!




Sad story the GPU is still the issue. Only reaches the SD8 Gen 2 performance






Is pending to see a device using it properly. But for a Full Big Core processor looks decent.

 

[Tup3x]

Sad story the GPU is still the issue. Only reaches the SD8 Gen 2 performance

That's perfectly adequate for 8000 series SoC.

 

[hemedans]

Indeed, for emulation will be poor...

Mali based devices are poor anyway, with or without 1T.

 

[DZero]

Mali based devices are poor anyway, with or without 1T.

Actually is everything non Adreno, is like nVIDIA and AMD in GPUs or Intel and AMD in CPUs.

 

[hemedans]

Actually is everything non Adreno, is like nVIDIA and AMD in GPUs or Intel and AMD in CPUs.

Xclipse is very competitive it just lack custom drivers, even Midrange like Exynos 1480 can run many switch games.