Kurnal

Kurnal

A brief discussion on the Tianji 9200.

Here is Kurnal
Thanks to the British Shorthair for the Die

Let's talk briefly about the Dimensity 9200
First of all, the Die we received is from the Vivo X90
which is a mass-produced version
Its TopDiemark is MT6985, the previous generation Dimensity 9000 was MT6983
image
Insulting MTK here

Diemark#

This is the Diemark of MTK
IMG_5671(20230720-181406)

image
24FEB2022
AHJ11296B
The first line is the Time
The second line is an unknown number
Time is Day Month Year
So it was produced on February 24, 2022
It was released on February 12, 2022

Dieshot#

Let's take a look at the Dieshot

D9200-web1
The Diesize is 11.34x10.63mm, which is exactly 120.5
image
It is known that TSMC N4 is 146mtr (6T)
There is a difference of 5e

image
And because of the 9-year compulsory education, the chickens and rabbits are in the same cage
It is known that the density of 6T is 146MTR, what about 7.5T?
What is the usage of the 7.5T library in D9200?
The answer is around 25%-20%
So let's do a graph extraction

CPU#

So this generation of Dimensity uses a 1+3+4 three-cluster configuration
This is its CPU Cluster
cpumtk
The layout is basically the same
What's special is that the sandwich structure in D9K has become a structure similar to "n"
The CPU L3 has become a CPU wrapping the L3
This may have some improvements in cache access latency
And the layout of the CPU cores has changed from lying down to standing up, reducing the overall width of the CPU Cluster

In terms of L3 Cache
From the previous inability to distinguish between L3 tag logic and irregular cache shapes
It has evolved into a regular-shaped cache with obvious tag logic, and its cache design is more like SDM's design
L3mtk

In terms of microarchitecture, it has changed from the previous generation's
X2+A710+A510c
to
X3+A715+A510c

It is still made by TSMC N4, and it is speculated that it uses the density library
After all, compared to the core of the same generation of competing products
CPU Core

Then there is the issue of Core shot
In fact, you can see that the cores of each company are different
Let's take the common A510 as an example
A510c
The L2 Cache, FE, and shared FP below are all different
This is enough to prove that although each company uses the Arm microarchitecture, they have made changes instead of staying the same
The so-called Kyro also proves that it is not purely a public version, and the specific changes are not clear

In terms of SLC

SLCmtk
There is no difference from the previous generation, still 2x3M 6M SLC, only the tag logic has changed

GPU#

图层 117

So in terms of the GPU, it has been upgraded from the previous generation's Mali G710 MP10 to Mali-Immortalis-G715

GPU Core

Although they belong to the same Valhall architecture, there are additions

image
Among them, the Ray Tracing unit (RTU) module
is placed in the shader core
It only occupies 4% of the shader core area
but achieves a 300% improvement in gaming performance (ppt)

Similarly, in its core configuration

Core Config
There is no change in G715i compared to G715

Although the G710 doubled the number of single-core compute units compared to the previous generation G78
the G715 doubled the number of single-core compute units again
Each core has 128 FMA
which can complete 256 FP32 operations per clock
while the pixel and texture capabilities remain unchanged.

In terms of texture units, there is no change compared to G710
In the G710 architecture

Texturing

There is also no change in the ISA configuration

图层 114

There is no change in the GPU Cache in the Dieshot, it is determined to be 3MiB

APU#

MTK actually used a large area to write this APU

图层 115

The previous generation was 4+2
and this generation is also 4+2
There is not enough information to see the difference

APU
Only a slight change can be seen in the big core, the general core remains basically unchanged
From MTK's official website:
The sixth-generation MediaTek APU has upgraded the shared memory engine
which improves the computational efficiency of the APU's VPU, DMA, and DLA processors.
The new generation network architecture search technology brings better performance and power consumption for machine learning (ML) applications.
Thanks to the eXtreme power-saving mode and APU hardware upgrades
Compared to the fifth-generation APU (APU 590)
APU 690's AI performance has increased by up to 35%
AI video super-resolution (AI-SR) efficiency has increased by 45%
AI noise reduction (AI-NR) efficiency has increased by 30%.

I'm not talented enough to annotate

Modem#

There have been some changes in the Modem section
But the GPU's Core is still learning and hasn't learned how to draw the layout yet
So let's just take a look

图层 116

That's it
Here is Kurnal
Please indicate the source as Kurnal

Insert a footnote (Easter egg)
Actually, we also got a media sample of the D9200
After X-ray, it was found to be pure plastic without any Si elements
微信圖片_20230629212726

Hilarious
That's it

Loading...
Ownership of this post data is guaranteed by blockchain and smart contracts to the creator alone.