This may be the most rigorous article you can find discussing SM8550
Please indicate the author when reposting
Cause#
The author obtained a SM8550 (Honor Magic5Pro) motherboard
So let's start
This product is a POP package (Package on Package)
TOP Package#
The top package mark is
2WB77 D8CCP
According to the query:
The manufacturer's code is MT62F1536M64D8CL-023 WT,
which means that the particle is LPDDR5 96G 1.5GX64 FBGA 8DP
So let's decap it
It can be known that
this Micron D8CCP particle is actually a product of Samsung SEC
SEC refers to Samsung Electronics Company
Very strange
Bottom Package#
After opening the top, it's time for the bottom
This time we got SM8550, which is 8Gen2
I won't write a detailed performance analysis, you should be able to find it yourself
This time we got the mass-produced SM8550,
which is SM8550 002-AB
The complete code is
Qualcomm-SM8550-002AB- FX215S2Y
Top Mark Analysis#
SM stands for Snapdragon Mobile
8550 is the product model, which is the SM 8 series 5th generation, also known as 8gen2
002-AB is the variant
FX215S2Y is the top mark code
It is xx/x xx/xxx
The first xx is the manufacturer + packaging factory
The second xx is the Year + Week
The third xxx is the traceability code
F stands for TSMC production, X stands for Amkor, Japan packaging
2 stands for 2022, 15 stands for the 15th week
S2Y is the traceability code, the specific meaning is unknown
These are the data that can be obtained from the top mark
Bottom Package Decap#
Next is the bottom package decap
Unfortunately, it was damaged
Considering cost-saving, let's see if we can salvage some data
Obviously, this magnification is not enough
So let's increase the magnification
For easier viewing, let's switch to a good die
The die mark is in the center of the image
Qualcomm HG11-31145-2
The specific meaning is unclear
Marked area
The topmost layer is the fanout metal layer
Good Die#
The reason we can't see deeper is because there is a metal layer blocking the view of the poly
Measurement
The die size is 11.20mm x 10.57mm
which is 118.384mm2
It is known to have 17 billion transistors
The average density is calculated to be 143.6004MTr, which matches the TSMC N4 density
The actual number of transistors is calculated to be 170.117808 billion
Process Analysis (SEM) (not yet written)#
Although it is known that this is TSMC N4
verification is needed, which requires SEM (still pending)
Process Analysis (not yet written)#
The analysis here is simple
This is TSMC's technology roadmap
It can be seen that N4 is actually a branch of N5
Inherits the 6TUHD/7.5THD library from N5
The Fin High of 6T UHD is
Specific data
Fin Pitch:
Yield Analysis#
So let's simply calculate the yield using the dpw formula
Still using this formula
According to the calculation based on N5 D0
When D0=0.75, the yield is 43.99%, 238/540 pieces
Of course, there is also thermal design power simulation
Of course, this is just a rough calculation, no need for something like Hi36a0 to calculate capacity/yield analysis)
(from ISSCC 2023 / SESSION 2 / DIGITAL PROCESSORS / 2.2)
Dieshot Layout#
So it's time for acid etching and layout drawing
Because the dieshots I've been getting are all low-resolution garbage, I decided to create a high-resolution dieshot myself
Okay, it's done
Now let's extract the image separately
This is the size of the CPU Cluster+L3+SLC
Now let's extract the image
CPU#
This is a comparison of the Core Size of X3 A710 A715 A510 A510Cluster, it can be clearly seen that Sdm is actually a 5-cluster design, not 143 or 1223, but how it is tuned specifically, you can find out by looking at an 8gen2 phone, I don't have one
GPU#
This is the GPU of the 8G1+ series, it can be clearly seen that it is a 2-cluster structure, each cluster has 4 cores (2CU)
The characteristic of Adreno is that its Tmus are not inside the Alus Core, so the labeling is approximate
The advantage of Adreno 740 over Adreno 730 is the addition of 1 cluster, reaching 12 cores (6CU 3 clusters)
The number of Alus reaches 1536, the frequency is marked incorrectly but I'm too lazy to change it, the actual frequency is 680Mhz, and the theoretical performance is 1536x0.68Ghz x2 = 2088.9Gflops
This is a comparison of the GPU size of this generation/previous generation, the smallest is actually the GPU of SM8475
Please measure the specific size yourself, I haven't learned it yet
Process Judgment (SEM) (not yet written)#
Although it is known that this is TSMC N4
verification is needed, which requires SEM (still pending)
Process Analysis (not yet written)#
The analysis here is simple
This is TSMC's technology roadmap
It can be seen that N4 is actually a branch of N5
Inherits the 6TUHD/7.5THD library from N5
The Fin High of 6T UHD is
Specific data
Fin Pitch:
Yield Analysis#
So let's simply calculate the yield using the dpw formula
Still using this formula
According to the calculation based on N5 D0
When D0=0.75, the yield is 43.99%, 238/540 pieces
Of course, there is also thermal design power simulation
Of course, this is just a rough calculation, no need for something like Hi36a0 to calculate capacity/yield analysis)
(from ISSCC 2023 / SESSION 2 / DIGITAL PROCESSORS / 2.2)
Dieshot Layout#
So it's time for acid etching and layout drawing
Because the dieshots I've been getting are all low-resolution garbage, I decided to create a high-resolution dieshot myself
Okay, it's done
Now let's extract the image separately
This is the size of the CPU Cluster+L3+SLC
Now let's extract the image
CPU#
This is a comparison of the Core Size of X3 A710 A715 A510 A510Cluster, it can be clearly seen that Sdm is actually a 5-cluster design, not 143 or 1223, but how it is tuned specifically, you can find out by looking at an 8gen2 phone, I don't have one
GPU#
This is the GPU of the 8G1+ series, it can be clearly seen that it is a 2-cluster structure, each cluster has 4 cores (2CU)
The characteristic of Adreno is that its Tmus are not inside the Alus Core, so the labeling is approximate
The advantage of Adreno 740 over Adreno 730 is the addition of 1 cluster, reaching 12 cores (6CU 3 clusters)
The number of Alus reaches 1536, the frequency is marked incorrectly but I'm too lazy to change it, the actual frequency is 680Mhz, and the theoretical performance is 1536x0.68Ghz x2 = 2088.9Gflops
This is a comparison of the GPU size of this generation/previous generation, the smallest is actually the GPU of SM8475
Please measure the specific size yourself, I haven't learned it yet
Modem#
The modem part of X70 is not significantly different from the previous generation X65,
Of course, there may be some minor changes, but no one studies basebands
Size comparison
The baseband of D9200 may be M80, but I'm not sure
ISP#
Qualcomm always has 3-core ISPs every year, but they are always elusive, so I just made a random mark, and this year's ISP has no movement
Strange Unit#
This is a very strange unit, we only know that the top is Internal Memory, the bottom is unclear, the area has increased by 1.5-1.7x this generation, very strange
Theoretical Performance#
Detailed performance tests can be found in many media sources, no need to read from me, so I won't write it
That's it
Explanation#
Please indicate the author as @Kurnal when reposting
Feel free to repost the images, just indicate the author as @Kurnal
Why create Dieshot: Because other people's images are too low resolution
The first update is on https://kurnal.xlog.app/SM8550, others are reposted/translated by me
Thank you for the Alipay v50, meow