—
The first photo of the mainstream GPU Ampere has appeared. It is interesting that it is said to have a version of the chip with more shaders, instead of which a different and weaker one will eventually be used.
Today we wrote about photos of extremely complex GPU for Intel Xe HPC supercomputers (probably contains at least 36 chipsets). It seems that coincidentally we now have something similar here again from Nvidia. This time, unofficially and perhaps unintentionally, she ran away for a change in a photo of her GPU GA106, which is, however, intended for the completely opposite end of the market. This is a relatively inexpensive GPU for the Ampere architecture that will run on – hopefully – the mass-available GeForce RTX 3060. Now we can see what it looks like.
The picture of the GA106 chip was published by the VideoCardz website, which is very fruitful in this discipline. The image shows a chip powered by a desktop card, probably a GeForce RTX 3060. You may notice that there are six GDDR6 chips around, indicating a 192-bit bus.
Tip: Nvidia introduces Ampere for the masses: GeForce RTX 3060 has 12 GB of memory
In addition to the GPU, there are two other unoccupied positions on the PCB, so the board itself may have a 256-bit bus. Why would that be so? According to some reports, the contacts on the underside of the GA106 chip housing are designed to fit on the card PCB. GeForce RTX 3070/RTX 3060 Ti with GA104 chip. With the fact that a quarter of the 256-bit bus that the board counts on is simply not used. Such compatibility allows you to save on PCB design, or allows the use of scrap from the production of higher models in RTX 3060 cards.
The GDDR6 chips in the photo are 16Gb (2GB), which proves the 12GB capacity of the graphics, which has already been officially confirmed by Nvidia. It should even be chips that officially support the 16.0 GHz frequency (effectively).
The GPU itself is estimated by VideoCardz to have an area of about 276 mm², which is similar to how much the 12nm TU116 chip occupied in the GeForce GTX 1660 (Ti / Super). GeForce RTX 2060 used a significantly larger GPU TU106 with an area of 445 mm².
Did Nvidia chop the GPU at the last minute?
The photo has one peculiarity. GeForce RTX 3060 already has a confirmed configuration of 3584 shaders, which is truncated against the full version, which probably has 3840 shaders. This configuration should be labeled GA106-300. Originally, Nvidia reportedly planned to use a variant of GA106-400, which could probably have 3840 shaders. The chip in the picture is marked GA106-400, so it is probably this variant, which in the end will not be in GeForce RTX 3060 desktop cards.
VideoCardz states that the photo captures an ES sample, but the GPU says “Qual Sample”, which would mean a qualifying sample. While ES are test chips from an earlier stage of development for testing and development, the qualification sample should match the serial product in almost everything and is used for verification. If there are qualifying samples of GeForce RTX 3060 cards with 3840 shaders, then it probably means that the decision to reduce the number of shaders came very late, maybe it was made almost at the last minute. The question is whether Nvidia has a problem producing enough chips with 3840 functional shaders, or just found out that 3584 will be enough against competing AMD cards.
Except for the inscription, however, the appearance of the chip version cut to 3584 shaders will not differ, so this photo is still “valid”.
Tip: Architecture and details of Nvidia Ampere / GeForce RTX 3000 GPU and GA102 and GA104 chips
Gallery: Architecture and details GPU Nvidia Ampere
Source: VideoCardz
