NVIDIA Ampere GA106-400 GPU pictured – VideoCardz.com

NVIDIA Ampere GA106 GPU pictured

VideoCardz was the first to give you a look at GA104 GPU, so now it is time for GA106.

NVIDIA is set to launch its GeForce RTX 3060 graphics card by the end of February. The graphics card will feature a new GPU codenamed GA106. This is a fourth member of the NVIDIA Ampere family after GA100, GA102, and GA104.

This GPU pictured below is codenamed GA106-400, the actual full chip. NVIDIA was originally planning to launch RTX 3060 with 3840 CUDA cores but as we approached the announcement the manufacturer decided that RTX 3060 will use GA106-300 instead with 3584 cores. This is why the picture has a GPU labeled as 400 – it’s an early engineering sample, and also proof that this was the original configuration of the card. Do note that we also have a picture of GA106-300, and they look identical.

NVIDIA GeForce RTX 3060 will feature K4ZAF3258M-HC16 memory clocked at 16 Gbps. There are 6 modules in total, each carrying 2GB of memory. It appears that GA104 and GA106 GPUs use the same package size, so they are likely pin-compatible. This greatly reduces the development cost for AIBs. This is also why we are seeing two unoccupied memory slots. The RTX 3060 Ti with GA104 GPU uses all eight slots, but memory is limited to 1GB per module and 14 Gbps speed. In other words, this board could easily be adopted for RTX 3060 Ti or even RTX 3050 Ti, which is also rumored to use GA106 GPU in the future.

NVIDIA GA106-400 GPU, Source: VideoCardz

We measured the GPU and the GPU appears to have a die size of 276 mm2:

NVIDIA GA106-400 vs GA104-300 GPU, Source: VideoCardz

In comparison to Turing GPUs, the GA106 is noticeably smaller than TU106 (445 mm2), and just a bit smaller than TU116 (284 mm2).

NVIDIA Ampere Graphics Processing Units
VideoCardz GA102 GA104 GA106 TU106
Architecture NVIDIA Ampere NVIDIA Ampere NVIDIA Ampere NVIDIA Turing
Fabrication Node 8nm SAMSUNG 8nm SAMSUNG 8nm SAMSUNG 12nm TSMC
Die Size 628 mm2 392mm2 ~276 mm2 445 mm2
Transistors 28.3B 17.4B TBC 10.8B
CUDA Cores 10752 6144 3840 2304
Tensor Cores 336 192 120 288
RT Core 96 48 30 36
Memory Bus 384-bit 256-bit 192-bit 256-bit

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *

%d bloggers like this: