A photo of the crystal revealed the details of the architecture of the flagship graphic processor Nvidia Blackwell GB202

An image of the crystal of the flagship graphic processor NVIDIA GB202 appeared on the network, which the GEFORCE RTX 5090 graphics card, published by the Kurnal user from the X/Twitter, showed the location of the key components of the Blackwell architecture, ry (SM), memory controllers and other elements. A comparison of the sizes GB202 and AD102 – the chip of the previous generation used in the GeForce RTX 4090 was also published.

Image source: Kurnal/X

Image source: Kurnal/X

Around the perimeter of the L2 cache are 12 rasterization and 3D conversion units that convert the geometric representation of the scene (for example, polygons) into an image on the screen. At the center of the GB202 die, from top to bottom, are Nvidia video encoders and decoders (NVDEC and NVENC), enabling efficient video playback and recording.

Next, we see that eight 64-bit memory controllers are located on both sides of Kesh L2, and 32-bit physical GDDR7 interfaces connecting the GDDR7 memory modules are located on top, left and right of the crystal. At the bottom of the chip there are a physical interface PCIe 5.0 X16 and a display controller.

If we compare the dimensions of the GB202 and AD102, then the image clearly demonstrates the impressive difference between the chips from each other. The Crystal area GB202 is 761.56 mm², which is 24 % more than that of AD102 (616 mm²). And since Blackwell uses the TSMC N4P process process, an improved version of N4 used in the ADA LoveLace architecture, this explains the significant increase in the physical size of the GB202 compared to AD102.

Image source: Kurnal/X

Image source: Kurnal/X