AMD’s much anticipated Navi GPU has finally been scrutinized by the impatient. Well known leaker on Twitter, KOMACHI posted the block diagram of the Navi chip, although it has now been taken down. However, as you know, once something leaks out to the internet, it’s pretty much impossible to contain. The GPU in the block diagram features 8 Shader engines, each packing 5 Compute Units, making a total of 40.
If this is indeed legit, then that means the Navi graphics card will have 2560 streaming processors (shaders or cores), as each Compute Unit (CU) contains 64. And since there are 40 CUs, distributed across 7 Shader Engines, you get that figure. The shader engine count here is an important tit-bit, as till now, AMD’s GCN architecture has been able to stuff a maximum of 4 SEs per die, limiting the shader and ROP count to 4096 and 64, respectively. Due to only 64 Render Output Units, AMD’s shaders otherwise known as “Streaming Processors” have been less efficient compared to NVIDIA’s CUDA Cores and that’s why Radeon GPUs need to have more of them than their GeForce counterparts.
If AMD leaves the block unchanged, then Navi will have a total of 128 ROPs which is a first, along with 2560 cores for better utilization. Now, from earlier leaks we already know that the graphics cards succeeding Polaris will come with 8GB of GDDR6 VRAM paired with a 256-bit wide bus, resulting in around 400Gb/s of bandwidth. That pretty much paints the complete picture of the Navi specs-sheet. It’s not far now, just four days left to AMD’s Computex press