Fiji is the first graphics processor to feature High-Bandwidth-Memory. In anticipation for second generation HBM (HBM2) which will allow 8 DRAM modules stacked together, AMD had to make a decision whether to wait for HBM2 or help Hynix test HBM1 with first processor. The results of this cooperation is Fiji.
AMD Fiji die area is 596 mm2
Almost a month ago, we concluded that Fiji die area could be somewhere in between 525 – 598 mm2. Well I’m glad to tell you that we finally know how big it is. The die area of 596 mm2 makes it the biggest GPU ever made by AMD, and we’re only talking about the graphics processor not the interposer area which is even larger: 1011 mm2. Fiji is almost as big as GM200, which has the area of 601 mm2. Fiji has 8.9 billion transistors, which means it’s even denser than Hawaii, Tonga or GM200.
Fiji is essentially dual-Tonga fit into one die.
|Graphics Processors Comparison|
|Die Size||359 mm2||438 mm2||596 mm2||601 mm2|
AMD Fiji GPU Block Diagram
If die area does not mean anything to you, this will. Fiji’s block diagram is so big, it hardly fits into one slide.
Fiji has 4 Shader Engines, each with 16 Compute Units. Each Compute Unit has 64 Stream Processors, so in the end we have 4*16*64 = 4096 unified cores. There are eight memory controllers, connected in pair to each of 4 HBM stacks. Unless each of those controllers would be enough to operate with one stack, I don’t believe we are going to see Fiji with more than 4 stacks.