AMD is obviously not capable of incorporating HBM2 memory into their video cards a year from now
It's similar to the evolution of GDDR memory. AMD just happens to adopt HBM1 first (aka HD4870 with GDDR5) while NV managed to skip it due to more efficient colour-compression mechanisms and 2-3MB of L2 cache on Maxwell. Once HBM replaces GDDR5, both AMD and NV will move on to HBM2 and then HBM3, etc. To imply that AMD will be stuck on HBM1 while NV will be on HBM2 is absurd. I've seen quite a few posters on other boards/forums already try to compare Pascal HBM2 to R9 390X HBM1, not realizing that 390X is a competitor to GM200, not to Pascal. Whatever memory technology becomes mainstream, AMD/NV will adopt it. Right now it seems very difficult to do 3D RAM stacking on the GPU die which is why Pascal will also use 2.5D stacking. I don't know why NV called it 3D memory because their Pascal picture shows they are stacking the VRAM modules on an interposer next to the GPU, not on top of it.
This doesn't make any sense. AMD will also be using HBM2, you know. How is GDDR5 better than HBM1?
GDDR5 is better for lower and mid-range chips because it's more cost effective. I think he is implying that because GM200 isn't fast enough, GDDR5 was sufficient for that Maxwell design, meaning had NV used HBM1 for GM200, it would have wasted 50% of the extra memory bandwidth over 336GB/sec. What he doesn't seem to realize is that GM200 and 390X are completely different architectures, completely different chips in terms of transistor composition/density. Historically speaking NV makes 500mm2+ while AMD's largest chip was Hawaii at 438mm2. It's not as simple as just taking Hawaii and making it a 600mm2 GDDR5 card. I am sure AMD's engineers had their reasons for why going HBM1 was superior to GDDR5 for their chip/design architecture. I think both NV's and AMD's engineers made smart choices for different reasons.
However, the earlier head start with HBM1 for AMD should benefit them in theory. Pascal will be NV's 1st generation HBM memory controller design but AMD's 2nd. Last time AMD went GDDR5 first, they were able to get much higher GDDR5 clocks than NV did.
HD5870 (AMD already had GDDR5 for nearly 2 years) = 4.8Ghz GDDR5
GTX480 (NV relative new to GDDR5, with only GT240 having it prior) = 3.7Ghz GDDR5
There is a lot more to engineering than simply choosing a technology for today's products. Sometimes you need to take more risks earlier to make transition in the future easier. The fact that it's the more financially vulnerable AMD that continues to take greater risks (first to AIO CLC with 295X2, first to GDDR5 with 4870, first to HBM with 390X) is more remarkable to me. It takes a lot of b**lls to go for the riskiest and latest tech when you are not on top of the world like Intel or Apple is. When NV goes HBM2 with Pascal, it won't be as impressive as AMD going to HBM1 first because someone else has already done it 1.5 years before you and
actually took that risk.
This is why right now the Samsung's Exynos 7420 is so impressive - they actually took the risk of going 14nm in the SoC world before anyone else! Sure, there will be faster SoCs with time (perhaps A9, etc.) but from a technological leap point of view, 7420 started the next breakthrough wave of SoCs first just like it will be AMD who will start the revolution for next generation graphics memory.
Facebook
Twitter