The problem is that if you need 1½GB. Then the 1GB card in your case will crawl.
You missed his point completely. For 6 months you keep repeating this:
HBM1 @ 1Ghz = GDDR5 512-bit 8Ghz
1) What if it's
not physically possible to design a card 40-60% faster than R9 290X with a GDDR5 512-bit 8Ghz controller in a sub-300W power envelope? What if the 3X performance/watt with HBM1 over GDDR5 is 100% true? That would mean 50-60W+ of extra power usage that would have been wasted on GDDR5 but can now be used for higher clock speeds and enlarging the die (i.e., since you can now accommodate an extra 50-60W of power towards performance transistors)?
2) What if it's not physically possible to fit a 512-bit memory controller inside the same die space without sacrificing "performance transistors" allocated for shaders, TMUs, ROPs, L2 cache, etc.? In other words, a 512-bit GDDR5 memory controller could have meant 20-30mm2 die space wasted over HBM1.
3) What if R9 390X has 640GB/sec memory bandwidth (i.e., HBM1 @ 1.25Ghz), then how how would you make a GDDR5 card with as much memory bandwidth? Not possible.
4) What if HBM1 allows AMD to make a miniITX card with 40-50% more performance over R9 290X due to reduced PCB complexity and moving all of the memory closer to the GPU die? That would result in the world's fastest miniITX card.
5) Extending from point #4, what if going HBM and miniITX allows AMD to release a card where the memory, the GPU and the VRMs are all 100% water-cooled by a 120mm AIO CLC? That would be the most epic design anyone has done. Even EVGA's 980 Hybrid can't claim this. The only way today to have the memory, VRMs and the GPU all water-cooled is to put a full-sized block on a videocard. If the HBM1 is positioned closely to the GPU and the VRM is not far out, it's possible to cool all of them in 1 shot with an AIO CLC and a 'mini' water-block since the PCB is no longer 11-12.5", which means it becomes more cost effective to pursue such a solution.
In that context, if a GDDR5 512-bit Fiji would be way too power hungry OR to keep it at a reasonable TDP AMD would have needed to reduce performance by 15-20% (smaller die size, less transistors, lower GPU clocks), then no doubt 4GB HBM1 >>>>>>>>> 8GB GDDR5 for AMD's flagship card.
It's better to have a card 30-40% faster than GTX980 with 4GB of HBM1 than a card just 10-20% faster than a 980 with 8GB of GDDR5 because performance sells flagships cards
before VRAM is even a part of the discussion. If you have 32GB of GDDR5 but your card is barely faster than a 980, no one cares to spend $700 on it. That's why all the arm-chair engineers here suggesting that 4GB HBM1 is an automatic fail over 6-8GB GDDR5 are missing the critical pieces of the puzzle because they can't quantify all the key benefits of HBM1 over GDDR5.
Excellent post JDG. :thumbsup: