Exactly see, as if DDR power consumption was ever at a head? It will now be as it never existed!
The sad part is that the usual suspects are ready to crap on HBM because it supports their previous positions how HBM was a waste of time, but the reality is just like Maxwell / Fiji would have benefited tremendously from a 20/22nm node, Pascal would be way better off from a perf/watt if all cards top-to-bottom were HBM2 because HBM is superior to GDDR5X because it's impossible to compete in perf/watt and PCB space when HBM2 just needs 1-1.25Ghz clocks to reach 1TB/sec. Who knows if NV can execute even better than AMD and even manage to reduced latency in the process.
People keep regurgitating how HBM brought nothing to the table over GDDR5 without realizing they are not comparing apples-to-apples architectures. This is akin to making a blanket statement during the early eras of hybrid cars that they bring little over the best diesel/gasoline cars, but over time the 3 best hypercars in the world all have hybrid powertrains and use all the benefits of electric power to improve performance. In that sense, we do not have the full picture if we just compare GM200 to Fiji and make statements how HBM is barely better than GDDR5 because GM200 is a far more efficient GPU architecture than GCN 1.2 to start with.
To truly understand the incredible impact HBM has achieved in its 1st iteration, we need to compare GCN to GCN in GPU limited scenarios to see what's what:
Perf/watt
1080P
Fury X = 100% (+33% vs. 290X)
290X = 75%
285 = 72%
1440P
Fury X = 100% (+39% vs. 290X)
290X = 72%
285 = 67%
4K
Fury X = 100% (+43%)
290X = 70%
https://www.techpowerup.com/reviews/AMD/R9_Fury_X/32.html
While it's fair to say that not all of the perf/watt improvement is fully attributable to HBM1 over GDDR5, anyone who says HBM brought
nothing/little to the table over GDDR5 is spreading FUD. If engineers could have integrated HBM1 on GM200 Maxwell, it may have had another 20-30% improvement in perf/watt that could be utilized toward a faster GPU or lower power consumption. Not having HBM2 on all Pascal cards top-to-bottom doesn't mean Pascal won't be good but it for sure would mean that not all Pascal chips would be as great as they could have been
assuming this rumor is true.
More worrying is the write-up implies that a GP104 may have 256-bit bus with 448GB/sec GDDR5X, while the flagship GP100 will be positioned as a professional Titan X/Tesla/Quadro product at first with HBM2 1TB/sec. That implies
3rd consecutive generation of milking the mid-range chip and making it a Marketing Flagship. D: If NV follows through with this strategy for a 3rd generation in a row we might as well forget decades of how GPUs were released because it would mean marketing has won and it will officially become a new era of GPU generations = bifurcating a generation where the mid-range acts as a flagship for the 1st part of the milking process of a new architecture. :twisted:
With HBM2, HMC, or other similar technologies you can get better than linear scaling on power consumption for increasing both bit density and bandwidth ...
If Big Daddy GP100, the true flagship Pascal, has HBM2 then that in itself would be 100% confirmation from NV themselves that GDDR5X is technologically inferior in all key aspects other than price and time to mass market production compared to HBM2. Despite the detractors here, HBM is the future for the next 5 years even if it has a bumpy road along the way.
This reminds me of all the doom and gloom I read on DDR4 when X99 first launched last summer. In barely more than a year, it's now possible to buy 16GB of DDR4 3200 for about
HALF the price of 16GB DDR4 2133 when X99 just launched. In fact, it took just over a year and we already have consumer
DDR4 4000 on Newegg, nearly double the speed of the 1st generation DDR4.
HBM might have a rough ramp-up but there is no doubt that GDDR5 or any of its derivative is just a short-term band-aid to what is outdated and budget technology. The amount of benefits HBM brings over GDDR is just too immense to ignore, from perf/watt, to total bandwidth, to ability to stack so much more memory in compact space, to reduced PCB size and the memory itself is cooled directly by the GPU heatsink. With time, like DDR4, HBM2 will obsolete GDDR5 just like GDDR5 obsoleted DDR3 on GPUs, etc.
GDDR5X is faster than HBM1.
Even at 16Gbps @ 384-bit bus, that's 768GB/sec, well short of 1TB/sec of HBM2 and that's just the 1st revision of HBM2.
What matters is bandwidth/watt and absolute memory bandwidth. It seems that in your dreamworld, GDDR5/X will scale linearly with memory bus regardless of the required complexity, but in the real world it's probably not that simple and it could be an engineering roadblock to go from a 256-bit memory controller with 16Gbps GDDR5X and to a 384-512 bit buses with the same GDDR5X memory speeds. Also, let's not ignore that perhaps a 256-512 bit memory controller could take up much more transistor die space than the HBM2 one.