[EDIT]hmmm.....whats going on at AMD? (and history of 3Dfx Rampage)

[ocre]

Yo Ocre, I've been thinking that AMD didn't get to completely solve the power consumption issue on GloFo's new 14nm process.  If AMD didn't diverge from TSMC and stuck with the same 16nm process as NV, I think that AMD had their RX 480 ready for launch maybe 6 months ago...  with lower idle power consumption than that of Tonga R9 285 (and lower load consumption than that of Pitcairn XT R9 270X -- less than 120W on average).  See how great the new Pascal cards are at idle consumption:




Could it also be that 8GB GDDR5 clocked at 8GHz effective would consume about 5W more than 7GHz that AMD is apparently setting lots of these 380 cards at default (not just for 4GB but 8GB cards as well)?  Furthermore, how much would 8GB consume over 4GB?  My guess is around 15W.

Okay,let's see. AMD went with GloFo's 14nm and we see Nvidia on 16nm running far higher clocks and far better perf per watt. So.....that's it!!
It must be the node, solved....

Well not so fast...
a couple things- Nvidia and AMD both used the same exact node last generation, TSMC 28nm. There was also something that should be noted: maxwell was running at higher mhz and was more efficient, on the same node. Maxwell clock speeds were in another league, actually so much so that it could have been a different node.

Amd was far behind in those attributes, already. In an attempt to compete, they stripped their bus and bolted HBM on their gcn. This saved them a large chunk of power, I heard estimates of 50watt saved by dropping their bus. This brought the efficiency closer but still Amd was far far behind in clock speed capabilities. The efficiency gained with Fiji was due to HBM, Amd was not able to get the great results Nvidia did in regards to clock speed and efficiency out of the 28nm tsmc node.

It may be easy to say that glofo or 14nm is their issue, but I tend to think that perhaps all that stuff Nvidia talked about during pascal event...it might mean something. Nvidia claimed they sunk a lot of money in tackling efficiency...maybe they did. Sure looks like they did something special with maxwell, could it not be true that this focus has a lot to do with pascal as well? I believe that the clock speed of pascal is 100% a result of their fine grain engineering and it's not just a byproduct of the node. That Nvidia did put a massive amount of cash and effort in making pascal. There is no way Amd expected such results, I believe they were truly caught off guard. There were also plenty of rumors on the 14nm that Amd was moving to. If you remember, this was supposed to be Samsung's node. Many believed it to be superior to tsmc 16nm. You might expect that Nvidia would take that possibility very seriously, that this possibility would have some concern. It only makes sense that Nvidia would put the extra effort in, to try to get the best efficiency they could as well as the highest clock speeds.

I just can't believe that these results are accidental, that it is simply a byproduct of the super tsmc 16nm node. We already seen Nvidia pulling away in 28nm when both of them were on the same node. I don't see why Nvidia would stop pushing in those same directions. Nvidia claims they put billions in pascal, they talk about the massive effort and the results which were from thousands of ideas chipping away. They claim that this is how they achieved pascal perf per watt and super high clock speeds.

A lot of people down played pascal but it's far more than a die shrink maxwell. Even if the block layout may be the same, Nvidia has been working down to the transistor level and it's been paying off for them in big ways. It took a node shrink for Amd to catch maxwell efficiency and if you imagine that Nvidia pumped billions since the to furthered in these directions- on top of a node shrink- there is little reason to think Amd would not be significantly behind. I doubt that 14nm glofo is the only difference here. We have seen Samsung vs tsmc 14 vs 16 in apple chips, the results were not dramatic. There was good reason to think that Amd might have the node advantage over Nvidia with their 14nm glofo.

We heard that Amd went to full automated chip design for their cpus, which is cost effective but never as good as what can be achieved by engineer teams working on transistor level to build a chip. I can only assume that Amd cpu tactics has worked it's way into their gpu department. It doesn't matter now though. You really have to think about it....Amd is now significantly behind Nvidia in very critical ways. While they continue to outspend Amd by larger and larger amounts every month. I believe most low hanging fruit is gone and from now on, big gains are only achieved by tens of thousands of small ones costing billions and billons just to inch forward. Maxwell leaped followed by pascal.
I am just not sure how Amd ever catches up...look at the efficiency of pascal...Nvidia is now further ahead than Intel was when they dropped the core 2. It seems unrealistic to me,...I mean, we are now seeing close to 2x efficiency pascal vs Polaris. That's not a gap, it's the grand Canyon. Look how long it took to just go from 28 to 14nm... and even then, I am not sure if Amd gained 2x efficiency