Originally Posted by Hardware Hoshi
The 1060 9Gbps and the 1080 11Gbps are nothing more than a gap-filler. They will not be able to keep Vega off for very long. Nvidia has to bypass the time until Volta finall comes. If Q3 to Q4 2017 is remotely true, that can be another 6-8 months until any successor arrives. To see which card series comes out first is quite exciting. In the past Nvidia started with the x70 and x80 and brought the x60 a few months later, then the big chip like a Titan slightly behind.
The battle against Vega will not be about performance but also price. The 1080 had price cuts recently. Since it is a middle-class sized chip, Nvidia might be able to go down way more. Vega can never win a price-war with expensive HBM2, which is limited to 8GB and this HBCC nonsense. Doesn't sell well on a GPU wrapping.
On the manufacturing side Nvidia could use tricks like with the 1080 Ti. Deactivating a cluster for odd or uneven memory configurations should be easier to simulate beforehand. The Geforce cards can get higher bandwidth / throughput without touching the memory speeds themselve. Although I wonder if this 16 Gbps number is fixed or if there are further improvements in the near future we don't know about yet.
This is their reasoning behind releasing the 9Gbps and 11Gbps enhancements to these cards. The RX 580 is already out and ties, slightly wins or slightly loses (TPU has it 7% faster on average) versus the 1060. Finally, the RX 580 is actually a viable option whereas with the 480 most people said, "might as well get a 1060," because the performance difference was substantial enough to dismiss the 480 as an option (and many other variables). Whereas today, with the 580, the card is incredibly competitive in comparison to the 1060. Here though, NVIDIA doesn't want to leave that performance parity open and by simply increasing the memory speed on the revamp; the 1060 can be back to being 5-10% ahead of the 580. While you describe this as a gap-filler, it is still a swift move or play by NVIDIA. Sure it isn't a game changing move, but it is enough to carry them over to the next generation. They already made the sales they needed to make and with even a slight performance advantage, alongside brand recognition, it will allow NVIDIA to continue to make those impulse buying sales from potential customers. With the 1080 11 Gbps, again the same applies (the videos I saw of the comparisons between Vega and the 1080 showed them basically having identical performance, with maybe a slight advantage going to Vega). Most recent leaks have Volta at around Q3 2017 which is around September/October right in time for holiday. Recently, we have seen that release leak estimations and dates have been pretty accurate. That gives them half of a year to use Pascal to compete with the Vega stack. Now here's where things get problematic. If this is the case and we see a 2080 card dropped before 2018 or at least we have all the information available about it (like NVIDIA has been doing lately with their press releases) with availability around Q4 2017 / Q1 2018 that means that now, AMD has a whole slew of problems. This is because the 2080 is more than likely going to be Titan Xp performance for around $600. Now if the Vega card is also $600 because of having HBM2 and only is around 1080 performance, you can see how detrimental this could be; because as you say with Vega it is going to have HBM2, and today, if it only competes with the 1080, it will be difficult for them to get people to go over to the Red Team if the pricing isn't aggressive. AMD still has the issue of being the value option, or price to performance option. People only want to go AMD if it makes sense from a monetary standpoint. Aside from AMD enthusiasts, it is hard for them to convince potential buyers because their cards and products are still stigmatized. So here, while Pascal is still available, this can become an issue and then even worse if Volta is indeed a 2017 product.
There is a possibility that the card AMD demonstrated is not their fasted card, but based on math it seems unlikely:
An RX 580 has 2306 SPs, as far as I'm aware the maximum amount of SPs possible on GCN is 4096. That's a 78% increase in SPs. If we are to estimate that 78% increase in SPs means 78% linear increase in performance; we arrive at 12.7% faster than a GTX 1080 @ 1440p (we saw with the Fury X that HBM didn't really affect performance as much as we thought it would, when compared to the 980 Ti. Therefore, I'm not adding more performance percentage points for HBM 2 for that exact reason):
This unfortunately means, that AMD doesn't have a competitor against the 1080 Ti / Titan Xp which again means they aren't actually competing in the high end segment for yet another generation. Translating to, again, NVIDIA doing anything they freaking want with prices for the rest of Pascal's life and for Volta's lifespan whether that means for better or for worse (undercutting AMD or price gouging).
Now that we see where Vega will most likely end up, even though the 1080 is a mid-sized chip, it still is going to easily compete with full Vega, especially after adding the 11 Gbps memory. Hopefully AMD has some magic sauce that allows them to price Vega @ $550 or something along those lines and retain a 5-8% performance advantage over the 1080 @ $500 where then it could be a pretty hopeful battle. However, again, NVIDIA will not leave that performance delta open for too long, they must always retain the performance crown. Bringing us back to our Volta release estimations in Q3/Q4 2017 as even more probable. As a 2080 that is as fast as a Titan Xp for say $600-650 means that the 2080 will be around 30% faster than the best Vega SKU for $550-600, while only being 10% more expensive. I mean there is a chance that AMD blows us away and can offer the biggest Vega SKU for 1070 prices, but then that leaves no price available for cut down products that are supposed to compete with the 1070 instead of the 1080. I mean the best versions of the RX 580 are already $300. So what would a Vega card that is a 1070 and 1080 competitor cost. Doesn't leave much room for price wars.
They will most likely manipulate the amount of VRAM additionally, as I agree with that and they could utilize more memory compression technology, but of course that isn't as good as simply increasing the memory bandwidth by increasing the speed. Of course there are future improvements with the 16 Gbps; GDDR5 started @ 900MHz and today it is @ 8000-9000MHz. Surely there is advancements to be made with the technology. Which brings us back to discussions we had around 2 years ago, is HBM going to survive or ever going to even be necessary for consumer GPUs? Or is this going to be a case where Radeon Technologies Group consistently uses HBM technology in their GPUs and NVIDIA goes a different route by sticking to modern day GDDR technology; whether GDDR6 or G5X.Edited by BiG StroOnZ - 4/25/17 at 1:11pm