GPU Some predictions for 2017

[IntoxicatedPuma]

I was a little bit slow today at work and was looking through a database of video cards, and more importantly the release dates. I noticed something:

Nvidia releases a new architecture about every 2 years fairly consistently.

200 (Tesla) - June 2008
400 (Fermi) Mar 2010
600 (Kepler) Mar 2012
750 (Maxwell) Feb 2014
10 (Pascal) May 2016

So obviously Pascal won't be replaced until around May 2018. However the release of mid series is not so much like clockwork. It seems Nvidia will release the refresh depending on the strength of competition. For example, AMD was strong in 2010-2011 so Nvidia released the GTX 580 only 8 months after the 480. That card lasted almost 1.5 years until the 680 came out. The 680 lasted around 1 year before the 780, and 780 about 1.5 years till the 980, and 980 over 1.5 years till 1080. During this time Nvidia's position was strengthening and Nvidia was weak.

So not to state the obvious but, if AMD's Vega is truly strong, I would expect that Nvidia won't release a GTX 1080 ti, but will start fresh with a new series and release the GTX 1180 by the summer.

I also have a feeling that Nvidia may not release an 11 series, partly because the name sounds weird but also for another reason: Nvidia hasn't released a sub 50 series card in almost 3 years aside from the GT 710. The naming convention isn't so great anymore because of this, i mean it does make their cards sound good because even the lowest end GTX is a "50" out of "100" right?

So two predictions:

• Nvidia will drop the current naming convention and come up with something new since Mobile and desktop cards are the same, and desktop cards don't have lower level configurations anymore.
• Nvidia will release the new generation (not new architecture) by summer time - except if Vega sucks, then I predict Nvidia will wait until October, and a 1080 ti will come out before May.
• And a third for the heck of it: Nvidia will release or announce an SOC for Windows powered laptops by the end of summer 2017. Maybe a Denver quad core CPU + Pascal graphics?

If i'm wrong on at least two of these, I'll double the money I donate for the site this year as an apology for posting stupid crap.

Do you guys have any predictions for this year?

 

[|||]

The next generation for Nvidia is Volta. For what is being rumored, it will have HBM2 on the high end and GDDR6 on the lower end cards. HBM2 has the issue that there is only limited availability and AMD has priority access to volume, due to their participation in the development of HBM memory. Micron just announced that they are pushing forward the production of GDDR6 to the end of the this year, making it look like they are readying it for cards hitting the market next year.

As for the 1080 Ti, the original Titan was launched Feb 21, 2013 followed by the 780 Ti on Nov 7 2013 (7 1/2 months after) and the Titan X was launched March 17, 2015 followed by the 980 Ti on June 1, 2015 (2 1/2 months after). Since the Titan X (Pascal) was launched August 2nd last year, we're still in a reasonable timeframe. Besides, Nvidia would want to reap as much in terms of premiums it's getting from the sales of the Titan X, especially now that it is being done exclusively with direct sales. The 1080 Ti will go through traditional sales channels and will probably be dropped before Vega gets released, so the early estimates I've seen are PAX East, but I wouldn't be surprised if it gets pushed to April, as Vega is supposed to be out in Q2 (talks have been May).

 

[Phuncz]

I can certainly agree with your analysis and prediction. I was expecting them to ditch the naming this generation. Nvidia has made indications they're not interested in making entry-level GPUs any more since they compete with iGPUs too much. Since OEMs occupy that market mainly and minimize cost religiously, it makes sense that an iGPU is much cheaper than a second GPU in almost any case.

I don't expect Nvidia to release much before Vega launches. I'd be amazed if Vega is even that much better than say Pascal +20% performance, which Nvidia should indeed be able to counter easily. I'd wish it was amazing, but I'm not expecting a $300 GTX 1080 equivalent.
I'm expecting the "RX490" to be around GTX 1070 performance for $350 with an "RX Fury" coming in at GTX 1080 performance for $450 and an "RX Fury X" above GTX 1080 performance at $600.

At that point I expect Nvidia to "unleash hell", introduce a new series within the month, all one-upping AMD once again. Atleast that's what I think will happen if Vega is just another evolution and not a revolution. Their marketing campaign seems to be hyping already known facts as some kind of secret news like Volta not coming this year (duh) and 4K @ Doom capped at 60fps being some kind of impossible mountain to climb.

tl;dr
I'd love to see Vega to be a true surprise in performance per dollar, but they've been doing the same game for a few years now, hyping and just being good but not great.

 

[EdZ]

I expect that in response to Vega, we'll see a 1080Ti as a 'full fat' (or close enough) GP102-based card arrive as the range-topper, and the 1080/1070/etc move 'down' a price point (if necessary) to match Vega. I suspect Vega will launch pretty close in price/perf to Pascal anyway: as seems to be the pattern, sitting a bit behind in everything except Vulkan/DX12, and trading blows for DX12/Vulkan depending on how any given game has optimised their engine.

Volta is probably quite a way off, and IIRC has already been slated to turn up in the HPC role quite some time before consumers get it. Summit will be gobbling up ~27,000 dies at full rollout, with Sierra grabbing another ~8,000. That's a lot of production volume dedicated to dies on a new architecture (and big ones too). Depending on whether Nvidia can find a HPC market for the binned dies from that production run, the 'off cuts' from that production push may end up in other supercomputers, or possibly make their way as a 'halo' card (I'd put this as more likely if yields are very good and there are few rejects). A year of HBM2 volume production for HPC may push prices down enough that it can make its way to consumer high/upper-mid range cards, but no game as of yet has been anywhere close to memory bandwidth limited, so it may not be worth bothering with anything other than GDDR5X or GDDR6.

On the normal consumer side, I could see a 'Pascal 2.0' refresh slot in before Vega for the consumer side of things unless Vega really pushes performance forward. Probably no change in process, so more in the way of optimisation (Similar to Maxwell => Pascal) and added features (likely more VR stuff. native non-rectilinear processing would be awesome, but unlikely to be possible with just a minor update).