• Hey Guest. Check out your NeoGAF Wrapped 2025 results here!

RUMOR: Is this NVIDIA TITAN Volta?

500mm2 HBM2 equipped chip beating only 314mm2 1080 is a super pessimistic expectation.

It's mostly based on the showcases AMD has put on. They have shown off a "top of the line" Vega card playing (I believe) Doom at 4K at 60fps, and we know from dozens of third party benchmarks that a 1080 can almost hit that mark with a 1080ti passing it and Titan Xp going well above it. Not 100% certain on the game, but I know it was a specific title from the last 9-12 months at 4k/60fps, and there are already cards on the market that can manage that.

In all likelihood the Vega cards will compete with the 1070 and 1080, and trade blows with the 1080ti at the high end. The problem for AMD is, I see no way HBM2-equipped cards make them more money per card than Nvidia makes off the 1070, 1080, and 1080ti. Not to mention those 3 cards will have been out for months and even years before the Vega cards hit shelves--and the first batch is a very limited release.

500mm2 is probably due HBM2 not GPU units.

There's a Reddit post that estimates the size since AMD has not released any solid numbers. The 525mm^2 is Computerbase.de's estimate based on the documented size of the HBM2 stacks courtesy of JEDEC. So it's somewhere between 475 and 525 by most estimates, which still makes it smaller than Hawaii, and that's just for the GPU die. Including the HBM2 stacks it's probably closer to 600mm^2. AMD cards tend to be less space efficient than Nvidia cards because they feature a lot more compute hardware on there--hence why the 7970/290X/390X were all much better at Bitcoin mining. Nvidia striped a lot of that away going from Fermi to Kepler, and focused instead on streamlined gaming performance since it was a lot easier to market and cut cost substantially.

AMD is basically chasing the same markets...but whether they can execute their plans is anyone's guess. Almost all of the semi-conductors made bad bets once in awhile (intel with Netburst architecture and choosing Rambus as partner in P4 days, Nvidia with mobile via Tegra) but AMD is famous for making so many bad bets that i always have large reservation for them (the bulldozer being one of those duds)

AMD has been making efforts to expand. The Frontier line of Vega cards that they announced is the first major push outside of the traditional GPU markets that they work within--a challenger to Nvidia's Tesla cards which make them a lot of money. AMD has also announced some deals for datacenters as far as I remember.
 
Are we expecting these cards to be a massive leap over the current 1080Ti's?

Usually the xx70 variant of the next generation is competitive with the Ti from the last generation. The 1070 was on par with--and often beat--the 980 Ti, the 970 was slightly below the 780 Ti and third party variants were on par with it. The thing is, when Nvidia does release Volta (I'm thinking Q3 this year maybe even Q4) they will launch with the xx80 variant first, then after a few weeks the xx70, then after a few months the xx60, and then after several months the Titan Volta (should just call it Titan V). So if this is the Nvidia Titan Volta, it won't be for sale for about a year, maybe longer.
 
That goes against all historical knowledge where the Titan launches first, then the 80/70, then the 50/60, and finally the 80 Ti.

Since we still don't really know when or if Nvidia ever plans on transitioning to HBM for consumer level cards, and quite frankly they haven't needed HBM because for them GDDR5X has been more than adequate, there may be a very significant delay before consumer Volta appears. There's no evidence that Vega is threatening Nvidia's current top shelf at all so they will be in no hurry whatsoever to release consumer Volta when they could be focusing their attention on their exploding datacenter and deep learning/AI markets.

Also Switch is juicing up their Gaming revenues nicely so they aren't under nearly as much pressure to ship more Gaming products just to maintain their quarterly numbers. A second year of Maxwell on the market uncontested didn't hurt their growth much if at all, a second year of Pascal on the market will likely be the same.
 
Nvidia GTX 1080: Mid-May 2016
Nvidia GTX 1070: Late-June 2016
Nvidia GTX 1060: Mid-July 2016
Nvidia Titan X Pascal: Mid-August 2016 (3 months after 1080)
Nvidia GTX 1080 Ti: March 2017

Assuming they keep the same structure as they did with Pascal for Volta, we won't see a Titan variant for quite a while. Even if we look at the Maxwell releases:

Nvidia GTX 980: Mid-September 2014
Nvidia GTX 970: Mid-September 2014
Nvidia GTX 960: Late-January 2015
Nvidia Titan X: Mid-May 2015 (9 months after 980)
Nvidia GTX 980 Ti: Early-June 2015

Titan's always come out before the Ti variant but after the 80/70/60.
 
Are you expecting 104 this year still?

Yeah, I'm thinking Sep-Nov for GV104 launch, GV102 as a new Titan in Dec-Feb and GV106 as a GP104 replacement somewhere along. GV104 could probably get away with G5X@11Gbps (or maybe Micron will be able to hit 12-14 Gbps as they promised for GDDR5X initially) but for GV102 they'll probably need GDDR6 @ 14Gbps.
 
Yeah, I'm thinking Sep-Nov for GV104 launch, GV102 as a new Titan in Dec-Feb and GV106 as a GP104 replacement somewhere along. GV104 could probably get away with G5X@11Gbps (or maybe Micron will be able to hit 12-14 Gbps as they promised for GDDR5X initially) but for GV102 they'll probably need GDDR6 @ 14Gbps.

Haven't they said in the past that Volta would be HBM2?
 
Haven't they said in the past that Volta would be HBM2?

Volta already is using HBM2. V100 was revealed earlier this month.
HBM2 doesn't make sense for Geforce when more affordable alternatives exist. Titan XP is already achieving peak memory bandwidth of 547 GB/s using GDDR5X. If the technology can continue to produce the performance needed without drawing too much power, then Nvidia will continue to use it.
 
Haven't they said in the past that Volta would be HBM2?

No reason to use HBM2 for consumer cards. Vega's memory bandwidth with HBM2 will be in the same ballpark as 1080Ti/TitanXp's with GDDR5X and a 384 bit bus for example while HBM2 is significantly more expensive in both production and usage.

Stacked DRAM makes sense only in highly bandwidth critical environments and these are covered by GP100 and GV100 which do use HBM2, and in four stacks configuration. Gaming markets are well covered by G5X and G6 for now.
 
No one even knows why AMD has such a hard-on for the expensive, low-availability HBM instead of just using GDDR5X in Vega. No lessons learned from the Fury X.

Nvidia GTX 1080: Mid-May 2016
Nvidia GTX 1070: Late-June 2016
Nvidia GTX 1060: Mid-July 2016
Nvidia Titan X Pascal: Mid-August 2016 (3 months after 1080)
Nvidia GTX 1080 Ti: March 2017

Assuming they keep the same structure as they did with Pascal for Volta, we won't see a Titan variant for quite a while. Even if we look at the Maxwell releases:

Nvidia GTX 980: Mid-September 2014
Nvidia GTX 970: Mid-September 2014
Nvidia GTX 960: Late-January 2015
Nvidia Titan X: Mid-May 2015 (9 months after 980)
Nvidia GTX 980 Ti: Early-June 2015

Titan's always come out before the Ti variant but after the 80/70/60.

Ah, I see. I was mistaken on the order when Titans are released.
 
No one even knows why AMD has such a hard-on for the expensive, low-availability HBM instead of just using GDDR5X in Vega. No lessons learned from the Fury X.

GDDR5X seems to be pretty much NV exclusive as no one but NV is using it. AMD do have plans to use GDDR6 though as soon as it will become available:

D8vc.png


As for why AMD is so locked to HBM - it's also not a secret as GCN is very bad in perf/watt and HBM allows them to push some or the final board power towards the GPU instead of VRAM meaning that they are likely able to provide higher performance with HBM by spending more power on driving the GPU.

NV is so far ahead in perf/watt that the additional available power provided by the HBM compared to G5/G5X/G6 is not worth it to them when the prices of the said memory solutions are taken into account.

Basically, HBM allow AMD to better compete against NV in perf/watt and right now this is the main reason why they opt for HBM over GDDR.
 
Yeah, I'm thinking Sep-Nov for GV104 launch, GV102 as a new Titan in Dec-Feb and GV106 as a GP104 replacement somewhere along. GV104 could probably get away with G5X@11Gbps (or maybe Micron will be able to hit 12-14 Gbps as they promised for GDDR5X initially) but for GV102 they'll probably need GDDR6 @ 14Gbps.

I would not expect any consumer cards this year and with GDDR6 expected to ship in volume early next year, I would position GV104 in March-May 2018. Based on past releases, I expect GV102 (GDDR6) and GV106 (with GDDR5x) in the summer of 2018. I would love to be wrong.
 
correct me if this is not correct regarding Titan cards.

Kepler generation
*GTX Titan
*GTX Titan Black
*GTX Titan Z (dual GPU)

Maxwell generation
*GTX Titan X

Pascal generation
*GTX Titan X
*GTX Titan Xp

Volta generation
?????????
 
Top Bottom