NVIDIA Pascal Thread

[Aristotelian]

I've been following the thread, but for a quick recap:

Nvidia may be making a more serious Pascal announcement shortly at a conference;
It's unlikely that any cards released during 2Q or early 2H 2016 will have HBM 2 (impossible) or GDDR5X (unlikely but possible);
Nvidia too will start with smaller dies, due to yield issues, and eventually go with a flagship-esque 600mm2 Pascal late in 2017 or early in 2018 (the real titan).

So the best we can hope for is the 1080 in 2016. Did I get this right? With the way the 980Ti stomped the 980 I don't see how the 1080 would be a compelling option, even someone for me on 580s in SLI. I want to buy 2, and put waterblocks on them, and go nuts. So I think I'm either going to see a baby titan + 1080 + 1070 in 2016, and then the real flagships next year, or just the 1080 and 1070 - which are, yes, much faster than what I have now, but somehow leave a bad taste. I'm interested in the Ti and Titan X successors, and it seems even AMD is waiting until next year to put out their Fury X successor.

Whoever puts out a real high end successor first gets my money. But I don't see any high end successors in 2016 - do you?

 

[Sweepr]

Clickbait website, so huge grain of salt required:

For the first time ever, we have spotted what is without argument the jewel in the next generation lineup that Nvidia will be presenting at GTC this April. The insurance value listed on the shipping manifest of the GPU indicates that this is the full uncut die (whether GP104 or GP100 remains unclear at this point). What we can say for sure is that we are looking at either the GTX 1080 or the GTX 1080 Ti (or equivalent) graphics card and that it is going to be unveiled at the upcoming Nvidia GPU Technology Conference Keynote in April.

http://wccftech.com/nvidia-flagship-gtx-1080-gtc-2016

 

[jpiniero]

I would not expect any consumer cards to be launched at GTC. Just Tesla.

 

[Sweepr]

Could be unveiled or at least mentioned at GTC.

For the first time ever, we have spotted what is without argument the jewel in the next generation lineup that Nvidia will be presenting at GTC this April. The insurance value listed on the shipping manifest of the GPU indicates that this is the full uncut die (whether GP104 or GP100 remains unclear at this point). What we can say for sure is that we are looking at either the GTX 1080 or the GTX 1080 Ti (or equivalent) graphics card and that it is going to be unveiled at the upcoming Nvidia GPU Technology Conference Keynote in April.

http://wccftech.com/nvidia-flagship-gtx-1080-gtc-2016

 

[DooKey]

I better get some Lite salt because you have to use lots of it when you are talking about wccftech. I hope I don't overdose.

 

[Sweepr]

Talking about salt...

The alleged pictures of NVIDIA’s upcoming GeForce GTX 1080 and GeForce GTX 1070 have just appeared on the web. The pictures which show the cooler shrouds of two upcoming Pascal based cards shows that NVIDIA might be using a revised version of the NVTTM cooler on their upcoming graphics cards.

However, since there’s no confirmation by NVIDIA on either the cooler or the naming scheme of the upcoming Pascal cards, we have to take this information with a grain salt.

http://wccftech.com/nvidia-pascal-geforce-gtx-1080-gtx-1070-cooler-leak/#ixzz44L7XK15F

 

[jpiniero]

HAHA, they really are shameless about being clickwhores. They just make up crap that's not believable because people will click even knowing it's fake.

 

[Maverick177]

That's a relatively big CNC machine, you just can't fake that.

 

[MrTeal]

HAHA, they really are shameless about being clickwhores. They just make up crap that's not believable because people will click even knowing it's fake.

I don't see how you can say that.

Whatever the case maybe, if the insurance value is anything to go by, we are looking at an absolute monster in terms of performance with the Geforce GTX 1080 GPU.

That logic is beyond reproach.

Speaking of, where else have we seen this cooler wccftech is claiming belongs to 1080/1070?

Oh yeah, same GPU linked above.

 

[Sweepr]

NVIDIA GeForce GTX 970 and GTX 980 to be replaced by Pascal with GDDR5X at Computex

A new generation of GPUs from AMD and Nvidia is imminent. For the latter it is about the architecture Pascal, the successor to today's Maxwell, to be manufactured on a 16 nanometer technology from TSMC. Both points are expected to lead to both higher performance, better energy efficiency.

Previously SweClockers has been able to uncover that Nvidia is aiming for a launch of Pascal for laptops at Computex 2016, which takes place May 31. Now it appears that the plans would include desktops as well, however, neither the highest or lower performance segment will be included which many first predicted.

According to information for SweClockers, Nvidia launches a primarily direct replacement for the Geforce GTX 970 and GTX 980. A source also claims that the first wave should include a replacement for the GTX 980 Ti, however, the performance in terms similar to the GTX Titan X will remain available a foreseeable future.

Another detail is that the graphics cards use memory standard GDDR5X, a further development of GDDR5. The first generation will clear rates of 10 and 12 Gbps, from 7 Gbps used in GTX 970 and GTX 980. Together with a 256-bit memory bus, this would provide bandwidth 320 and 384 Gb / s, in line with the GeForce GTX 980 Ti and AMD's Radeon R9 390X.

That Nvidia is aiming at Computex could be interpreted as their partner gets free reign to demonstrate self-designed solutions in place. Whether that happens remains a mystery, as several partner manufacturers have not yet received any chips by Nvidia or even had access to the full specifications. It could mean that graphics cards will initially only with Nvidia's reference design.

- Uncertainties about the naming
In previous leaks have been coming graphics cards appeared as the Geforce GTX and 1070 GTX 1080, which is in line with Nvidia's current naming scheme. Rumors suggest that the company abandon the GTX prefix and instead combine the Roman numeral X (10) with the Arabic which is the standard today - such as Geforce X80.

- Pascal for laptops and desktops launching at Computex
- GDDR5X (@ 256-bit? for 320-384 GB/s) - in line with GeForce GTX 980 Ti and Radeon R9 390X
- Not entry level nor highest performance (enthusiast) parts expected
- Direct replacement for Geforce GTX 970 & GTX 980 parts, which hints at GP104
- First wave of Pascal cards should include a replacement for Geforce GTX 980 Ti, similar performance to the Geforce Titan X - GP104 offering similar performance to GM200 at more affordable price points?
- Several partners haven't received chips by Nvidia, some don't even know the full specifications yet
- Possibly only NVIDIA refecence design available at launch
- To be named Geforce X instead of Geforce G 1000

www.sweclockers.com/nyhet/21927-nvi...80-ersatts-av-pascal-med-gddr5x-till-computex

Very similar info to the recent Bench-life article.

 

[Adored]

How can GDDR5X be ready in time for any of these cards when it only just started sampling this month?

 

[Sweepr]

How can GDDR5X be ready in time for any of these cards when it only just started sampling this month?

Perhaps they're planning a paper launch at Computex, with limited availability initially.

 

[DooKey]

How can GDDR5X be ready in time for any of these cards when it only just started sampling this month?

Risk production.

 

[Adored]

Perhaps they're planning a paper launch at Computex, with limited availability initially.

The issue I have with the whole GDDR5X rumour is that both IHV's would need to have developed GPUs with its availability in mind. It's such a game-changer at the midrange, but not one that either AMD or Nvidia could reasonably risk designing a chip for.

My point is that a Titan X/Fury X class of GPU with normal GDDR5 and 256-bit bus is likely to be hamstrung by bandwidth. Designing that GPU on the hope(?) that GDDR5X would be ready in time would be very foolish in that case, given the difficulties of the new node. The less risky (less reward also being true) option would be to design a smaller GPU that made the most of regular GDDR5 + compression.

If both companies did design smaller midrange GPUs around the assumption that they would only have regular GDDR5 then the extra bandwidth from GDDR5X is likely to be superflous. I guess it depends on cost, and whether it's worth waiting on...

 

[Kris194]

NVIDIA GeForce GTX 970 and GTX 980 to be replaced by Pascal with GDDR5X at Computex

www.sweclockers.com/nyhet/21927-nvidia-geforce-gtx-970-och-gtx-980-ersatts-av-pascal-med-gddr5x-till-computex

First, GDDR5x mass production starts in Summer, second, if GP104 would offer similiar performance to GTX Titan X it would mean that "GTX 1080" would be only 30% faster than GTX 980 GTC starts on Tuesday and then we will find out more about new graphics cards from Nvidia.

 

[Samwell]

The issue I have with the whole GDDR5X rumour is that both IHV's would need to have developed GPUs with its availability in mind. It's such a game-changer at the midrange, but not one that either AMD or Nvidia could reasonably risk designing a chip for.

My point is that a Titan X/Fury X class of GPU with normal GDDR5 and 256-bit bus is likely to be hamstrung by bandwidth. Designing that GPU on the hope(?) that GDDR5X would be ready in time would be very foolish in that case, given the difficulties of the new node. The less risky (less reward also being true) option would be to design a smaller GPU that made the most of regular GDDR5 + compression.

If both companies did design smaller midrange GPUs around the assumption that they would only have regular GDDR5 then the extra bandwidth from GDDR5X is likely to be superflous. I guess it depends on cost, and whether it's worth waiting on...

You overestimate the problems with GDDR5X. It's as if you were saying, how can Nvidia/AMD Design with HBM, when they don't know when it launches. But HBM is at least a totally new tech which might make problems. GDDR5X is very simple. It's the same as GDDR5 with small changes by doubling prefetch. You can assume the timeframe for availability of these chips pretty easy.

But nevertheless i can't believe in GDDR5X coming so early. Micron starts production in summer, it makes no sense that we get gpus with GDDR5X in summer. Assuming june for production we look at september launching gpus with gddr5x.

 

[Pinstripe]

I still believe a 384bit bus with 6GB GDDR5 RAM is more likely. GDDR5X is just not ready, and 256bit with GDDR5 is too much of a bottleneck. But we shall see.

 

[Erenhardt]

You overestimate the problems with GDDR5X. It's as if you were saying, how can Nvidia/AMD Design with HBM, when they don't know when it launches. But HBM is at least a totally new tech which might make problems. GDDR5X is very simple. It's the same as GDDR5 with small changes by doubling prefetch. You can assume the timeframe for availability of these chips pretty easy.

But nevertheless i can't believe in GDDR5X coming so early. Micron starts production in summer, it makes no sense that we get gpus with GDDR5X in summer. Assuming june for production we look at september launching gpus with gddr5x.

What? It is already out. Samsung is in mass production since January.

https://news.samsung.com/global/sam...on-newest-high-bandwidth-memory-hbm-interface

 

[Samwell]

What? It is already out. Samsung is in mass production since January.

https://news.samsung.com/global/sam...on-newest-high-bandwidth-memory-hbm-interface

He was talking about risc in development which was happening some years ago. I know that samsung is already producing HBM.

 

[Erenhardt]

He was talking about risc in development which was happening some years ago. I know that samsung is already producing HBM.

Sorry, I lost the track then...

 

[LTC8K6]

https://www.micron.com/about/blogs/2016/february/gddr5x-has-arrived

I think Micron is probably just about ready to go.

Also, they were already running at over 13gb/s according this February post.

 

[Adored]

You overestimate the problems with GDDR5X. It's as if you were saying, how can Nvidia/AMD Design with HBM, when they don't know when it launches. But HBM is at least a totally new tech which might make problems. GDDR5X is very simple. It's the same as GDDR5 with small changes by doubling prefetch. You can assume the timeframe for availability of these chips pretty easy.

HBM was a lot different because it was for the high-end, and for AMD's high-end which is traditionally very low volume anyway. In many respects it was a real test chip.

When designing a new GPU for the midrange on a new node, you can't risk the launch on stuff like GDDR5X availability 2 years later. Now granted you could use normal GDDR5 anyway and only lose performance due to bandwidth constraints, but right now every bit of mm2 is a prisoner.

But nevertheless i can't believe in GDDR5X coming so early. Micron starts production in summer, it makes no sense that we get gpus with GDDR5X in summer. Assuming june for production we look at september launching gpus with gddr5x.

It makes a lot more sense that both companies will refresh mid 2017 with bigger chips and GDDR5X. I'll be really surprised to see it this year though.

 

[Adored]

I still believe a 384bit bus with 6GB GDDR5 RAM is more likely. GDDR5X is just not ready, and 256bit with GDDR5 is too much of a bottleneck. But we shall see.

I believed that Nvidia would go that route but the latest rumour suggests not. In many ways I hope not as I'd really like to see a "fair" fight in the midrange again.

 

[Sweepr]

Could be fake but they're saying NVIDIA is about to phase out the Geforce GTX 980 Ti in favour of the new Pascal cards. Will start shipping the GPUs to manufacturers soon, still enough time to demo the new cards at Computex. Launch might happen 1-2 months later in July/August.

www.zolkorn.com/news/nvidia-starting-stop-delivery-out-gm200-for-gtx-980ti

 

[Headfoot]

I could see a paper launch / risk production midrange on GDDR5X happening, especially if Polaris is more competitive than we know (but that nVidia would have known due to their position in the market). Better to keep potential customers waiting for very limited stock to replenish than to buy a competitors product. This is essentially what Fiji did with HBM, I don't think its far fetched. I'm not ruling out simple high-clocked GDDR5 though either.

I'll be buying whichever single new 14/16 GPU is fastest while I can sell my 290s on the ethereum price bubble