So when do the next process node video cards come out?

[Fallen Kell]

Now that both Nvidia and AMD have essentially released their last 28nm cards, what and when will the next cards really come out? TSMC 16nm should be out in a few months at this point, but Samsung has already been making low power/performance chips for months at 14nm. What are the odds that Nvidia and/or AMD finally get fed up with TSMC enough to move to Samsung or similar?

 

[happy medium]

I'm guessing Nvidia's Pascal to release Q3 next year.
And they will use TSMC again.

AMD's Fury 2 a few months after.

 

[alcoholbob]

They are the last new cards before 16FF+ but in 6-9 months there should be a 28nm refresh of Titan X/Fury X likely with higher clocks.

 

[beginner99]

I'm pretty sure firts 14/16 nm cards will be small ones for mobile. Till we actually get new high-end GPUs, it will probably be 2017 and they will almost certainly up the price level even more. So I don't think it will be all that great.

Why will it take that long? Yields and pricing. Apple and the likes can and will pay more and hence get first access. Then the process must first mature to allow >400 mm^2 dies for large GPUs. Yes, a 300 mm^2 one would probably already beat 980Ti/Fury but would not be a huge jump worth upgrading to or able to do 4K.

 

[ShintaiDK]

I would guess Q4 2016/Q1 2017.

Small dies and high prices.

 

[Silverforce11]

I would guess Q4 2016/Q1 2017.

Small dies and high prices.

I agree with this prediction.

Both NV/AMD are going to a new unproven immature node. Very risky to go with a new uarch big GPU with also new unproven HBM2 vram.

GloFo/Samsung has 2 separate 14nm ff production, one for low power (SoC/Mobiles) and the other for performance. This has been in the public record for a long time now.

 

[Fallen Kell]

So everyone really believes it will be end 2016 or early 2017 even though both AMD and Nvidia had been working on different designs for 22nm which never were released? I would expect that those 22nm designs would simply be scaled down to the 16nm process, with some minor tweaks (possibly add more stream processors, etc., in unit/groups that make sense to make use of the additional space savings from the 16nm process).

 

[Alatar]

Rumor has it that big Pascal has already taped out recently.

I'm gonna go against the common "Q4 2016" sentiment and say that I think that's pessimistic. Imo we should be looking at H1 2016 for the first design making an appearance. What type of chip and which market segments is another question entirely though.

 

[Grooveriding]

At least a year, maybe longer.

 

[Kippa]

Some are predicting Q4 in 2016? That is a long time to wait for an upgrade. I might consider getting a new GPU if it is going to be that long. I don't mind waiting till Q1 of 2016 but not all the way to Q4 2016.

 

[Elixer]

Some are predicting Q4 in 2016? That is a long time to wait for an upgrade. I might consider getting a new GPU if it is going to be that long. I don't mind waiting till Q1 of 2016 but not all the way to Q4 2016.

Could even be longer... it took a heck of a long time to get to where we are at now.
28nm had tons of issues as well.
Small chips will have a much easier time with the new node, but, chips like GPUs & CPUs are enormously more complex, and require much more resources.

If I was to bet, I would bet AMD having HBM2 Furys before Nvidia has their first HBM product.

 

[Despoiler]

Rumor has it that big Pascal has already taped out recently.

A not very credible rumor considering it was just some random poster on a forum and no foundry has confirmed anything. Foundries tend to like to brag about their milestones.

We really have 2 technologies that have to line up to get Pascal out the door. 14/16nm and HBM gen2. My prediction is that both will be late if Nvidia stays with TSMC. Q3 2016 earliest. Q4 2016 more likely.

 

[Kippa]

Could we be seeing an AMD 28nm with HBM2 card before the 14nm cards come out? I wouldn't mind waiting a few months if there was a 28nm HBM2 card on the table.

 

[Qwertilot]

HBM 2 isn't due for a little while it seems, so you'd be more likely to see an early die shrink with HBM1.

Say AMD deciding they're fed up with not having any mid range cards and plain up die shrinking fury. HBM1/4GB would then be roughly OK as a mid range thing.

Must be decent chances of seeing an early 750ti ish equivalent etc from NV. We'll find out

 

[krumme]

Some are predicting Q4 in 2016? That is a long time to wait for an upgrade. I might consider getting a new GPU if it is going to be that long. I don't mind waiting till Q1 of 2016 but not all the way to Q4 2016.

Its not like the new 14nm gpu will be far faster than the 600mm2 we have today. Nv says 14/16nm is more expensive per transistor than 28nm.
We have to remember the world has changed and mobile is eating the good stuff. Say end h1 2016 there will be zen and samsung and qcom own new custom cores to compete for the good 14nm stuff on top of the usual arm a72 whatever. its stiff competition.

So anyway if you can get 350mm2 14 nm cards they might be very impressive but damn expensive anyway.
I can see one reason to wait and that is for 28nm to be less expensive. So a 390x and 970 will be 230 usd cards. 3 months perhaps but the new highend 600m cards will stay high especially for nv.

 

[krumme]

Do we have any confirmation on if amd new 28nm fijj cards is tsmc or gf?

 

[Elixer]

Maybe early Q2 '16 Hynix will start shipping HBM 2, so, realistically, Q3/Q4 '16 you will see products with HBM 2.

 

[Samwell]

As GP100 has taped out already i expect it to come in Q2 2016 for professional market. First consumer gpu will be a successor to GM107 without HBM i think. They need a gpu for the middle mobile segment, which has better DX12 support than Maxwell V1. This should come pretty early, End of Q1 2016 maybe if we're lucky.

AMDs was talking so much about HBM, that i don't expect Gpus without HBM made by AMD. Questions is just, when will HBM2 be ready for mass production?

@Krumme
In the Anandtech article TSMC was mentioned, also read the same somewhere else. It seems no GF again.

 

[JDG1980]

AMD desperately needs FinFET GPUs; their current generation consists largely of uncompetitive rebrands. Fury looks to be decent, but won't be available at the most popular price points. And AMD has nothing worthwhile for the laptop market on 28nm.

Furthermore, AMD has always been at the forefront of adopting new GPU technology; they were the first to adopt GDDR5, the first to 28nm (beating Nvidia by 3 months), and now the first to ship HBM. For this reason, I expect that they will come out with FinFET GPUs before Nvidia.

I think AMD is going to aim to bring their first GPUs on the new node to market no later than 1 year from now. If they can get them sooner, they will. I wouldn't be surprised to see them follow the strategy that Nvidia used with Maxwell, and lead off Arctic Islands with a smaller, low-end chip.

 

[Silverforce11]

Maybe early Q2 '16 Hynix will start shipping HBM 2, so, realistically, Q3/Q4 '16 you will see products with HBM 2.

Per Hynix, HBM2 schedule for production in Q2 2016. Not volume production until Q3 2016.

How long to ship and integrate/stack on interposer with the GPU? Probably a quarter to sort out issues.

Do you think NV or AMD would go big-die first on an unproven node with brand new uarchs with brand new vram? Would be epic-fail waiting to happen.

You could be hopeful and expect a mid-range Pascal first that offers ~Titan X performance at much reduced power usage. But the big Pascal would need to be for Tesla/Quadros first. Then there's a question of yields.

Pretty much the stars have to align for us to get a consumer big GCN/Pascal in early Q1 2017.

 

[Stuka87]

Do we have any confirmation on if amd new 28nm fijj cards is tsmc or gf?

AMD has not said. But we should know tomorrow when reviews come out.

 

[moonbogg]

I bet its 6-9 months away.

 

[Elixer]

Per Hynix, HBM2 schedule for production in Q2 2016. Not volume production until Q3 2016.

How long to ship and integrate/stack on interposer with the GPU? Probably a quarter to sort out issues.

Do you think NV or AMD would go big-die first on an unproven node with brand new uarchs with brand new vram? Would be epic-fail waiting to happen.

I really think AMD already has Fury with HBM2 in the labs, and as soon as Hynix can ramp up production I can see them having products out.
So, think of the current Fury with 8/16/32GB available.
Since AMD will have experience with the new node with Zen, then, then can do a new node for Fury 2, which will be around Q1 '17/Q2 '17.

For Nvidia, it is much more complex, since they are jumping to a new node + new RAM tech, which is why it seems they will have something late Q4 '16/ Q1 '17 for Pascal (if all goes well), and Q1 '18 for Volta assuming everything goes OK with Pascal.

That just means we are stuck with the current round(feature wise) of cards from both camps for about a year.

 

[JDG1980]

Per Hynix, HBM2 schedule for production in Q2 2016. Not volume production until Q3 2016.

How long to ship and integrate/stack on interposer with the GPU? Probably a quarter to sort out issues.

Do you think NV or AMD would go big-die first on an unproven node with brand new uarchs with brand new vram? Would be epic-fail waiting to happen.

For those reasons, I wouldn't be surprised if AMD's first 14nm FinFET product is a low-end chip that still uses GDDR5. (Even though they probably want to push HBM further down the stack, it won't be economical to do so all the way to the bottom until the technology fully matures. That's why there will be no HBM APUs until 2017.)

AMD really needs a replacement for the aging Pitcairn chip. If we assume that 14nm FinFET has roughly twice the transistor density of 28nm, then AMD should be able to pack 1280 shaders into a chip about the size of Cape Verde (a bit over 120 mm^2). With GCN 1.2 color compression, a 128-bit GDDR5 memory bus should suffice. Even if costs per transistor are a bit higher than on 28nm, such a chip wouldn't be much more expensive to manufacture than Pitcairn, and with a narrower bus and lower power requirements it would be cheaper to build boards, so the street price could still be around $150 for AIB cards. This would be a strong competitor to GM107, beating it in performance by 20%-30% with similar levels of power consumption. This might help AMD get some much-needed laptop design wins.

Q1 2016 might be too aggressive a timeline for this, but AMD should bring it out no later than Q2.

 

[Elixer]

We are developing 22nm FD-SOI process technology right now.” The plan is to qualify the process by early 2016 and start volume production by the end of 2016. Asked about what the company’s 22nm FD-SOI process entails, Teepe explained, “We are using STMicroelectronics’ 14nm FD-SOI in our frontend, while using 28nm FD-SOI in the backend.”
GlobalFoundries appears to believe that their initial FD-SOI customers will be designers of big volume SoCs. While those customers need to push the performance of their SoCs, what they really need is a “good price,” said Teepe. The promise of GlobalFoundries’ 22nm FD-SOI solution, he concluded, is to offer his customers “almost 14nm FinFET performance at almost 28nm cost.”

Gloflo will have a talk about this stuff...Wednesday, June 24 at 10 am, PDT.
Who needs to see Fury reviews, when you can watch this!?

http://ci43.actonsoftware.com/acton/form/6140/0008:d-0001/0/index.htm?sid=pC58tqysP

GlobalFoundries appears to believe that their initial FD-SOI customers will be designers of big volume SoCs. While those customers need to push the performance of their SoCs, what they really need is a “good price,” said Teepe. The promise of GlobalFoundries’ 22nm FD-SOI solution, he concluded, is to offer his customers “almost 14nm FinFET performance at almost 28nm cost.”

Late 2016? Delays...delays...delays.