[Ars] AMD confirms high-end Polaris GPU will be released in 2016

[Ma_Deuce]

Nobody is saying anything like that.

Suggesting a company should lose money to gain market share. Giving away your products.

Potato. Potato.

 

[Glo.]

http://videocardz.com/58237/amds-project-f-is-232mm2-discrete-gpu-made-in-14lpp-process

232 mm2 It is the bigger one, for sure.

 

[jpiniero]

232 mm2 does sound like the size of Polaris 11 and it does match up with what was said about Polaris 10. So Polaris 11 would be price and performance comparable to Grenada, maybe a tad faster.

I still don't know why people are allowed to post confidential information on Linkedin so blatantly.

 

[Glo.]

To be honest 232 is perfectly in line with what I have been writing about die size shrink between TSMC28/GloFo/Samsung14NMLPP. Shrinking die of Fury X on 14 nm, along with a little bit more dense architecture itself would get us into similar size. Not that It is the reality here, but it is a possibility.

 

[iiiankiii]

232mm2.... There goes my hope of having a Titan X killer in 2016.

 

[tviceman]

To be honest 232 is perfectly in line with what I have been writing about die size shrink between TSMC28/GloFo/Samsung14NMLPP. Shrinking die of Fury X on 14 nm, along with a little bit more dense architecture itself would get us into similar size. Not that It is the reality here, but it is a possibility.

Eh, close but not quite. Looking at the last big release AMD made on 40nm (Cayman), that die had a density of ~6.8m transistor/mm2. First gen 28nm Tahiti had a density of ~12.2m transistors/mm2, about an 80% increase in density. All the press info we read about transistor density and perf/w on specific nodes is usually above average scenarios, and regardless, is very architecture dependent. Anyways, if AMD gets similar scaling improvements moving from their latest 28nm product to first gen FF as they did when going from their best 40nm to their first 28nm, then a chip with 8.9 billion transistors would be right around 330 mm2. I think with 232mm2, we'll be looking at stock ~R9 390 performance, or about double Pitcairn.

 

[Arachnotronic]

232mm2.... There goes my hope of having a Titan X killer in 2016.

I dunno, FinFET can drive frequencies up much higher than on 28nm.

 

[raghu78]

I think with 232mm2, we'll be looking at stock ~R9 390 performance, or about double Pitcairn.

232 sq mm on 14LPP with R9 390 performance would be pathetic. :thumbsdown:

 

[AtenRa]

If Polaris 11 is 232mm2, Polaris 10 is then exactly half the size at 110-120mm2.

Polaris 11 at 232mm2 with HBM2 could be very close to Fiji.

 

[tviceman]

232 sq mm on 14LPP with R9 390 performance would be pathetic. :thumbsdown:

That would be double Pitcairn's performance at nearly the same die size as Pitcairn. What is pathetic about that? Tahiti was 15% smaller than Cayman and didn't come anywhere remotely close to doubling performance at the time of it's release. You must be expecting miracles.

 

[Arachnotronic]

That would be double Pitcairn's performance at nearly the same die size as Pitcairn. What is pathetic about that? Tahiti was 15% smaller than Cayman and didn't come anywhere remotely close to doubling performance at the time of it's release. You must be expecting miracles.

raghu78 has a history of making shall we say "optimistic" predictions about future AMD products

 

[itsmydamnation]

raghu78 has a history of making shall we say "optimistic" predictions about future AMD products

Remember there will also be utilization per unit improvements as well as higher clocks with A*B neither A or B have to be massive to get a good deal of C.

But 230nm sounds about right ( if a little on the low side) for a new process, rv770 256nm, cypress 330nm, Tahiti 350nm. given HBM2 throughput improvements and more memory compression technology we might see a narrower then expected HBM interface(512bits) giving more die space to ALU:MTU:ROP, would mean a smaller imposer as well.

 

[Arachnotronic]

Remember there will also be utilization per unit improvements as well as higher clocks with A*B neither A or B have to be massive to get a good deal of C.

But 230nm sounds about right ( if a little on the low side) for a new process, rv770 256nm, cypress 330nm, Tahiti 350nm. given HBM2 throughput improvements and more memory compression technology we might see a narrower then expected HBM interface(512bits) giving more die space to ALU:MTU:ROP, would mean a smaller imposer as well.

All I care about is getting a pair of nice new GPUs for one of my systems that's currently sitting GPU-less, stuck with an integrated GPU after I ripped out its Titan X's and turned them into cash money in anticipation of 14/16nm GPUs. Hope either AMD or NVIDIA can deliver on something worth my money.

 

[MrTeal]

I wouldn't even be so sure a 232mm² die would get HBM, but if it does I could easily see it being a two stack solution. Given the larger size of HBM2 relative to HBM, a four stack solution would have an interposer basically as large as the one on Fiji.

 

[Tuna-Fish]

given HBM2 throughput improvements and more memory compression technology we might see a narrower then expected HBM interface(512bits)

HBM comes in 1024 bit increments.

 

[HurleyBird]

I wouldn't even be so sure a 232mm² die would get HBM, but if it does I could easily see it being a two stack solution. Given the larger size of HBM2 relative to HBM, a four stack solution would have an interposer basically as large as the one on Fiji.

I would be incredibly surprised if you *didn't* see a ~200mm2 (or even smaller) die with HBM. That's about as large of a chip as you're going to want to put in a laptop, and HBM has massive power benefits over conventional GDDR. A ~200mm2 HBM-equipped laptop chip could command incredibly high margins (especially if only one company has the foresight to produce something like it), similar to what Intel enjoys with EDRAM-equipped Iris Pro branded IGPUs.

 

[raghu78]

That would be double Pitcairn's performance at nearly the same die size as Pitcairn. What is pathetic about that? Tahiti was 15% smaller than Cayman and didn't come anywhere remotely close to doubling performance at the time of it's release. You must be expecting miracles.

It would be pathetic because it would mean AMD did nothing to improve architectural efficiency beyond what the FINFET process node shrink brings. It would mean AMD have done nothing from 2012 to 2016 to improve perf/sp perf/watt, perf/sq mm purely from an architectural standpoint. It would also mean GP106 will murder it. So if AMD are so dumb as to not anticipate Nvidia improving Maxwell further and try to compete with that then they might as well shut down their business.

raghu78 has a history of making shall we say "optimistic" predictions about future AMD products

If AMD have been twiddling their thumbs for 2 GPU generations (4 years) then they deserve to go out of business.

 

[itsmydamnation]

HBM comes in 1024 bit increments.

sry meant 2048 not 4096 like fury, no idea why i typed 512, off with some kind of fairy i suppose D:

 

[Arachnotronic]

If AMD have been twiddling their thumbs for 2 GPU generations (4 years) then they deserve to go out of business.

IMHO, they are betting their company on Zen, GPU is secondary priority. You can see it in their financial analyst day when they are pumping data center/Zen while saying that dGPU is a slow-growing opportunity.

NVIDIA is 100%, all-in on GPUs, making them a very dangerous opponent. The people who are sitting there claiming that NV has its thumb up its rear-end just waiting to get railed by AMD/Polaris just don't have a clue, IMO.

 

[raghu78]

IMHO, they are betting their company on Zen, GPU is secondary priority. You can see it in their financial analyst day when they are pumping data center/Zen while saying that dGPU is a slow-growing opportunity.

NVIDIA is 100%, all-in on GPUs, making them a very dangerous opponent. The people who are sitting there claiming that NV has its thumb up its rear-end just waiting to get railed by AMD/Polaris just don't have a clue, IMO.

AMD needs Zen to succeed for their long term survival. But that does not mean AMD gave up on the GPU business. What they did was to prioritize their budget on the FINFET products - Zen and Polaris. Thats the reason for AMD not having a new GPU generation with significant improvements to compete with the Maxwell generation. Anyway I just want a close fight like the HD 7970 and GTX 680. I want AMD to show up for the fight like they did in the HD 4870 / HD 5870 generation. I believe they will compete with Polaris.

 

[JDG1980]

That would be double Pitcairn's performance at nearly the same die size as Pitcairn. What is pathetic about that? Tahiti was 15% smaller than Cayman and didn't come anywhere remotely close to doubling performance at the time of it's release. You must be expecting miracles.

FinFET should bring higher clocks (the FinFET A9 SoC in the iPhone 6S clocks 32% higher than the planar A8 SoC in the iPhone 6). In addition, AMD will be bringing out architectural improvements with Polaris.

Just a blind die shrink of Hawaii, run at the same clock speed, would be 199mm^2. (Hawaii is 438mm^2, Samsung/GloFo 14LPP is supposed to be 2.2x as dense as 28nm, so 438 / 2.2 ~= 199.) Add 30% for clock speed improvements, another 20%-30% for better architecture, maybe bump up the shader count from 2816 to an even 3072, and we should be expecting a 40%-50% increase in performance over Hawaii on a Pitcairn-sized FinFET GPU from AMD.

 

[JDG1980]

NVIDIA is 100%, all-in on GPUs, making them a very dangerous opponent. The people who are sitting there claiming that NV has its thumb up its rear-end just waiting to get railed by AMD/Polaris just don't have a clue, IMO.

But Nvidia's area of focus with Pascal, from what we've seen so far, is a bit different from what AMD is doing with Polaris. AMD seems to be focusing on mobile first - a very sensible strategy given the current patterns of usage. They are developing small, efficient dies that will work well in laptops and all-in-ones. Considering how well Nvidia's GM107/GM108 sell, this should be an effective strategy. On the other hand, Nvidia seems to be focusing on the HPC and embedded market this generation. The former has been neglected since Kepler, and Nvidia desperately needs a new Tesla chip. That's why GP100 is coming first. At thousands of dollars each, it doesn't matter how low the yields are. And the focus on "deep learning" and FP16 seems to be part and parcel of a strategy to get Nvidia GPUs in more embedded systems, as we saw with the Drive PX 2 presentation.

These are different strategies, but both have the potential to be successful.

I don't think that Nvidia "has its thumb up its rear-end" but I do think AMD is beating them to FinFET in time to market. AMD has demonstrated working silicon, while Nvidia has not, and had to substitute a Maxwell part in the PX 2 presentation. This won't stop Nvidia from being successful in their chosen fields (those supercomputer chips will pay a lot of bills), but it will mean AMD has the chance to get a lot of design wins in laptops with Polaris 10 unless Nvidia brings out GP107 a lot faster than I think they will.

If we really are getting just Polaris 10 and 11 this year from AMD, and if Polaris 11 is the 232mm^2 chip discussed here, gaming enthusiasts might find AMD's new products to be somewhat disappointing. They may well beat the past flagships, but only by 20%-30% or so (in large part due to higher clocks enabled by FinFET). Nvidia will probably have the performance crown once they release GP104 (probably October or so), at least until AMD brings the big guns to the table in 2017. But AMD will still be able to make bank from smaller, more efficient chips that come to market first.

 

[Arachnotronic]

Anyway I just want a close fight like the HD 7970 and GTX 680. I want AMD to show up for the fight like they did in the HD 4870 / HD 5870 generation. I believe they will compete with Polaris.

AMD did show up for the fight, they just lost.

 

[3DVagabond]

IMHO, they are betting their company on Zen, GPU is secondary priority. You can see it in their financial analyst day when they are pumping data center/Zen while saying that dGPU is a slow-growing opportunity.

NVIDIA is 100%, all-in on GPUs, making them a very dangerous opponent. The people who are sitting there claiming that NV has its thumb up its rear-end just waiting to get railed by AMD/Polaris just don't have a clue, IMO.

For AMD to be what they once were (which I believe they are shooting for) they need their CPU business to be successful. Even when their GPU business turns a profit it gets eaten up by the CPU side. That's why they are so focused on Zen's success.

As far as GPU being secondary, there is nothing that points to that. They know they need GPU to realize their overall vision. They wouldn't have put as much into Mantle, LiquidVR, Freesync, HBM... if they considered it "secondary". Also, nVidia is 100% all in on GPU's? Who are you kidding? Remind me what it was they were hawking at CES? Automotive computers, wasn't it?

 

[tviceman]

It would be pathetic because it would mean AMD did nothing to improve architectural efficiency beyond what the FINFET process node shrink brings.

You're not being accurate or realistic. After I just showed you that netting an 80% increase in transistor density, and doubling perf/mm2 would massively outstrip what that last node leap accomplished, you answer in disbelief. AMD has said time and again 2x perf/w improvement. Look at Pitcairn; double the performance and keep the power envelope the same. That gets you 390 performance at Pitcairn power levels, 2x perf/w.

FinFET should bring higher clocks (the FinFET A9 SoC in the iPhone 6S clocks 32% higher than the planar A8 SoC in the iPhone 6). In addition, AMD will be bringing out architectural improvements with Polaris.

Just a blind die shrink of Hawaii, run at the same clock speed, would be 199mm^2. (Hawaii is 438mm^2, Samsung/GloFo 14LPP is supposed to be 2.2x as dense as 28nm, so 438 / 2.2 ~= 199.) Add 30% for clock speed improvements, another 20%-30% for better architecture, maybe bump up the shader count from 2816 to an even 3072, and we should be expecting a 40%-50% increase in performance over Hawaii on a Pitcairn-sized FinFET GPU from AMD.

40-50% you say? Can I quote the bolded part in my signature so we can revisit all these crazy claims when reality sets in?