videocardzAMD’s official GPU Roadmap for 2016-2018

[Headfoot]

Yes the US has massive problems. We just act oblivious and better than everyone

what?

 

[Headfoot]

Aha, thank you.
Is it possible to renew the patent ?

Yeah, i just read some more about PIM. I find it very interesting indeed.
https://en.wikipedia.org/wiki/Computational_RAM



I had to think of the xbox360 gpu with edram right away.

No you can't renew a patent, once the patent's term has expired it is dedicated to the public. That is "the deal" so to speak, in exchange for 20 years exclusivity you agree to provide the patent explaining (kinda) how to do the thing, and the rights for anyone to do it in 20 years. In reality, you can get patents on improvements to something previously patented so as long as you keep innovating you can continue to get patents around the same invention via improvements though never on the exact same thing

 

[RussianSensation]

I'd say that's pretty much confirmed, hint scalability :sneaky:

Maybe, maybe not. AMD has used the term scalability before to describe building upon existing products by making them faster (i.e., growing die size, GPU, memory clocks). It doesn't necessarily mean multi-GPU scalability. It could mean growing the die size from 438mm2 R9 290X to 596mm2 Fury X while scaling GCN architecture, using next gen memory (GDDR5 replaced with HBM1 in this case).

Applying this idea to Vega and Navi could mean:

450mm2 Vega -> 600mm2 Navi
1TB/sec HBM2 Vega -> 1.5TB/sec HBM2 Navi

I'd consider faster HBM2 as still next generation memory.

 

[Qwertilot]

Not sure if I'd quite call it small die first. It looks/feels more a bit more like a console first strategy. Well semi custom stuff in general too of course.

It would be very rational- that stuff has been something of a lifeline to them, but they really haven't had a gpu designed to do it really well as yet. This stuff shows every sign of being that.

 

[RussianSensation]

Not sure if I'd quite call it small die first. It looks/feels more a bit more like a console first strategy.

In the past, AMD got away with small die strategy because they priced their GPUs much lower and they obliterated NV in perf/mm2 and perf/watt, while older gen GTX200/Fermi were not very competitive in these categories. Now that NV is leading in perf/mm2, perf/watt and are making 600mm2 flagships, small die strategy will not work. If Vega is at best a GP104 competitor, market share will go down to 15%. AMD may launch small die Polaris 10, but not a chance they are going small die strategy if they plan on regaining back market share in the $300-650 segments.

Furthermore, it makes no sense whatsoever to take a 250-300mm2 die and slap on 1GB of HBM2 on it. That's overkill. This means Vega will be a much larger die than Polaris 10 to take advantage of all that added bandwidth.

What's happening with Polaris 10 isn't small die strategy. It's launching lower-end SKUs first in the market segments where you have been weakest since 750/750Ti launched. In the sub-$300 mobile and desktop dGPU spaces, NV cleaned up with these weak but highly perf/watt efficient parts.

Why do you think AMD pitted GTX950 against Polaris 10 and showed the delta in power usage? They could have shown a part with 110W TDP and 30-50% more FPS but they didn't. Clearly Polaris is their perf/watt play, not their performance play. Remember, a new node shrink + new architecture tends to bring 60-100% more performance. Even if Polaris 10 matches Fury X, by next gen standards, that's no better than a low mid-range part. After-all, 1080 GP104 (a straight up mid-range product) is going to beat 980Ti by most likely 20-30%. So what would we call Polaris 10 other than next gen low mid-range?

By extension, that means Vega and GP102 (or whatever it'll be called) are the real deal, aka the real flagships of 1st gen 14nm/16nm generation.

What AMD and NV are doing now is launching GTX650/650Ti/660/660Ti/670/680 level cards first and pricing them much higher than in the past. NV has been far more successful with this strategy though.

 

[tential]

what?

Us patent laws are out of date but that's a whole other thing entirely.

 

[Bryf50]

Why do you think AMD pitted GTX950 against Polaris 10 and showed the delta in power usage?

That was Polaris 11.

 

[Headfoot]

Us patent laws are out of date but that's a whole other thing entirely.

No, they aren't. There was a colossal update, the biggest update in the history of US patent laws in 2011. America Invents Act, which since you are so up on patent law you obviously already knew. Not to mention the sea change post-Alice in 2014. Biggest change in software patents since State Street Bank, which is a case I'm sure you're familiar with. You can say a lot of things about US patent law, but being out of date is not one of them. This is off topic however.

 

[RussianSensation]

That was Polaris 11.

Thanks for correcting me. Ya, I meant AMD decided to show off how their Polaris 11 would deliver similar performance to a 950 with less power, rather than focus on more performance at the same power as a 950. This suggests they are really gunning for utmost efficiency with lower end Polaris chips, while I presume the performance focus in the $350-650 markets will be attained with Vega. I somehow doubt hat Polaris 10/11 were designed around PS4K for example. I think they may have targeted a specific TDP to allow Polaris 10/11 to fit into smaller PCs but not necessarily benchmarked Polaris's performance against PS4. That's not a bar high enough when your competitor already has a GTX970 in a 145W TDP reference design using 'last generation' architecture and node!

 

[swilli89]

Thanks for correcting me. Ya, I meant AMD decided to show off how their Polaris 11 would deliver similar performance to a 950 with less power, rather than focus on more performance at the same power as a 950. This suggests they are really gunning for utmost efficiency with lower end Polaris chips, while I presume the performance focus in the $350-650 markets will be attained with Vega. I somehow doubt hat Polaris 10/11 were designed around PS4K for example. I think they may have targeted a specific TDP to allow Polaris 10/11 to fit into smaller PCs but not necessarily benchmarked Polaris's performance against PS4. That's not a bar high enough when your competitor already has a GTX970 in a 145W TDP reference design using 'last generation' architecture and node!

At the same time I doubt AMD wanted to give the competition too much insight into outright Polaris performance. They merely wanted to show the world that they had working silicon way before nvidia and that they were able to utilize 14nm to save power. Polaris 11 will at least double 950 performance, no doubt about that.

 

[Silverforce11]

That's why I think for Hawaii users, Polaris 10 isn't really meant to be a true generational upgrade. That's most likely Vega.

AMD is not even trying to hide that Polaris 10 isn't a next gen flagship card.

They haven't hide anything, they've been quite clear from the start of this year, Polaris is aimed solely to bring a huge perf/w leap with 11, and with 10, it's to deliver a better VR experience at a lower price point than the $339 970/290/390 minimum requirements.

What this means if you break it down is simple. Polaris 10's target is Hawaii+, at a lower price point.

Polaris 11's target is raw perf/w for notebooks and entry gaming. It may not even hit 380X performance, but 380 performance at <50W? That's pretty good for the market it's aimed at.

Vega 11 is where it's at, for current owners of 970/980/390/X to upgrade and Vega 10 will be for Fiji/980Ti owners to upgrade to.

 

[RussianSensation]

Vega 11 is where it's at, for current owners of 970/980/390/X to upgrade and Vega 10 will be for Fiji/980Ti owners to upgrade to.

Very, very risky strategy though because many 970/980/390/390X owners will upgrade to GP104 and not wait 6+ months. Also, to preserve resale value, I bet many 980Ti owners will get 1080, even for 20-30% more performance.

 

[Silverforce11]

Very, very risky strategy though because many 970/980/390/390X owners will upgrade to GP104 and not wait 6+ months. Also, to preserve resale value, I bet many 980Ti owners will get 1080, even for 20-30% more performance.

That's fine, as long as 750Ti, 950, 960 and 380/X (and whoever still has 7870, 7950/70 etc) owners upgrade to Polaris 10 (IMO, 390X+ performance, ~110W, $299).

And those even on older stuff, who want a cheap low power, can go with Polaris 11.

Vega will be very competitive when it arrives (more perf/w than Polaris, bigger chip, higher TDP), no worries.

ps. Did you see the leaked prices for the 1070/1080? The 1080 is a CUT GP104, the full die is 1080Ti. Despite that, the 1080 is ~$649 and the 1070 is ~$499. Not in the same segment as Polaris 10, not a competitor.

 

[R0H1T]

That's fine, as long as 750Ti, 950, 960 and 380/X (and whoever still has 7870, 7950/70 etc) owners upgrade to Polaris 10 (IMO, 390X+ performance, ~110W, $299).

And those even on older stuff, who want a cheap low power, can go with Polaris 11.

Vega will be very competitive when it arrives (more perf/w than Polaris, bigger chip, higher TDP), no worries.

ps. Did you see the leaked prices for the 1070/1080? The 1080 is a CUT GP104, the full die is 1080Ti. Despite that, the 1080 is ~$649 and the 1070 is ~$499. Not in the same segment as Polaris 10, not a competitor.

I'd like GDDR5x & nothing less, for people (like me) who upgrade less often there has to be more "innovation" & value in the "next gen" GPU's & I think a lot of prospective buyers, in the 150~250$ range, will also hold onto their GPU's for a bit longer. This doesn't apply to brand loyal customers but I see people going slow with this new release, from both the camps.

 

[maddie]

I'd like GDDR5x & nothing less, for people (like me) who upgrade less often there has to be more "innovation" & value in the "next gen" GPU's & I think a lot of prospective buyers, in the 150~250$ range, will also hold onto their GPU's for a bit longer. This doesn't apply to brand loyal customers but I see people going slow with this new release, from both the camps.

Waiting for GDDR5X. Do you buy by tech or by performance and cost factors?

Also, waiting till when? Volta or Navi generation?

 

[R0H1T]

Waiting for GDDR5X. Do you buy by tech or by performance and cost factors?

Also, waiting till when? Volta or Navi generation?

So you're saying people would buy Vega &/or 1080Ti even if it were GDDR5? There is a reason why the high end will be HBM2 & if there are some (same?) models of Polaris being sold with (8GB?) GDDR5x then I'd definitely prefer them over the plain old GDDR5 ones & I'm not the only one who thinks that way.

Cost & perf/w as well as OC headroom, in no particular order, but if I had to choose just one that'd be cost since GPU's sell for a major premium in my part of the world & they rarely come down in price, if ever.

 

[maddie]

So you're saying people would buy Vega &/or 1080Ti even if it were GDDR5? There is a reason why the high end will be HBM2 & if there are some (same?) models of Polaris being sold with (8GB?) GDDR5x then I'd definitely prefer them over the plain old GDDR5 ones & I'm not the only one who thinks that way.

Cost & perf/w as well as OC headroom, in no particular order, but if I had to choose just one that'd be cost since GPU's sell for a major premium in my part of the world & they rarely come down in price, if ever.

I think the FuryX / 980Ti match-up should answer most of those questions.

If some Polaris models have GDDR5X modules, I would expect a $ increase commensurate with the performance increase. In other words, you get what you pay for. You don't get to pick between GDDR5 and GDDR5X cards at the same price and same performance, so if you can only afford $X, then you can only get the card in that price range. Waiting for the tech to trickle down, means skipping a generation most likely.

Remember, there is always something newer coming next year.

 

[R0H1T]

I think the FuryX / 980Ti match-up should answer most of those questions.

If some Polaris models have GDDR5X modules, I would expect a $ increase commensurate with the performance increase. In other words, you get what you pay for. You don't get to pick between GDDR5 and GDDR5X cards at the same price and same performance, so if you can only afford $X, then you can only get the card in that price range. Waiting for the tech to trickle down, means skipping a generation most likely.

Remember, there is always something newer coming next year.

Of course that's what I expect nothing less, but I doubt we'll get the same GPU being massively overpriced just because it's equipped with GDDR5x, even though the bandwidth will likely be doubled.

That's what I usually do, unless I need something urgently, & pay just about (what I feel is) the right price for a given piece of tech, same goes for smartphones & every other disposable electronic item I purchase.

I will however pay a slight premium for durability/reliability & after sales service from certain brands, that are good at it. Not all of us have unlimited funds to pursue a hobby & being realistic, or conservative, in not overspending on them is the right thing to do IMO.

 

[maddie]

Of course that's what I expect nothing less, but I doubt we'll get the same GPU being massively overpriced just because it's equipped with GDDR5x, even though the bandwidth will likely be doubled.

That's what I usually do, unless I need something urgently, & pay just about (what I feel is) the right price for a given piece of tech, same goes for smartphones & every other disposable electronic item I purchase.

I will however pay a slight premium for durability/reliability & after sales service from certain brands, that are good at it. Not all of us have unlimited funds to pursue a hobby & being realistic, or conservative, in not overspending on them is the right thing to do IMO.

Radeon 4850 and 4870. GDDR3 and GDDR5. $199 and $299.

http://anandtech.com/print/2556/

I have to tell you that this Polaris launch reminds me very much of that launch.

 

[railven]

That's fine, as long as 750Ti, 950, 960 and 380/X (and whoever still has 7870, 7950/70 etc) owners upgrade to Polaris 10 (IMO, 390X+ performance, ~110W, $299).

And those even on older stuff, who want a cheap low power, can go with Polaris 11.

Vega will be very competitive when it arrives (more perf/w than Polaris, bigger chip, higher TDP), no worries.

ps. Did you see the leaked prices for the 1070/1080? The 1080 is a CUT GP104, the full die is 1080Ti. Despite that, the 1080 is ~$649 and the 1070 is ~$499. Not in the same segment as Polaris 10, not a competitor.

Be interesting if those leaked prices are in a sense to get AMD to launch high. I recall the Fury X price leak/rumors and then NV punched them in the stomach.

20-30% perf increase for >$700? Hard pill to swallow. And my gravy train is running out here since I'm ready to change careers. Decisions decisions.

 

[maddie]

Be interesting if those leaked prices are in a sense to get AMD to launch high. I recall the Fury X price leak/rumors and then NV punched them in the stomach.

20-30% perf increase for >$700? Hard pill to swallow. And my gravy train is running out here since I'm ready to change careers. Decisions decisions.

My belief lately is that AMD will allow Nvidia to launch first, set their prices and then launch.

Remember Koduri never said that AMD would launch first overall. This is what he said per the venturebeat article.

"We believe we’re several months ahead of this transition, especially for the notebook and the mainstream market. The competition is talking about chips for cars and stuff, but not the mainstream market."

GP104 is neither.

 

[MrTeal]

At the same time I doubt AMD wanted to give the competition too much insight into outright Polaris performance. They merely wanted to show the world that they had working silicon way before nvidia and that they were able to utilize 14nm to save power. Polaris 11 will at least double 950 performance, no doubt about that.

Double 950 performance is 970 level. P11 might be a good GPU, but it's just too small and low power to hit that. The only time I could see P11 doubling the 950's performance would be if it ships with 4GB and you tested at a resolution/setting mix that tanked performance on the 2GB 950.

 

[Erenhardt]

What if... what if they secretly agreed behind the scenes to take their slice of cake for themselves? Tinfoil mode on:

AMD gets mainstream market and price guage $400 for 200mm sq and $200 for 100 mm sq chip.

NV gets high end with 350mm sq chips @ $600

They cover separate markets, not competing on price. Everyone gets milked while they enjoy higher margins without even trying...

Grim...

 

[railven]

What if... what if they secretly agreed behind the scenes to take their slice of cake for themselves? Tinfoil mode on:

AMD gets mainstream market and price guage $400 for 200mm sq and $200 for 100 mm sq chip.

NV gets high end with 350mm sq chips @ $600

They cover separate markets, not competing on price. Everyone gets milked while they enjoy higher margins without even trying...

Grim...

Didn't NV already coerce ATI to price fix and both got fined for it?

 

[railven]

Radeon 4850 and 4870. GDDR3 and GDDR5. $199 and $299.

http://anandtech.com/print/2556/

I have to tell you that this Polaris launch reminds me very much of that launch.

If AMD pulls this off again, they'd have me back in their team without question.

It's when AMD tried to charge the same as NV for a product that on most accounts was inferior that I couldn't support big red.

80-90% performance of whatever NV has as their top card for 60-80% of the price? Yes, please.