updatevideocardzAMD Polaris 10 engineering sample ‘pictured’

[tential]

I am really excited about Polaris 10 if it really is a 390X that uses less power.

I might get one when they are released at the jacked up early price and then get another in a year or so to crossfire in preparation of the Oculus gen 2.

If it's not as fast as Fury X/Fury then I'm not really excited. It needs to be faster than 390x significantly. We already have a 390x a 390x that uses less power is meh. Fury/Fiji that is low power rocks though.

That's what I'm expecting.

 

[poofyhairguy]

If it's not as fast as Fury X/Fury then I'm not really excited. It needs to be faster than 390x significantly. We already have a 390x a 390x that uses less power is meh. Fury/Fiji that is low power rocks though.

That's what I'm expecting.

I don't know. The 390x is like the perfect card for 1080p console ports. I mean I wouldn't mind Fury speeds, but I want 8GB of ram not 4GB.

 

[tential]

I don't know. The 390x is like the perfect card for 1080p console ports. I mean I wouldn't mind Fury speeds, but I want 8GB of ram not 4GB.

Ah see, completely different worlds.

The 390x is NOT a 1080p card in my eyes.
The whole reason I'm excited for a new AMD release is AMD tends to target higher resolutions, so if this card was focused for 1080p, I wouldn't even look at it. It needs to do well at high resolutions just like previous AMD generations. So ya, I expect this card to do well at 4K just like the past and be a card that starts at 1440p just like the 390/x

 

[Qwertilot]

They're not releasing the 4k cards yet though That's presumably down to limited resources, but at least they're managing those sanely this time.

They need a low/mid end card that is actually small/around 390(X) power and can serve as the low/mid end of their line up for the next 3-4 years.

 

[tential]

They're not releasing the 4k cards yet though That's presumably down to limited resources, but at least they're managing those sanely this time.

They need a low/mid end card that is actually small/around 390(X) power and can serve as the low/mid end of their line up for the next 3-4 years.

Well considering the 390x in crossfire is capable of 4K, thats my reasoning with Polaris 10. In crossfire, it should be capable of 4K with Vega being the king obviously...

 

[3DVagabond]

They're not releasing the 4k cards yet though That's presumably down to limited resources, but at least they're managing those sanely this time.

They need a low/mid end card that is actually small/around 390(X) power and can serve as the low/mid end of their line up for the next 3-4 years.

Where are you hearing this? I figured it was down to HBM2 not being quite ready yet and the process not being tried for giant chips before. Who wants to come out with something now, that will likely be bandwidth bottlenecked when the proper tech will be along in a few months? And who wants to risk trying a huge die on a new process?

 

[raghu78]

They're not releasing the 4k cards yet though That's presumably down to limited resources, but at least they're managing those sanely this time.

They need a low/mid end card that is actually small/around 390(X) power and can serve as the low/mid end of their line up for the next 3-4 years.

Raja Koduri said Vega with HBM2 in early 2017 will fill out the upper range of AMD's next gen GPU stack. HBM2 volume availability is scheduled for Q3 2016. It takes atleast another quarter to make its way through GPU production (2.5D stacking). So its a more a case of HBM2 high volume production ramp timelines than any budget constraints which are dictating the launch timeframe of higher GPUs. The production yields at TSMC 16FF+ and GF 14LPP are also not good enough to ramp production of large GPUs (350 - 450 sq mm) at this point in time. AMD will benefit from yield learning and improvement which happens till end of 2016. This is critical for large GPUs which have a much higher rate (%) of GPUs per wafer with defects.

 

[Headfoot]

We have a rather optimal case with GPUs that are currently selling and a developer with a vested interest in showing DX12 on a game that just launched. Yet the issue still occurred and was tested and supported prior to launch?

What happens in the 2018 time frame when we will have another set of GPU architectures added into the mix combined with games possibly at 1yr+ post release? What is the real support situation going to be?

I'm surprised this potential issue with low level APIs is not getting more attention.

I've been thinking about this too. My take is that if you look at today's gaming scene, every AAA release since 2014 has come with a season pass / DLC option. Studios are planning for their releases to have an effectively longer lifespan via DLC. I imagine they will build in the cost of updating for new cards into the calculation for the P/L on the DLC and support budgets.

Indies will probably continue to use 3rd party engines (Unity, Unreal, CryEngine) so that task is abstracted away from them provided they at least do the engine level patching. Either that, or they will continue to use DX11.

The unknown is the midrange game release -- not quite AAA in budget and DLC planning, but bigger than indie.

I think you wont be able to expect up to date DX12 card optimizations past 2 years after release, and probably not even compatibility with newer cards long past that for the average developer. Most multiplayer games have nobody playing them 2 years later anyways. Such is the state of modern AAA gaming on PC

 

[3DVagabond]

Raja Koduri said Vega with HBM2 in early 2017 will fill out the upper range of AMD's next gen GPU stack. HBM2 volume availability is scheduled for Q3 2016. So its a more a case of HBM2 high volume production ramp timelines than any budget constraints which are dictating the launch timeframe of higher GPUs. The production yields at TSMC 16FF+ and GF 14LPP are also not good enough to ramp production of large GPUs (350 - 450 sq mm) at this point in time. AMD will also benefit from yield learning and improvement which happens till end of 2016.

Raja gave a date?

 

[Piano Man]

so scandalous

showing a healthy amount of ankle with a little calf for teasing

LOL, I love it

 

[Qwertilot]

I think you wont be able to expect up to date DX12 card optimizations past 2 years after release, and probably not even compatibility with newer cards long past that for the average developer. Most multiplayer games have nobody playing them 2 years later anyways. Such is the state of modern AAA gaming on PC

A bit absurd if they won't run though. A lot of people play stuff for ages

The obvious answer is probably to include a 'DX11 style' path in the games for stuff going forwards. Not optimised, but newer cards >> performance anyway so will survive.

 

[3DVagabond]

A bit absurd if they won't run though. A lot of people play stuff for ages

The obvious answer is probably to include a 'DX11 style' path in the games for stuff going forwards. Not optimised, but newer cards >> performance anyway so will survive.

They likely won't require the latest and greatest hardware to run acceptably and will do just fine on new HW even unoptimized.

 

[Qwertilot]

Where are you hearing this?

AMD's (relatively!) limited resources have surely been patently obvious in recent years?

I figured it was down to HBM2 not being quite ready yet and the process not being tried for giant chips before. Who wants to come out with something now, that will likely be bandwidth bottlenecked when the proper tech will be along in a few months? And who wants to risk trying a huge die on a new process?

That's why no giant chips, yes. Makes absolute sense. NV will quite possibly do it with Tesla's but those command huge prices!

Thing is, the absolutely optimal strategy here is to get the top card of the first wave of releases a chunk past the current top end. Not hugely so, as it has to be able to drop a tier later on but enough to get some upgrade sales. ~170W say.

It sounds like AMD are maybe not going to do that, which is absolutely fine. Much saner than their planning in the past few years.

 

[Qwertilot]

They likely won't require the latest and greatest hardware to run acceptably and will do just fine on new HW even unoptimized.

Hopefully. There are real worries though. Stuff has been falling over entirely pre patch on even GCN 1.2 and the companies won't be able to patch it via drivers half as easily.

It will also make benchmarking new ranges of cards quite amusing if they're up against games micro optimised for the previous gen of cards......
(Or of course a brilliant way to make quite sure people upgrade!).

 

[raghu78]

Raja gave a date?

No. But its an educated guess based on HBM2 production timelines. btw anandtech's take was similar

http://www.anandtech.com/show/10145/amd-unveils-gpu-architecture-roadmap-after-polaris-comes-vega

"Vega is currently scheduled to come relatively quickly after Polaris. Depending on how literal you interpret this chart, the far left edge of the Vega box does fall into 2016, though obviously AMD intends to leave themselves some wiggle room here and not tie themselves down to specific dates. The fact that Vega comes this soon after Polaris is interesting; it seems hard to believe that it’s a direct successor to Polaris – I can’t see AMD replacing Polaris parts in less than a year – so this points to Vega being more of a cousin, and is where AMD’s naming system isn’t especially helpful in deciphering anything further.

With Polaris confirmed to use GDDR5, Vega is notable for being the first AMD architecture to use HBM2, and the first parts in general to use HBM tech since Fiji. I’m presuming these are higher-end GPUs to complement the Polaris GPUs (the smaller of which we know to be a low-power laptop design), which is where HBM would be more cost-effective, at least at current prices.

Meanwhile AMD has also confirmed the number of GPUs in the Vega stack and their names. We’ll be seeing a Vega 10 and a Vega 11. This follows Polaris GPU naming – which has finally been confirmed – with Polaris 10 and Polaris 11. I have also been told that Polaris 11 is the smaller of the Polaris GPUs, so at this point it’s reasonable to assume the same for Vega."

 

[3DVagabond]

Raja Koduri said Vega with HBM2 in early 2017 will fill out the upper range of AMD's next gen GPU stack. HBM2 volume availability is scheduled for Q3 2016. It takes atleast another quarter to make its way through GPU production (2.5D stacking). So its a more a case of HBM2 high volume production ramp timelines than any budget constraints which are dictating the launch timeframe of higher GPUs. The production yields at TSMC 16FF+ and GF 14LPP are also not good enough to ramp production of large GPUs (350 - 450 sq mm) at this point in time. AMD will benefit from yield learning and improvement which happens till end of 2016. This is critical for large GPUs which have a much higher rate (%) of GPUs per wafer with defects.

No. But its an educated guess based on HBM2 production timelines. btw anandtech's take was similar

http://www.anandtech.com/show/10145/amd-unveils-gpu-architecture-roadmap-after-polaris-comes-vega

"Vega is currently scheduled to come relatively quickly after Polaris. Depending on how literal you interpret this chart, the far left edge of the Vega box does fall into 2016, though obviously AMD intends to leave themselves some wiggle room here and not tie themselves down to specific dates. The fact that Vega comes this soon after Polaris is interesting; it seems hard to believe that it’s a direct successor to Polaris – I can’t see AMD replacing Polaris parts in less than a year – so this points to Vega being more of a cousin, and is where AMD’s naming system isn’t especially helpful in deciphering anything further.

With Polaris confirmed to use GDDR5, Vega is notable for being the first AMD architecture to use HBM2, and the first parts in general to use HBM tech since Fiji. I’m presuming these are higher-end GPUs to complement the Polaris GPUs (the smaller of which we know to be a low-power laptop design), which is where HBM would be more cost-effective, at least at current prices.

Meanwhile AMD has also confirmed the number of GPUs in the Vega stack and their names. We’ll be seeing a Vega 10 and a Vega 11. This follows Polaris GPU naming – which has finally been confirmed – with Polaris 10 and Polaris 11. I have also been told that Polaris 11 is the smaller of the Polaris GPUs, so at this point it’s reasonable to assume the same for Vega."

That's fine. I just read the bold and wanted to be certain if it was your educated guess or if he actually stated that.

 

[Slaughterem]

Maybe some of you could explain this information that was in a GloFo webinar about 14nm LPE and LPP. https://vimeo.com/142694548 {Go to 36 min mark}They did a design test case on an ARM cortex A9 Falcon-Neon graphics and showed the Area, Performance Power advantages. The results indicated an Area reduction of around 75%. So my question would be if Fiji 8.9 billion transistors = 596 mm2 on 28nmHPP then a direct copy of that on 14LPP would be 8.9 billion = 150mm2 and then a Polaris 10 232mm2 chip would be capable of containing somewhere around 13.6 billion?

 

[3DVagabond]

Maybe some of you could explain this information that was in a GloFo webinar about 14nm LPE and LPP. https://vimeo.com/142694548 {Go to 36 min mark}They did a design test case on an ARM cortex A9 Falcon-Neon graphics and showed the Area, Performance Power advantages. The results indicated an Area reduction of around 75%. So my question would be if Fiji 8.9 billion transistors = 596 mm2 on 28nmHPP then a direct copy of that on 14LPP would be 8.9 billion = 150mm2 and then a Polaris 10 232mm2 chip would be capable of containing somewhere around 13.6 billion?

I'm not that optimistic.

 

[Silverforce11]

Maybe some of you could explain this information that was in a GloFo webinar about 14nm LPE and LPP. https://vimeo.com/142694548 {Go to 36 min mark}They did a design test case on an ARM cortex A9 Falcon-Neon graphics and showed the Area, Performance Power advantages. The results indicated an Area reduction of around 75%. So my question would be if Fiji 8.9 billion transistors = 596 mm2 on 28nmHPP then a direct copy of that on 14LPP would be 8.9 billion = 150mm2 and then a Polaris 10 232mm2 chip would be capable of containing somewhere around 13.6 billion?

It doesn't scale perfectly for some chip parts like MC, interlinks.

Also, last I saw, reported density increase was 2.2 to 2.4x compared to 28nm.

 

[Glo.]

Slaughterem, You compare nodes between GloFo and GloFo. Fury was built on 28 nm TSMC process. And by the look of some slides on SemiWiki https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html
It looks like GloFo 28nm process was a bit more dense than 28 nm TSMC.

So the differences may be bigger. Also what you have provided looks like it is hybrid process between 14nm SOI from IBM and 14nm LPP from Samsung. Otherwise it would not get that die are conversion and leakage reduction.

But overall, yes it would be that small, however, there are other factors of core and uncore area. Uncore area may not scale linearly with process.

P.S. A bit of personal note. Lately I have started expecting huge surprise from Polaris GPUs. In terms of power and performance.

P.S. http://semiaccurate.com/forums/showpost.php?p=257667&postcount=122
http://semiaccurate.com/forums/showpost.php?p=257736&postcount=127
http://semiaccurate.com/forums/showpost.php?p=257772&postcount=131
http://semiaccurate.com/forums/showpost.php?p=258067&postcount=144
Also posts from Zen CPU thread of this guy on this forum are related to this. He may be dead right on this.

 

[raghu78]

Slaughterem, You compare nodes between GloFo and GloFo. Fury was built on 28 nm TSMC process. And by the look of some slides on SemiWiki https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html
It looks like GloFo 28nm process was a bit more dense than 28 nm TSMC.

So the differences may be bigger. Also what you have provided looks like it is hybrid process between 14nm SOI from IBM and 14nm LPP from Samsung. Otherwise it would not get that die are conversion and leakage reduction.

But overall, yes it would be that small, however, there are other factors of core and uncore area. Uncore area may not scale linearly with process.

P.S. A bit of personal note. Lately I have started expecting huge surprise from Polaris GPUs. In terms of power and performance.

P.S. http://semiaccurate.com/forums/showpost.php?p=257667&postcount=122
http://semiaccurate.com/forums/showpost.php?p=257736&postcount=127
http://semiaccurate.com/forums/showpost.php?p=257772&postcount=131
http://semiaccurate.com/forums/showpost.php?p=258067&postcount=144
Also posts from Zen CPU thread of this guy on this forum are related to this. He may be dead right on this.

seronx is the last person who is likely to say anything true or even plausible. Most of his posts are a product of his wild imagination with very little factual basis or even technical plausibility. GF 14LPP will be a huge leap over TSMC 28nm (which is the node on which AMD GPUs have been built upto now). Qualcomm is quoted as saying TSMC 16FF+ and GF 14 LPP are quite similar in electrical characteristics and performance. GF 14LPP is slightly more denser than TSMC 16FF+.

http://semiaccurate.com/2015/11/14/qualcomm-shows-us-the-snapdragon-820/

"Weighting on the foundry debate Qualcomm said that it does not believe that there is a significant performance difference between TSMC and Samsung. They chose Samsung’s LPP 16nm FinFET process over TSMC’s 16nm FinFET alternative for cost and volume reasons."

 

[Arachnotronic]

"Weighting on the foundry debate Qualcomm said that it does not believe that there is a significant performance difference between TSMC and Samsung. They chose Samsung’s LPP 16nm FinFET process over TSMC’s 16nm FinFET alternative for cost and volume reasons."

Hah, nice PR spin

 

[raghu78]

Hah, nice PR spin

We will know this year for sure if Qualcomm's statements was PR spin as Polaris is 14LPP and Pascal is 16FF+. We should also get some transistor structural and electrical characteristics details from the likes of chipworks or techinsights as they analyze the Snapdragon 820.

 

[Arachnotronic]

We will know this year for sure if Qualcomm's statements was PR spin as Polaris is 14LPP and Pascal is 16FF+. We should also get some transistor structural and electrical characteristics details from the likes of chipworks or techinsights as they analyze the Snapdragon 820.

They are pure spin. As I understand it, Qualcomm basically chose Samsung for two reasons:

1. Wafer prices - Samsung priced 14nm wafers fairly aggressively relative to TSMC.

2. Quid pro quo -- It is said that Samsung Mobile has mandated that its flagship devices use only Samsung-built chips.

Right now, Samsung Foundry is basically betting on Qualcomm to sustain its leading edge volumes. Apple is going away with the A10 and I strongly suspect for the 10nm A11 and beyond. MediaTek, HiSilicon, Spreadtrum, NVIDIA, etc. are all seemingly choosing TSMC.

TSMC also has the benefit of having all of the 2nd and 3rd wave customers (i.e. SSD controllers, display driver ICs, etc.) generally choosing them.

IMO, Samsung did a very good job of running a nice fancy PR campaign by being "first" to 14nm (even though the technology is widely believed to have been ripped off from TSMC), but I think at 10nm and 7nm you will start to see TSMC really pulling away from Samsung.

BTW, I wouldn't buy the spin from AMD about 14nm being "better" than TSMC's 16nm (see: Robert Hallock comments on reddit). This is taken from AMD's latest form 10-K:

GLOBALFOUNDRIES Inc. On March 2, 2009, we entered into a Wafer Supply Agreement (WSA) with GLOBALFOUNDRIES Inc. (GF). The WSA governs the terms by which we purchase products manufactured by GF, a related party to us. Pursuant to the WSA, we are required to purchase all of our microprocessor and APU product requirements, and a certain portion of our GPU product requirements from GF with limited exceptions.