Tuesday, March 15th 2016

AMD Unveils GPU Architecture Roadmap, "Polaris" to Skip HBM2 Memory?

Alongside its big Radeon Pro Duo flagship graphics card launch, AMD unveiled its GPU architecture roadmap that looks as far into the future as early-2018. By then, AMD will have launched as many as three new GPU architectures. It begins with the launch of its 4th generation Graphics CoreNext architecture, codenamed "Polaris," in mid-2016. Built on the 14 nm FinFET process, "Polaris" is expected to offer a whopping 2.5x increase in performance-per-Watt for AMD, compared to its current GCN 1.2 architecture on 28 nm.

Hot on Polaris' heels, in early-2017, AMD plans to launch the "Vega" GPU architecture. While this appears to offer a 50% increase in performance-per-Watt over Polaris, its highlight is HBM2 memory. Does this mean that AMD plans to skip HBM2 on Polaris, and stick to GDDR5X? Could AMD be opting for a similar approach to NVIDIA, by launching its performance-segment GPU first as an enthusiast product, giving it a free run on the markets till early-2017, and then launching a Vega-based big-chip with HBM2 memory, taking over as the enthusiast-segment product? Some time in early-2018, AMD will launch the "Navi" architecture, which appears to offer a 2.5x performance-per-Watt lead over Polaris, taking advantage of an even newer memory standard.
Add your own comment

46 Comments on AMD Unveils GPU Architecture Roadmap, "Polaris" to Skip HBM2 Memory?

#26
the54thvoid
Stefan Payne
1. There is only one Polaris chip on the roadmap.
AMD statet that there will be two chips. So nothing changed, just some inaccuracy of the roadmap...
So nothing new here...


2. nVidia can NOT just go for HBM, because HBM is a completely different beast than every other memory technology so far!
The manufacturing process is stil pretty damn complicated and for low long has AMD worked on HBM?! Do you really think that they just let nVdia use this technology?!
Besides:
AMD has already an HBM chip out and knows of the downsides - nVidia has not!

So isn't it safe to assume that nVidia may need some more R&D sessions than AMD because they don't know shit about HBM?!

Just having the spec is worthless this time!
Because how do you mount the memory onto the interposer!

You need some experimenting for that...

And also you need the machienery for that - and you can bet that AMD has exclusive concracts with everyone able to place the chips onto the interposer - especially since AMD is one of the HBM pioneers!!!
And nVidia is just again the parasite that comes around when everything is done - like they did with GDDR4 and GDDR5...


Buttom line:
There's a Polaris chip missing on the roadmap, the one that will have HBM.
And HBM insn't as easy to implement this time because it's completely different.

And have you any infos that AMD will use GDDR-5X?! I haven't herad anything about it...
And why should AMD use GDDR-5X?! There's no need for that - with the smaller chip they don't need it, that's something for notebooks and something like that. And with the bigger one they'll use HBM...
Lol. That's all I can say.
Posted on Reply
#27
Stefan Payne
Because you've seen the working nVidia HBM chip??
And you heard about the problems AMD had with Fiji??
And because you know you need some specialized test equipment for HBM??

Oh and if you just make one tiny mistake, you can flush a whole buch of working dice down the toilet...
Posted on Reply
#28
rtwjunkie
PC Gaming Enthusiast
Stefan, nice trolling. You're very good at it. Not so good with facts though. HBM doesn't "belong" to AMD, so there is not "let"ting NVIDIA just have it.
Posted on Reply
#29
the54thvoid
rtwjunkie
Stefan, nice trolling. You're very good at it. Not so good with facts though. HBM doesn't "belong" to AMD, so there is not "let"ting NVIDIA just have it.
I was crafting a reply but got bored. I'll countersign yours.

The one thing I will give is that yes, AMD innovated again with HBM. But without it, Fiji would be really bad compared to Maxwell. Slightly higher power draw with HBM (about 10%) and the higher res performance would have not improved as it does. However, pioneering isn't always great for the present. Sometimes pioneers get overtaken by someone who implements things better. Given how much shouting (PR wise) AMD is doing right now, it's a little bit disconcerting this is bluster to me. Got a bad feeling Polaris won't be the Godly card we're all expecting.
Posted on Reply
#30
HumanSmoke
btarunr
Does this mean that AMD plans to skip HBM2 on Polaris, and stick to GDDR5X?
...or AMD could simply continue to use HBM1 since it now seems to reaching volume production and the required metrology tooling for test and verification is now in place. As Ryan Shrout at PC Per noted:
Curious about the HBM2 inclusion in Vega on the roadmap and what that means for Polaris? Though he didn’t say it outright, it appears that Polaris will be using HBM1, leaving me to wonder about the memory capacity limitations inherent in that. Has AMD found a way to get past the 4GB barrier? We are trying to figure that out for sure.
Stefan Payne
Because you've seen the working nVidia HBM chip??
And you heard about the problems AMD had with Fiji??
And because you know you need some specialized test equipment for HBM??
Oh and if you just make one tiny mistake, you can flush a whole buch of working dice down the toilet...
This is third party manufacturing, not Nvidia (or AMD for that matter) with the exception of a demonstration of a working Nvidia HBM2 chip - but since HBM2 isn't slated for volume production for some months that is hardly surprising. Specialized equipment? I presume you mean X-ray metrology for checking the viability of TSV's within the HBM layers and microbump solder contacts in the microbumps connecting the HBM to the interposer? That is down to companies like KLA-Tencor and NORDSON Dage. The tooling has only recently reached production status hence the protracted delay in ramping AMD's Fiji. Something that is now less of an issue.

If you had been paying attention, you would know that HBM is basically a well understood DDR technology whose only real innovation is the stacked die nature of the memory IC's. The difficulty lies not in its implementation but its manufacture and assembly - neither of which are the down to Nvidia/AMD. Hynix and Samsung manufacture the HBM, UMC and TSMC manufacture the interposer, and Amkor and ASE (and probably TSMC) assemble the GPU+HBM+Interposer package.
Here is a graphic representation of who is actually responsible for the technology implementation if you are still in a state of confusion.
Posted on Reply
#31
buggalugs
I dont think there will be no HBM2 for polaris. I expected them to only use HBM2 on high end enthusiast cards for polaris then the whole range for vega.
Posted on Reply
#32
sweet
HumanSmoke
If you had been paying attention, you would know that HBM is basically a well understood DDR technology whose only real innovation is the stacked die nature of the memory IC's. The difficulty lies not in its implementation but its manufacture and assembly - neither of which are the down to Nvidia/AMD. Hynix and Samsung manufacture the HBM, UMC and TSMC manufacture the interposer, and Amkor and ASE (and probably TSMC) assemble the GPU+HBM+Interposer package.
Here is a graphic representation of who is actually responsible for the technology implementation if you are still in a state of confusion.
You forgot the R&D state where AMD played an important role. They also hold a patent on how to wire stacked DRAM to the interposer.

Back to topic. For AMD, using HBM for at least two generation makes sense though. There is also no chance for a switch to GDDR5X now after spending that much resource on HBM. They just have to deal with the 4GB limit though.

Meanwhile, there is no way nVidia would be allowed to use HBM for free. Hence, they would make a move on GDDR5X when HBM2 is not available for mass production. Still GDDR5X is different from GDDR5. They still need time to implement it perfectly or face the disaster of Fermi with GDDR5 again.
Posted on Reply
#33
HumanSmoke
sweet
There is also no chance for a switch to GDDR5X now after spending that much resource on HBM. They just have to deal with the 4GB limit though.
You think that AMD will absorb the increased manufacturing and assembly cost of GPU+HBM+Interposer for sub-$200 R5 and R7 cards? Unlikely IMO when revising a conventional GDDR5 IMC for use with GDDR5X is very likely to be a minor matter whose R&D should be offset by GDDR5X's doubled density and the ability to shave die space with fewer IMC's/PHY/GDDR5 interfaces if required. Half the number of chips equals shorter PCB and less overall traces. If AMD want to expend extra money on HBM for mainstream cards and the mobile discrete product stack they should be applauded, but it doesn't sound like optimum business sense. They could just continue to tap GDDR5 I suppose but unless AMD aren't boosting core/ROP/TAU counts, they are certainly going to run into either bandwidth issues and/or extra die space to maintain bus width with conventional GDDR5. Since uncore typically doesn't shrink as well as the core, AMD will be effectively devoting disproportionately more die space using the older standard.
sweet
Meanwhile, there is no way nVidia would be allowed to use HBM for free.
You can provide proof of this? Because this just sounds like a bunch of FUD to me, since AMD stated their position six months ago
AMD: We are actively promoting HBM and do not collect royalties
AMD is not involved in collecting any royalties for HBM,” said Iain Bristow, a spokesman for AMD. “We are actively encouraging widespread adoption of all HBM associated technology on [Radeon R9] Fury products and there is no IP licensing associated.”
Advanced Micro Devices owns a number of patents covering HBM, but as that intellectual property is a part of JEDEC’s JESD235 standard, it has to be licensed to applicants desiring to implement the standard “either without compensation or under reasonable terms and conditions that are free of any unfair discrimination.” Moreover, AMD and Nvidia have a broad cross-licensing agreement, which largely prevents royalty demands.
Note that this also covers subsequent revisions to the standard such as JESD235A (HBM2).
sweet
Still GDDR5X is different from GDDR5. They still need time to implement it perfectly or face the disaster of Fermi with GDDR5 again.
Not that different from what I've seen. A doubled prefetch is the biggest takeaway, and Micron have had GDDR5X in the wild for some time, and the GDDR5X emulation tools for memory controllers and physical layers has also been around for some time....

...of course if you have information to the contrary I'd love to see it....although I suspect just like AMD's supposed royalties keeping Nvidia away from HBM, it is more a case of wishful thinking on your part than actual fact.
Posted on Reply
#34
Octopuss
I don't even understand what's going on with the manufacturers with all the cryptic names for stuff. Polaris, Vega, Fiji, Arctic Islands... I don't know what's what. My brain is full of fuck!
Posted on Reply
#35
Stefan Payne
rtwjunkie
Stefan, nice trolling. You're very good at it. Not so good with facts though. HBM doesn't "belong" to AMD, so there is not "let"ting NVIDIA just have it.
The thing is completely different with HBM this time.

You can handsolder GDDR4 and 4 onto a PCB. Good luck with HBM...

The specs mean shit if you are unable to use it. And why the hell would you let nVidia just have HBM?!
Well, they can have it, but it does mean shit because they just can't put everything together!

I think you haven't thought about HBM at all and just say some shit without knowing anything!

With HBM you NEED this highly precise equipment and there aren't many companys out there able to put everything together. Bad luck for nVidia who let others develop the good stuff and come around when everything is done...

AMD and the development partners invested a great amount of money in this new technology and they should just give it to nVidia?! are you kidding?!

As said eariler, the spec means shit, because you have to put everything together, with GDDR4 and 5 you can do it by hand, with HBM that's impossible!

So they _NEED_ pretty damn expansive and precise equipment - something that wasn't around before, that's completely new and developed explicitely for HBM!


As for the Polaris chips:

AMD said there will be _TWO_ chips, one with HBM (the bigger one) and one without HBM...

Why the hell should there now be just one chip?!
Just because of some shitty roadmap?!
Posted on Reply
#36
Naito
Stefan Payne
You can handsolder GDDR4 and 4 onto a PCB. Good luck with HBM...
Good luck hand soldering a 150+ pin BGA GDDR5 package...
Posted on Reply
#37
Stefan Payne
Naito
Good luck hand soldering a 150+ pin BGA GDDR5 package...
Nothing to it.

All you need is a decent hot air station and off you go...
Posted on Reply
#38
rtwjunkie
PC Gaming Enthusiast
HumanSmoke
You can provide proof of this? Because this just sounds like a bunch of FUD to me, since AMD stated their position six months ago
AMD: We are actively promoting HBM and do not collect royalties
Stefan Payne
AMD and the development partners invested a great amount of money in this new technology and they should just give it to nVidia?! are you kidding?!
And they have shared this technology freely, as stated in the above quote. I present it to you since you are light on facts and apparently are not reading HumanSmoke's posts.

Both companies have been saying for the last year that they will be using HBM in the future. This is nothing new.
Posted on Reply
#39
64K
Stefan Payne
The thing is completely different with HBM this time.

You can handsolder GDDR4 and 4 onto a PCB. Good luck with HBM...

The specs mean shit if you are unable to use it. And why the hell would you let nVidia just have HBM?!
Well, they can have it, but it does mean shit because they just can't put everything together!

I think you haven't thought about HBM at all and just say some shit without knowing anything!

With HBM you NEED this highly precise equipment and there aren't many companys out there able to put everything together. Bad luck for nVidia who let others develop the good stuff and come around when everything is done...
Nvidia can't put everything together with HBM nor can they make HBM or the interposer and neither can AMD. Various other companies do that and they will want Nvidia's business too. In fact moreso because Nvidia sells more GPUs so that means more profits for these other companies.
Posted on Reply
#40
HumanSmoke
Octopuss
I don't even understand what's going on with the manufacturers with all the cryptic names for stuff. Polaris, Vega, Fiji, Arctic Islands... I don't know what's what. My brain is full of fuck!
Your brain isn't the only one full. I wonder whose bright idea it was to leave the "Mission Failed" screen running through the whole turgid presentation. Actually quite apt for Richard Huddy's abysmal attempt at humour.

Posted on Reply
#41
rruff
ZoneDymo
They could much easier then AMD just make the jump for all their future cards to HBM2 if they wanted to but nope.avi, rather just have another incremental update then giving consumers the full beans right away because MOAR MONEYZZZ.
You don't understand business very well, do you?
Posted on Reply
#42
EarthDog
rruff
You don't understand business very well, do you?
yup.. last i checked AMD was a 'for profit' organization. That said, there really isn't a need for hbm, nonetheless hbm2 on anything but the high end cards. Typically, at least the past couple of generations, midrange/upper-midrange came out first so...there is that as well.
Posted on Reply
#43
Naito
Stefan Payne
Nothing to it.

All you need is a decent hot air station and off you go...
Maybe I took it too literally - i.e. actually hand soldering not using something like a reflow station. In that case, you'd probably want something more advanced than hot air, perhaps infrared heat.
Posted on Reply
#44
bug
Stefan Payne
Because you've seen the working nVidia HBM chip??
And you heard about the problems AMD had with Fiji??
And because you know you need some specialized test equipment for HBM??

Oh and if you just make one tiny mistake, you can flush a whole buch of working dice down the toilet...
Do you honestly believe that just because nvidia hasn't built a consumer video card using HBM, they know nothing about building one?
Besides, it's non-sensical to complain about the lack of HBM2 in March, when HBM2 memory itself won't be available till September at the earliest.
Posted on Reply
#45
Slizzo
bug
Do you honestly believe that just because nvidia hasn't built a consumer video card using HBM, they know nothing about building one?
Besides, it's non-sensical to complain about the lack of HBM2 in March, when HBM2 memory itself won't be available till September at the earliest.
Gonna caveat this a little: HBM2 won't be available for mass production until September. There's HBM2 being manufactured right now, just not en masse yet.
Posted on Reply
#46
eidairaman1
The Exiled Airman
ZoneDymo
Not sure what you are having problems with understanding there.
Nvidia has a much larger market share, aka much more money to spend on development.
They could much easier then AMD just make the jump for all their future cards to HBM2 if they wanted to but nope.avi, rather just have another incremental update then giving consumers the full beans right away because MOAR MONEYZZZ

Company could give us A B and C in one new generation but they rather give us A first, then A and B, then A B and C because again, moar moneyz.
Sandbagging
Posted on Reply
Add your own comment