AMD: Speculation, Rumors, and Discussion (Archive)

[Entropy]

AMD announces replacement for display controller block: http://www.anandtech.com/show/9836/amd-unveils-2016-vistech-roadmap

Display Port 1.3, HDMI 2.0a, Freesync over HDMI, and HDR support. HDR support also requiring some other things, and what exactly is supported (there's a lot of argument over what 'HDR' should mean in the industry) isn't specified.

Be that as it may, this is very welcome. Or will be, once products actually ship. It will facilitate the next step in displays and may come in handy for VR.

And about air-cooling: The reviews at the recent launches of both the Fury and the 380x were made using AIB products. So it seems AMD has learned how to dodge the stigma of noisy reference coolers (while also ensuring that the review results are more relevant for end user enthusiast, who have no reason to care about TAM).

 

[FriendlyNeighbour]

What's interesting is that NVidia proved that it's possible to make a blower style cooler that works and doesn't sound atrocious. It doesn't wiggle within the tight constraints of some group of people who are clearly not enthusiast gamers. So what. It's an easy sell to the actual enthusiast gamers who've bought reference blower cooled 780Ti and still buy reference cooled Titan X and 980 Ti.

I literally laughed when I read your last two sentences. If you seriously think that anyone who doesn't like a noisy and useless blower fan isn't an "enthusiast" then you're absolutely hilarious

 

[Babel-17]

Are there any third party blower coolers? They tend to go for an open design.

The HIS IceQ Radeon HD 7950 now sitting in my other PC has a design that was well received. Tom's reviewed it, though mine was the non-turbo version. It had several versions on Newegg, and they got decent reviews.
http://www.tomshardware.com/reviews/radeon-hd-7950-review-benchmark,3207-2.html
http://www.newegg.com/Product/Produ...tion=HIS+IceQ++Radeon+HD+7950&N=-1&isNodeId=1
Came with great software, could even adjust the memory voltage, and monitor the VRM for overheating.

Summary And Conclusion
Overall, two very different graphics cards rise to the top. First, we have HIS' HD 7950 IceQ Turbo sporting a tall two-slot thermal design. It manages to edge out the other five boards, taking first place in performance and cooling, along with third place in acoustics. Second, there's Asus' HD 7950 DirectCU II, a heavier three-slot card that generates the least noise, finishes a close second in cooling performance, and still posts competitive benchmark results. PowerColor's HD 7950 PCS runs quietly, but suffers much warmer operating temperatures. And the other two factory-overclocked submissions are both hotter and louder. Finally, AMD's reference design demonstrated the worst performance, the least-attractive acoustic profile, and the highest temperatures. Its only redeeming quality is price.

Edit:

Best Cooling Performance: HIS HD 7950 IceQ Turbo
The DHE cooling solution on HIS' HD 7950 IceQ Turbo really works. Not only does it cool well, but it outperforms the thermal solutions of the other cards we tested. It also manages to beat them in the gaming benchmarks without generating that much extra noise.

Although this board won our recommendation in Europe, we're not able to give it an endorsement in the U.S. As we've seen HIS do in the past, it's only selling this higher-clocked Turbo model overseas. North American customers can get the HD 7950 IceQ (no Turbo), but it runs at a reference 800 MHz and costs more than $400. Sure, it's more attractive than AMD's reference design. But it's certainly not a front-runner at that lower speed.

 

[3dilettante]

The most visible trade-off in that case is the custom blower's utilizing more height than the standard dimensions.

 

[dbz]

The most visible trade-off in that case is the custom blower's utilizing more height than the standard dimensions.

Strangely enough, the GTX 970 is the beneficiary of quite a number of AIB custom blower shroud designs. Those from Asus (also available on the GTX 960) and Gigabyte both feature a through-and-through blower intake (drawing air from both sides of the board). Both, like the EVGA reference style card are geared towards multi-GPU setups. The other blowers used by MSI and PNY seem to be reduced BoM SKUs using recycled reference-style GTX 760/670 coolers.

 

[3dilettante]

Strangely enough, the GTX 970 is the beneficiary of quite a number of AIB custom blower shroud designs. Those from Asus (also available on the GTX 960) and Gigabyte both feature a through-and-through blower intake (drawing air from both sides of the board). Both, like the EVGA reference style card are geared towards multi-GPU setups. The other blowers used by MSI and PNY seem to be reduced BoM SKUs using recycled reference-style GTX 760/670 coolers.

That would seem to be consistent with the contention that blowers within the standard dimensions and specs are reaching the margins of their effectiveness, since the 970 at stock needs to dissipate half the power.
The benefits to a cheaper cooler, not penalizing poorer case airflow, or catering to multi-GPU setups would not be counterbalanced by an uncompromising physical requirement.

 

[dbz]

That would seem to be consistent with the contention that blowers within the standard dimensions and specs are reaching the margins of their effectiveness, since the 970 at stock needs to dissipate half the power.
The benefits to a cheaper cooler, not penalizing poorer case airflow, or catering to multi-GPU setups would not be counterbalanced by an uncompromising physical requirement.

I think that really only applies when the cards in question start breaching the PCI 300W spec. from my own experience, admittedly Maxwell is bit of an oddball, since regardless of cooler and power input over and above the nominal 250W board power, all overclocking limits are very similar whether you run a reference blower at 275W, or a 300W+ AIB custom like the EVGA Classified, Asus Matrix/MSI Lightning unless you are using sub-zero cooling. It should also be noted that aside from noise, AMD's current flagships are in much the same boat regarding overclocking headroom. I honestly think this trend will continue. AMD and Nvidia have more to gain from reduced warranty claims by clamping down on overclocking excesses for all but the most determined, and keeping the OEM vendors onside with conventional PCI/ATX specced blowers - not just the gaming OEMs, but those looking at bigger ticket hardware.

One thing that does need to taken into consideration is the proliferation of SFF and mATX builds as people migrate in ever larger numbers to smaller form factors. Both Nvidia and AMD have specifically targeted this area of late with short PCB cards and conventional blower shroud designs. Many SFF's don't react well to an open graphics cooler dumping hot air into a confined chassis. Indeed, many SFF's won't even accept AIB custom cards that butt up against the extreme limit of the ATX spec (312mm in length), nor accept 2.5 and 3 slot cooling.

 

[3dilettante]

I think that really only applies when the cards in question start breaching the PCI 300W spec.

There are signs that certain cards get tantalizingly close, and some like Fury X under Furmark where it is strongly indicated that they do.
The specification is somewhat forgiving in terms of what it takes to be considered operating above 300W non-transiently, less so than TDP would be.

It should also be noted that aside from noise, AMD's current flagships are in much the same boat regarding overclocking headroom.

The noise would be primary impetus for much of the criticism in the past and in this thread.
Another would have been the questions concerning quality control for the 290's fans, given the later patching to match Powertune's offsets to the measurable real-life differences in how the fans responded to changed controller outputs.
AMD can't seem to get a break on quality control for their coolers or someone else's, given what happened with Fury X.

There's more margin for error in the 970's power band, when a cooler doesn't need to fight for every CFM possible in smaller time increments. There can be more physical slack on the part of the case and cooler's heatsink mass as well.

 

[Deleted member 13524]

I have two MSI 290X Gaming for Crossfire. Even after investing in a high end case with two additional fans in the side, the setup isn't cool at all. Pun not intended, but appropriate.

Even with a cold room, Witcher 3 often takes the top card towards 94°C, and it throttles down from there.
Though Witcher 3 seems to heat my cards a lot more than anything else I throw at them. Even more with Hairworks enabled.

 

[Moloch]

I googled it out of curiosity and that seems to be a common enough problem with the MSI 290X cards, I believe that's a situation where the reference cooler might be a better choice.
You could always water cool them
I also noticed when I was playing the Witcher 3 it did a very good job of utilizing my R9 290, GPU load was pegged and temps were a bit higher than in other games. While playing it I would also yank out the bottom case filter for my nzxt H2 as the filter is restrictive enough to cause my PSU fan to spin all the way up making quite the racket.

 

[Esrever]

I think for crossfire, the blower design is actually better unless you have a motherboard that can leave a space between the cards. The first card blocks so much of the 2nd card's cooling if they are next to each other and the air is circulated inside the case.

 

[Frenetic Pony]

Greenland (and assumedly the two other Arctic Islands GPUs) to be built on GF 14nm, launch in late summer. http://wccftech.com/amd-greenland-14nm-production-q2-2016/

I suppose "Late Summer" means august. Keep in mind GF 14nm is the same as Samsung 14nm, which is technically smaller in transistors terms than TSMC's 16nm. But has also had more yield troubles than TSMC, and that's for Samsung, which came up with the process and has already been delivering chips for a while.

 

[Kaotik]

Greenland (and assumedly the two other Arctic Islands GPUs) to be built on GF 14nm, launch in late summer. http://wccftech.com/amd-greenland-14nm-production-q2-2016/

I suppose "Late Summer" means august. Keep in mind GF 14nm is the same as Samsung 14nm, which is technically smaller in transistors terms than TSMC's 16nm. But has also had more yield troubles than TSMC, and that's for Samsung, which came up with the process and has already been delivering chips for a while.

The original article mentions the cards to be in production for computex, so should release there in June

Also, Samsung will also do AMD GPUs http://english.etnews.com/20151222200002
Samsung will supposedly start producing Greenland after the new year

 

[Jawed]

Samsung Electronics secured AMD as its toll manufacturing (foundry) customer for semiconductor chip and will produce AMD’s next graphic processing unit (GPU) chip starting from New Year.

"from new year" looks like an auto-translation goof for something like "next year", tallying with

According to industries on the 21st, Samsung Electronics’ System LSI Foundry Business Department will start mass-producing AMD’s new GPU ‘Greenland (development code name)’ along with Global Foundry (GF) starting from 2nd quarter of 2016.

 

[Deleted member 13524]

Anandtech reports that Gemini, the dual-Fiji card, had its release postponed to Q2 2016. The original release date was November/December 2015.
The reason they state for the delay is to come out at the same time as the VR HMDs.

I think this supports my theory that Fiji will continue to be AMD's flagship throughout 2016 and the two new 14/16nm GPUs will be focused on power efficiency, one of them replacing Hawaii in its performance category and the other being something that approaches Tonga but is very notebook-friendly (~75W in the desktop version, 50W in the notebook version).

 

[Frenetic Pony]

Anandtech reports that Gemini, the dual-Fiji card, had its release postponed to Q2 2016. The original release date was November/December 2015.
The reason they state for the delay is to come out at the same time as the VR HMDs.

I think this supports my theory that Fiji will continue to be AMD's flagship throughout 2016 and the two new 14/16nm GPUs will be focused on power efficiency, one of them replacing Hawaii in its performance category and the other being something that approaches Tonga but is very notebook-friendly (~75W in the desktop version, 50W in the notebook version).

That doesn't make any sense, 14nm costs less than 28nm to produce and offers numerous other benefits. It's just that yields for the Fury X are still low, comprehensible as it relies on both HBM production and the maximum die size you can go with 28nm. It's delayed because of production problems, not because of a new strategy that could only hurt AMD. Releasing it "with the release of VR headsets" is just good PR spin.

"from new year" looks like an auto-translation goof for something like "next year", tallying with

So, August it is I guess. With the late release of the Fury X2 or whatever that makes sense. With Nvidia continuing to bring out new variations of Maxwell it doesn't sound like they'll be hitting the first half of 2016 with Pascal either. Unless they'll be releasing a compute only focused GP100 and saving binned cards for gaming until they get enough stock.

 

[ImSpartacus]

That doesn't make any sense, 14nm costs less than 28nm to produce and offers numerous other benefits. It's just that yields for the Fury X are still low, comprehensible as it relies on both HBM production and the maximum die size you can go with 28nm. It's delayed because of production problems, not because of a new strategy that could only hurt AMD. Releasing it "with the release of VR headsets" is just good PR spin.

So are you anticipating a relatively widespread rollout of 14nm AMD GPUs in the mid & low end as long as they use GDDR5 (or a derivative)? e.g. Something to replace Hawaii, Tonga & Pitcairn, but without HBM.

 

[Deleted member 13524]

14nm costs less than 28nm to produce

I beg your pardon?

It's just that yields for the Fury X are still low

Citation needed.

 

[Entropy]

I beg your pardon?

Per transistor. Cost per wafer is up, as is design costs.

 

[Frenetic Pony]

Per transistor. Cost per wafer is up, as is design costs.

Aye, the cost per transistor is down only 25% compared to 28nm I believe. A paltry sum compared to what used to scale for a nigh doubling in potential density. Regardless there's no reason whatsoever either AMD or Nvidia would go to the trouble of moving to a new node, and in AMD's case making architectural changes, for low end and low profit margin GPUs only. Greenland is the replacement for Fiji and will be out... whenever.