GPU Now this could be big: R9 Nano!

theGryphon

Airflow Optimizer
Original poster
Jun 15, 2015
299
237
I don't think anyone saw this coming. :eek:
Half the power of 290X, but more powerful? Very unlike AMD, lol. Could put Nvidia in trouble if this becomes the trend. Real competition will be welcome! :)

 

Vittra

Airflow Optimizer
May 11, 2015
359
90
Careful, they said 2x performance/watt.

It could be the same TDP envelope as the 290x, but twice the power, or half the power, with the same performance of the 290x.

Still great news, though I can fit a Fury X in my M1, so lets see how this all plays out. :)

For those who don't know, Fiji has a 4 card lineup:

Fury X - $650 - June 24th
Fury - $550 - July 14th
Fury Nano - Price/Availability TBA (late summer)
Fury Dual-GPU - Price/availability TBA (fall)
 

PlayfulPhoenix

Founder of SFF.N
SFFLAB
Chimera Industries
Gold Supporter
Feb 22, 2015
1,052
1,990
"2x perf-per-watt" is in the neighborhood of the sorts of efficiency gains nVidia got with Maxwell. But nVidia was already providing better power efficiency before Maxwell, so they'll still retain a lead - just, a much smaller one.

It's interesting (and great) that AMD seems to be embracing more compact form factors, but we'll have to see how that translates in performance, temperature and noise.
 

theGryphon

Airflow Optimizer
Original poster
Jun 15, 2015
299
237
Careful, they said 2x performance/watt.

It could be the same TDP envelope as the 290x, but twice the power, or half the power, with the same performance of the 290x.

Still great news, though I can fit a Fury X in my M1, so lets see how this all plays out. :)

For those who don't know, Fiji has a 4 card lineup:

Fury X - $650 - June 24th
Fury - $550 - July 14th
Fury Nano - Price/Availability TBA (late summer)
Fury Dual-GPU - Price/availability TBA (fall)


Well with such a small PCB and only air cooling, no way they can have the same TDP as 290X in there. ITX size has a clear message: low TDP. So, "most" likely, it should be around half the TDP of 290X with similar graphics power.

This is a huge deal for AMD if true. Unfortunately it's late in the game, and they couldn't come up with a whole line based on Fiji and similar power characteristics.
I believe the culprit is low yields for HBM. If you can't supply sufficient volume (at economic rates), how can you possibly have a product line? I believe they're doing what they can, and being on the shelves (albeit with older GPUs) is better than paper launches (which don't make any money, and turn-off consumers).

AMD seems to have caught up on the R&D side, but not yet on the supply chain side, which is not within their hands. 28nm node should die already for high-end GPUs. As the fabs reach sufficient yields, both AMD and NVIDIA can start reaping the benefits.

It seems though AMD benefited with this over-stretched lifetime of 28nm node. Counter-intuitive, I know, but imagine if Maxwell came on 16nm!!! AMD would be wiped-off for good. This gave AMD the time to catch-up in terms of power efficiency, and their fab partners to catch up in terms of node size.

Next year will be very interesting with almost everything leveled between AMD and NVIDIA on the HW side.
 
  • Like
Reactions: Phuncz

Vittra

Airflow Optimizer
May 11, 2015
359
90
Actually, AMD playing catchup has partly been because they've been hamstrung by 28nm ;). Although these plans are years in the making, I have no doubt the last two gens had planned for die-shrinks and subsequently got altered to accommodate TSMC's failings.
 

theGryphon

Airflow Optimizer
Original poster
Jun 15, 2015
299
237
Actually, AMD playing catchup has partly been because they've been hamstrung by 28nm ;). Although these plans are years in the making, I have no doubt the last two gens had planned for die-shrinks and subsequently got altered to accommodate TSMC's failings.

Nah, I don't buy that for AMD. See, what NVIDIA did? Maxwell was clearly designed for 16nm, but they made some changes (potentially involving crippled DP and GPGPU performance) but they made it happen on the 28nm. AMD has been riding the same GPUs for 4 years now. If they had something on the pipeline, they would bring it on 28nm somehow.

The only exception for AMD is Tonga, which curiously is a small GPU. I don't know why they wouldn't double it up and build the 300 series on that, instead of serving the same old GPU for the zillionth time. Funny thing is they're also shamelessly changing the codenames (Hawaii, Grenada, etc) as well as brand names of the same thing. I mean, come on, it's like lying to yourself.

Tonga was a good direction to go but they didn't go there, unfortunately.

Anyway, 300 series is already history now. They should look ahead and bring a fully HBM-based line-up for 400 (?) series.
 
  • Like
Reactions: Phuncz

Vittra

Airflow Optimizer
May 11, 2015
359
90
Correct. Arctics Islands (4xx) should see a top-down new architecture, most likely based on HBM2.

We'll have to agree to disagree regarding AMD and 28nm. :)
 

rawr

SFF Lingo Aficionado
Mar 1, 2015
137
10
Oh my god they made a top tier ITX air cooled card! The cooler design is also very attractive, imo.

Perfect for my next project...
 

theGryphon

Airflow Optimizer
Original poster
Jun 15, 2015
299
237
Correct. Arctics Islands (4xx) should see a top-down new architecture, most likely based on HBM2.

We'll have to agree to disagree regarding AMD and 28nm. :)


I can actually see what happened: they're lacking R&D power and they cannot lead two projects (Fiji and "large" Tonga) at the same time. Fiji obviously and rightfully and correctly took the lead, but volumes weren't enough to create a full line-up on that. It's also possible that they couldn't foresee low volumes for Fiji (due to HBM) and that their plan was actually to have a fresh line-up based on HBM. Eventually, unfortunately for AMD, they had to warm up the Hawaii GPUs, since, well, they had to have a line up on the shelves.

That's why I'm saying that the long-lived 28nm gave time to AMD to actually catch up, and with HBM2 and 14-16nm, things should look much brighter next year.
 

rawr

SFF Lingo Aficionado
Mar 1, 2015
137
10

Wow. Properly short as well; looks like it's at the magical 17cm mark, without any cooler overhang.
 

Vittra

Airflow Optimizer
May 11, 2015
359
90
They made several references to it being a 6" card in the livestream, so if it's exact, 152.4mm. I'd give it some leeway and say 155mm.
 

PlayfulPhoenix

Founder of SFF.N
SFFLAB
Chimera Industries
Gold Supporter
Feb 22, 2015
1,052
1,990
I'm eagerly awaiting reviews and real-world use/testing on that one. Having AMD come out with a good performer in the ITX graphics space would be a huge boon for many SFF builders.
 

iFreilicht

FlexATX Authority
Feb 28, 2015
3,243
2,361
freilite.com
Really looking forward to this! Not only will it help AMD to get back in the game but it will also step up the performance in uSFF systems, which is great! Not sure why they choose to make the card even shorter, tough. 170mm would've been perfect, why go even shorter?

Nah, I don't buy that for AMD. See, what NVIDIA did? Maxwell was clearly designed for 16nm, but they made some changes (potentially involving crippled DP and GPGPU performance) but they made it happen on the 28nm. AMD has been riding the same GPUs for 4 years now. If they had something on the pipeline, they would bring it on 28nm somehow.[...]

Actually, if this HBM stuff is what AMD was working on in the past few years, that makes a lot of sense. As HBM incorporates the main Graphics memory on the actual GPU die, it would be very hard to just upscale this to 28nm. There are good reasons why something like this has not been done before and modern Processors use multiple levels of cache: Because it wasn't feasible. Too expensive, too complicated, too high transistor count.
I don't know what AMD has done to make it possible to put a few Gigabytes of memory on a die, this has never been done before, but I would think that 16nm is maybe part of it.
 
  • Like
Reactions: Phuncz

iFreilicht

FlexATX Authority
Feb 28, 2015
3,243
2,361
freilite.com
Oh it also seems like Project Quantum is powered by an Intel processor. Video link, the board you can see in the timelapse seems to be a modified Asrock X99E-ITX/ac. It would be so awesome to work on a project like this!
Also very interesting is the way I/O is solved. The riser they use for the GPU flips the mainboard head over heels from front to back so the mainboard I/O you see on the back is mainly from the front USB headers and the single Ethernet port while the GPUs I/O is completely unmodified.
Another thing I noticed is the beefy power jack for the external PSU (at least that's what it looks like).
Really cool thing altogether, and while not that small in general, it really will push the Performance/Litre mark if they really use a Dual-GPU card in there.
 

EdZ

Virtual Realist
May 11, 2015
1,578
2,107
As HBM incorporates the main Graphics memory on the actual GPU die
It doesn't. The dies are completely separate, but sit on a shared 'interposer' (think a very fine scale PCB, fabricated in the same way as a large IC). If you really wanted you could put a GPU and a bunch of GDDR5 dies on an interposer, though there's not much reason to.
 

iFreilicht

FlexATX Authority
Feb 28, 2015
3,243
2,361
freilite.com
It doesn't. The dies are completely separate, but sit on a shared 'interposer' (think a very fine scale PCB, fabricated in the same way as a large IC). If you really wanted you could put a GPU and a bunch of GDDR5 dies on an interposer, though there's not much reason to.
Ah, that makes more sense. But the point still stands, maybe that sort of thing wasn't possible with the processes available to 28nm manufacturing.
EDIT: Bullshit, AMD already did that with the Radeon E8860, as SaperPL pointed out on [H].

Wow. This is so big and yet so small. I hope that card will have its connectors on the side, though. And it looks slightly longer than 170mm, which is a bummer. I mean it's still great that they've manage to make it this damn small, but if it was just a little bit shorter, it would fit in so many existing mITX cases, many of which would even be able to hold the radiator.
 
Last edited:
  • Like
Reactions: Phuncz

Phuncz

Lord of the Boards
SFFn Staff
May 9, 2015
5,827
4,902
VERY interesting about the power connectors being on the side instead of on the top, this can make SFF cases even smaller. And with a single dual-GPU card, mITX is looking better than ever ! Thank you AMD for sparking interest !

Also funny how memory is now contained within the package while the stuff around the core is now VRM's mainly. In contrast to the standard GPU package with RAM dies around it.