[Rumor (Various)] AMD R7/9 3xx / Fiji / Fury

[tential]

Fury Pro will be an interesting card, hopefully that triple fan cooler is good. The Fury X costs $100, but that's really only 18% more. For your extra $100 you get the CLC and 14% more shaders if the 4096/3584 split holds up. Depending on how they test, this might be one of the few times where the top end card is actually the better value and not just the better performer.

I'm getting wce no questions asked if it's 100 price difference. I think you'd be stupid not to. Like you said the benefits are big and amd cards get better with time. That's the thing people will judge off launch and initial game reviews.

These cards will be good, but I expect far better performance down the line vs the 980ti. If you hold onto cards I expect the fury x to be a better buy.

For me the fury x at 650 has too many positives to ignore. Will buy and attempt to eBay first but I'll probably keep it if these performance numbers hold up. Either way the 980ti is a no go. They reference cooler is a joke. I can't believe people hate the r9 290x reference cooler but are ok with using a 980ti cooler.

 

[tential]

I can only wonder what the comments would be if nV was the first to HBM1, and AMD would've had decided to wait until HBM2 next year.


Ahh, reviews, where are you?



EDIT: That Fiji X2 board looks amazing. The VRM area seems much less than what 295x2 needed:

This time at least the VRM should be watercooled too, considering how compact everything has become.

Lol amd late to market again.

If amd is first with something though it's lol amd needed it to compete.

Personally I don't care what you need to compete. Amd fits my needs by having a card perform better over time. If Nvidia cards got faster overtime vs amd I'd be happy but my 7950 just has gotten more value gaming wise vs my nvidia counterpart card while costing far less. If I resold cards though and purchased more frequently and played games day 1 I'd get nvidia though. Nvidia has a clear advantage on day 1 of game releases.

 

[looncraz]

I just can't believe 390X is going far with Fuji so close in price. I can't see any reason to buy 390X unless the price drops a lot.

The lower end Fuji cards should kill the 390X, unless something else is at play.

My predictions for the Fiji cards:

Nano: 48CU, 25% faster than 290X, 175W
Pro: 56CU, 35% faster than 290X, 225W
XT: 64CU: 55% faster than 290X, 300W

TDPs are max, not gaming load, and are slightly better than nVidia at the same performance level.

I also expect the XT to have about 20% overclocking headroom - less than the 980Ti, but both should be about equal with maximum overclocks, with the maximum overclocked Titan X just ing out a narrow lead.

 

[xthetenth]

I think Nano's going to run at lower clocks, and have more EUs than a proportional scaling would indicate to get those power savings.

 

[chimaxi83]

And? FP, EDO, SDR, DDR1-4, GDDR1-5 had zero difference. But somehow HBM changes all that? 4GB is 4GB nomatter what type of memory. I can guarantee you it was the original plan to only ship it with 4GB. But thats because Hynix couldnt deliver.

Can you cite a source for your "guarantee", or anything in this post? I'll wait.

 

[The Alias]

One thing you guys are discounting is that gcn usually does worse than maxwell and Kepler in fire strike. Usually if and matches nvidia in fire strike they're like 10% faster than them. So in my mind I'm thinking the fury is actually the 10% faster card and its $100 cheaper!

 

[crashtech]

That Fiji X2 card, is it still going to act like two Crossfired GPUs?

 

[The Alias]

And? FP, EDO, SDR, DDR1-4, GDDR1-5 had zero difference. But somehow HBM changes all that? 4GB is 4GB nomatter what type of memory. I can guarantee you it was the original plan to only ship it with 4GB. But thats because Hynix couldnt deliver.

you fail to remember that Macri said they're implementing memory usage optimizations so that 4gb will not be nearly as much of an issue

 

[jackstar7]

That Fiji X2 card, is it still going to act like two Crossfired GPUs?

In DX11, probably.

DX12... unknown.

 

[Bobisuruncle54]

you fail to remember that Macri said they're implementing memory usage optimizations so that 4gb will not be nearly as much of an issue

He also has no evidence for his assumptions either.

 

[IEC]

And? FP, EDO, SDR, DDR1-4, GDDR1-5 had zero difference. But somehow HBM changes all that? 4GB is 4GB nomatter what type of memory. I can guarantee you it was the original plan to only ship it with 4GB. But thats because Hynix couldnt deliver.

{Citation needed}

 

[iiiankiii]

One thing you guys are discounting is that gcn usually does worse than maxwell and Kepler in fire strike. Usually if and matches nvidia in fire strike they're like 10% faster than them. So in my mind I'm thinking the fury is actually the 10% faster card and its $100 cheaper!

yeah. If it's 10% faster at stock clocks and overclock gives another 10%, it should match an overclocked Titan X. We shall see.

 

[krumme]

2 x 8pins and very short length. Nice.

 

[Despoiler]

And? FP, EDO, SDR, DDR1-4, GDDR1-5 had zero difference. But somehow HBM changes all that? 4GB is 4GB nomatter what type of memory. I can guarantee you it was the original plan to only ship it with 4GB. But thats because Hynix couldnt deliver.

Yes it does. The memory strategy behind GDDR has been partially to increase the number of chips for more size, but also for more bandwidth. Essentially the size came with the bandwidth so the strategy was to load up as much of the VRAM as possible to maximize the bandwidth. Just because a game will load over 4GB of memory does not mean it's actively being used. That is Macri's findings when working on Fury's memory. This is aside the fact that we've seen some engines just plain have better memory management with the same fidelity compared to ones that use a lot of memory. That leads me to believe the game dev's have more responsibility to any size limitation issues than the hardware manufacturers.

As far as the HBM1 4GB limitation, SK Hynix split the HBM spec into version 1 and version 2 for development/availability purposes. I did get my hands on the JEDEC spec and it has nothing about HBM1 or HBM2. It's just HBM that supports upto 8Hi stacks. There is no limitation on the number of stacks arranged horizontally.

He also has no evidence for his assumptions either.

I don't know about where you work, but where I do people with engineering backgrounds don't tend to do work based off of assumptions without evidence. Not to mention you provide no evidence to support your "argument." That's the burden of proof logical fallacy. Your statement is flat out ridiculous.

 

[Elixer]

'Moor' stuff to read:

The AMD Caribbean Islands Family: Not Just a Rebrand
While AMD’s Fury X provides a halo for the Radeon brand, AMD still has other new
graphics cards in the Radeon 300 series including the following air cooled cards:
 R9 390 & R9 390X each with 8GB of GDDR5 and 275W for 4K gaming
 R9 380 2 or 4GB of GDDR5 and 190W for 1440p gaming
 R7 370 up to 4GB of GDDR5 and 110W for everyday gaming
 R7 360 up to 2GB of GDDR5 and 100W for everyday gaming
AMD has been hard at work over the past year-and-a-half optimizing and re-architecting
the microcontrollers within the ASICs themselves. Combined with the improvements to
their manufacturing process, AMD has been able to squeeze more performance out of
each of their cards and increase performance while maintaining the same price tier as
its predecessor.
 The R9 390 and R9 390X replace the R9 290 and R9 290X and are both 300
GFLOPS faster than their predecessors (5,100 GFLOPS and 5,900 GFLOPS
respectively) without increasing power in typical workloads.
 The R9 380 also benefits from the maturing of the 28nm process technology and
AMD’s optimizations. It gains roughly 200 GFLOPS in performance: from 3,290
GFLOPS to 3,480 GFLOPS in compute performance.
 The R7 370’s compute capability of 2,000 GFLOPS is also faster by 200
GFLOPS than its predecessor’s (R7 265) 1,800 GFLOPS.
 The R7 360 has a compute performance of 1,610 GFLOPS, slightly more than
the 1,536 GFLOPS of the R7 260.
In all cases, AMD increased performance and also added many features that previous
generations did not have. Some of those features are enabled through the driver and
others are done in hardware. But all of the GPUs listed above will support DirectX 12,
Vulkan, and Mantle graphics APIs.

http://www.moorinsightsstrategy.com/research-brief-amd-brings-new-value-to-radeon-2/

 

[ShintaiDK]

{Citation needed}

Like from Hynix?

HBM1 got downgraded twice before we got what we do today.

(Expectations in 2013)

And they believed in 8 hi stacks till last year.

In 2015 it looks like this:

(Note 2HI stacks are half speed).

 

[flopper]

That Fiji X2 card, is it still going to act like two Crossfired GPUs?

DX12 for VR who knows what they can do as developers have more control over hardware.

https://www.youtube.com/watch?v=dHuDTUcJURU

 

[Elixer]

OMG, Quantum, Want!!!!

I wish we could see what is inside of it...

 

[coercitiv]

I wish we could see what is inside of it...

Go one page back.

 

[monstercameron]

I wish we could see what is inside of it...

watch the video!

 

[Erenhardt]

Is this steam machines?
Water - check,
Lots of heat - check

 

[ShintaiDK]

Asrock MiniITX board inside.

A stripped version of the Asrock Z97E-ITX

 

[IEC]

snip.

No. Citations for your unsourced assertions, not the known Hynix information.

 

[Elixer]

Go one page back.

D'oh, dunno how I missed that...

 

[looncraz]

http://cdn3.wccftech.com/wp-content/uploads/2015/06/AMD-Radeon-R9-Fury-X-3DMark-Firestrike.png

http://cdn4.wccftech.com/wp-content/uploads/2015/06/AMD-Radeon-R9-Fury-X-3DMark-11.png

Not solid sourcing but at least something.

Fury X 10% better than 980ti at stock? We know 980ti has headroom for OC, need to see how much Fury X has now.

If that's true, the 390X isn't likely to be a direct rebrand unless there are some meaningful driver enhancements (though it only needs like 5% or so... ).

The previously 'leaked' 390X score was out of line with the R9 390 score, which many noted as odd. And the one owner of the R9 390 posted results that indicated a noteworthy per clock performance increase in addition to lower power draw...

The 390X was claimed to score 5570/2779, but this shows 6038/2967, which is a very significant difference ;-)

Most importantly, it is showing about 12% greater performance with a mere 5% overclock (and even extra 20% on the memory clocks don't bring 7% extra performance - more like 2~3%).

That would make sense with the price, and the earlier leaks of the R9 390.

Looks more likely to be a refresh then :thumbsup:

Very slight performance per clock increase, much more meaningful power efficiency increase would be my guess.