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

[RussianSensation]

You can probably get a couple more years out of the 2 you have if you wanted. Getting a little long in the tooth feature wise, but performance is still hanging in there. The compute is just the cherry on top. A free cherry with no premium charged.

Maybe but not for the reasons you think. I'll probably have very limited gaming time in the next 2 years due to a shift in job/educational priorities. Hardware wise, they won't be good enough though as I bought a BL3200PT on June 5th, which will hammer 7970s. I might time an entire PC rebuild with Skylake-E/14nm HBM2 around summer/fall 2017. Otherwise, I would have upgraded right now. We'll see.

BTW, an awesome AT member with an NV card tipped me about a Legit Review test where they overclocked a 980Ti on a reference blower. The noise level results aren't pretty, let's just put it mildly. Good thing AMD didn't even bother with a blower design on a 300W TDP card or they would have failed hard. :biggrin:

"When you overclock the card though by raising the power limits you’ll see than the fan speed jump up to over 60% (2800RPM) and you’ll hear the difference."
Source

Rumor has it AMD originally developed a prototype blower fan for their 300W TDP Fiji XT card but they decided it wasn't going to fit into a miniITX case.

I think for Pascal and 14nm Arctic Islands, AMD and NV should go AIO CLC for reference flagships and have 2nd tier with air like AMD is doing with Fiji XT (air) and Fiji PRO.

 

[StinkyPinky]

Impressive specs.....on paper (if true). The bigger test is how it goes in the real world.

I hope for AMD this generation is a home run because they really need it. Competition is good.

 

[ozzy702]

7970 had a good run - outlasted 580/680/770/960/285 and made the OG Titan look bad, while being a dual-purpose DP + mining champ. :thumbsup: I don't think I'll ever own a card as impressive as the 7970 was from how it balanced all aspects of a GPU well.

I have to agree 100%. I owned several 7970's and 7950's over the past few years and they sure have held up well and if you were smart paid for themselves many, many times over.

 

[tential]

If you were smart and mined.
I just enjoyed my card. My 7950 lived up to and exceeded expectations. Very happy I did not get the gtx 680 which was far more money. Especially after the last graphs posted if I remember right.

 

[JDG1980]

And to the others that asked why I would even need one for a HTPC and why not get a APU?? Well because I am using a Cheapie AM1 5350 in the living room as a HTPC and wanna do some light gaming in that living room at more then 20-25FPS at 720p and would rather play at 1080p at 30-60FPS and I dont want to break the bank or run up the electric bill doing so.

In that case, given your budget limit of $100, you should purchase a GTX 750. Newegg has a highly rated EVGA model for $89.99 after rebates, and if you need to stay under $100 without using a rebate, there's a Zotac model for $99.99.

The GTX 750 surpasses even fully enabled Bonaire (R9 260X) in most gaming benchmarks. It costs the same or less than a R9 260X, and uses far less power. And the upcoming R7 360 is not going to be a fully enabled Bonaire; the OEM site and the leaks released so far indicate it's going to be a cut-down part with 768 shaders, equivalent to the R9 260 (non-X). Although AMD has the advantage in perf/$ at certain price points if you don't care about power consumption, the $99 price point is one where Nvidia has the upper hand.

Older drivers had colorspace issues with HDMI on Nvidia cards which caused some problems for HTPC users, and this was one of several reasons why AMD cards have often been recommended in the HTPC space. However, this particular issue has now been fixed.

 

[richaron]

I'm feeling more and more that AMD's marketing/financial tactics are hurting themselves. Case in point; their Tonga strategy. They hold back a 4G card on release because of the 280X (must differentiate), we don't get the 4G card until almost a year later in the 'next' generation.

But I wanted a 4G card last year. Now, almost a year on, surely they'll sell me a full Tonga XT? Doesn't seem so. Can't eat into sales of the 290/390... Maybe next gen they'll sell me a full Tonga on 28nm whilst everything else is on '14'nm, gotta maximize return on investment after all.

Perhaps it's the only viable tactic ATM, I don't know. But it's not the sort of thing which makes enthusiasts like myself gush with brand loyalty and bring in more customers.

 

[3DVagabond]

I'm feeling more and more that AMD's marketing/financial tactics are hurting themselves. Case in point; their Tonga strategy. They hold back a 4G card on release because of the 280X (must differentiate), we don't get the 4G card until almost a year later in the 'next' generation.

But I wanted a 4G card last year. Now, almost a year on, surely they'll sell me a full Tonga XT? Doesn't seem so. Can't eat into sales of the 290/390... Maybe next gen they'll sell me a full Tonga on 28nm whilst everything else is on '14'nm, gotta maximize return on investment after all.

Perhaps it's the only viable tactic ATM, I don't know. But it's not the sort of thing which makes enthusiasts like myself gush with brand loyalty and bring in more customers.

I remember when Hawaii was released there was also talk of a 256 bit 2GB Tahiti being released as well. It didn't happen. Seems like they had this chip for a long time before it was released. just another very confusing release from when R. Read was there. Besides the hardware, I'm curious to see how this release goes.

 

[JDG1980]

But I wanted a 4G card last year. Now, almost a year on, surely they'll sell me a full Tonga XT? Doesn't seem so. Can't eat into sales of the 290/390...

AMD's refusal to sell full Tonga to the PC market, even in the 300 series rebrand line, is one of their more baffling decisions. The only possible justification for this that I can think of is that perhaps AMD signed some sort of deal with Apple to give them exclusivity on full Tonga (maybe for one year?)

There are several other odd decisions, like the decision to go with cut-down Pitcairn instead of the full chip (they can't possibly be having problems with yields on a 212 sq. mm. 28nm chip, can they?) It can't be justified as necessary to fit in a 6-pin connector, since the R9 270 (with the full complement of 1280 shaders) already is able to do that. Honestly, this looks like they're trying to pinch pennies by creating a standard that even their worst trash silicon can meet. We saw the same thing with Godavari and its grossly overvolted factory BIOS settings.

Maybe next gen they'll sell me a full Tonga on 28nm whilst everything else is on '14'nm, gotta maximize return on investment after all.

At this point, I'm more cynical; if AMD remains an independent company I do not think we will ever see another full GPU lineup from them again (full lineup here being defined as 3 or more chips, like we saw with Tahiti, Pitcairn, and Cape Verde back in 2012). I think that in late 2016 or early 2017 they're going to release one new Arctic Islands GPU on FinFET+ with HBM2, probably around 350 sq. mm. with 5120 or so shaders, 250W TDP, and a price tag of $599, and rebrand everything else yet again. It's unlikely we will ever be rid of Pitcairn until and unless Samsung buys AMD and they finally have an actual budget again.

 

[JDG1980]

Fiji pro has a lower tdp than 290x. Works for me. The performance difference between pro and X might be larger than that between the 290 and 290x unfortunately for the cheapos out there.

While this isn't actually confirmed yet, nothing in here is particularly surprising. It's in line with what we expected to see. The most unexpected part, honestly, is the mention of 3-fan coolers on the air-cooled reference cards. I wonder if they're going to use a modified version of the 7990 cooler, or if they made an agreement with Sapphire to use the Tri-X as a reference part. (This wouldn't be too much of a stretch, since Sapphire and AMD are joined at the hip; Sapphire is the exclusive vendor of FirePro boards, for example.) The Tri-X, especially, would work very well; it would extinguish the noise and overheating complaints we saw with Hawaii reference cards. One concern is that triple-fan coolers don't work as well as blowers in CrossFire configurations, especially in cases with less-than-perfect ventilation. When Tom's Hardware tested two 7990 cards in tandem, they got "hot enough to cook a pork shoulder in about seven hours". To be fair, that was a card with a 375W TDP, so the 275W-300W Furies shouldn't be as bad.

One interesting thing is that if this leak is true, it makes sense of a lot of the conflicting information: 20% faster than 290X vs. 50% faster, slower than GTX 980 Ti vs. faster than Titan X.

The leak indicates that the top Fury X SKU will be about 54% faster than the R9 290X. That's believable; if it has 4096 shaders, that's a 45% increase in shader power, and HBM, driver improvements, and GCN 1.2 architectural improvements not present in Hawaii could easily make up the remainder of the difference. So where does that put it in comparison to Big Maxwell? The R9 290X measured about 70% of the Titan X's performance according to TechPowerUp. So, 0.70 x 1.54 = 1.078. In other words, we'd be looking at a roughly 7.8% edge over the Titan X. The interesting question is whether AMD is pushing their card harder at stock settings than Nvidia. If so, then aftermarket parts could take back the performance crown. Aftermarket cards are going to have a hard time beating the watercooled Fury X (what are they going to use, Peltiers? LN2?) On the other hand, the EVGA Titan X Hybrid offers a 14% boost clock increase over the stock Titan X, which could easily make up the difference between the two cards. It looks very much like Fury X is going to be battling it out with GTX 780 Ti and Titan X for supremacy. It won't be able to charge the Titan X's premium price because AMD doesn't have Nvidia's premium brand image, and the AMD card is limited to 4GB of RAM compared to three times that on the Titan X. Instead, we're probably looking at a max of $699 for the top watercooled SKU.

The "20% faster than GTX 980" rumors probably referred to the Fury (non-X) with 3584 shaders. That's a fairly substantial cut, only 87.5% as many shaders as the full Fiji chip. It would fall short of the Titan X by around 6% (0.70 x 1.54 x 0.875 = 0.94325). So this may be where the "5% slower than GTX 980 Ti" rumors came from. If 275W is really the thermal limit (i.e. FurMark), then that means not only performance but also perf/watt would only be a few percent worse than Big Maxwell. (GTX 980 Ti tops out at 277W power usage, so we can assume that its thermal limit is set at 275W as well.) Still, the fact that AMD had to resort to exotic new technology to achieve this while Nvidia did it with plain old GDDR5 is not encouraging. It's like back when Intel beat AMD badly in IPC with Conroe despite AMD having their memory controller on-die and Intel not; when Intel followed suit in Nehalem, AMD fell irretrievably behind forever. For the sake of competition, let's hope that doesn't happen here.

I'd be interested to see what performance you'd get out of the non-X Fury if you lowered the power limit to 65% or so (around GTX 980 levels). If AMD is pushing it to the margins for competitive reasons, then lowering the power limit just might result in an actual victory in perf/watt over GM204.

 

[Kenmitch]

AMD's refusal to sell full Tonga to the PC market, even in the 300 series rebrand line, is one of their more baffling decisions. The only possible justification for this that I can think of is that perhaps AMD signed some sort of deal with Apple to give them exclusivity on full Tonga (maybe for one year?)

There are several other odd decisions, like the decision to go with cut-down Pitcairn instead of the full chip (they can't possibly be having problems with yields on a 212 sq. mm. 28nm chip, can they?) It can't be justified as necessary to fit in a 6-pin connector, since the R9 270 (with the full complement of 1280 shaders) already is able to do that. Honestly, this looks like they're trying to pinch pennies by creating a standard that even their worst trash silicon can meet. We saw the same thing with Godavari and its grossly overvolted factory BIOS settings.



At this point, I'm more cynical; if AMD remains an independent company I do not think we will ever see another full GPU lineup from them again (full lineup here being defined as 3 or more chips, like we saw with Tahiti, Pitcairn, and Cape Verde back in 2012). I think that in late 2016 or early 2017 they're going to release one new Arctic Islands GPU on FinFET+ with HBM2, probably around 350 sq. mm. with 5120 or so shaders, 250W TDP, and a price tag of $599, and rebrand everything else yet again. It's unlikely we will ever be rid of Pitcairn until and unless Samsung buys AMD and they finally have an actual budget again.

Be optimistic. Probably the current node is the cause of the lack of full round this time.

Next node I'd expect more offerings from them.

Pretty sure their currently in survival mode.

Diverting R&D $'s towards CPU/APU is better spent for long term I'd imagine.

Samsung could invest without buying AMD couldn't they. Maybe custom GPU for their SOC adventures? Kind of what I picture when I read Figi Nano.

 

[Silverforce11]

Specs suggest this is a 4K beast, 128x Tonga ROPs with massive vram bandwidth and low latency.

I think it could well be ~65% above R290X at 4K, ie. scaling better at higher resolutions.

Edit: Also its quite funny to see the SAME sites peddle such wild rumors such as "Fiji only 20% faster than 980!", to "Fiji to get 8GB HBM"... all sorts of angles covered, enough so that eventually 1 is correct.

For myself, from last year when SDesign posted on their website they are working on a >500mm2 GPU for AMD, we should have all known it was gonna be massive, AMD going for the performance crown (huge GPU die + new vram advantage!). It's quite ballsy but fully expected from a sudden change in CEO, AMD needs to ditch the value brand and go for the premium performance brand image.

 

[MrTeal]

Where has it been confirmed that they are using the Tonga ROPs?

 

[boozzer]

Where has it been confirmed that they are using the Tonga ROPs?

no where, just more wccf crap. I am actually super impress by how AMD kept a lid on everything.

 

[DrBombcrater]

I remember when Hawaii was released there was also talk of a 256 bit 2GB Tahiti being released as well.

The 2GB/256-bit Tahiti came out under the name 7870LE. It was fairly quick and could have been a good card but all the ones I saw were 1000MHz core @ 1.25v and ran very hot. It was plainly just a dumping ground for low-grade salvaged Tahitis that couldn't be used anywhere else.

 

[3DVagabond]

At this point, I'm more cynical; if AMD remains an independent company I do not think we will ever see another full GPU lineup from them again (full lineup here being defined as 3 or more chips, like we saw with Tahiti, Pitcairn, and Cape Verde back in 2012). I think that in late 2016 or early 2017 they're going to release one new Arctic Islands GPU on FinFET+ with HBM2, probably around 350 sq. mm. with 5120 or so shaders, 250W TDP, and a price tag of $599, and rebrand everything else yet again. It's unlikely we will ever be rid of Pitcairn until and unless Samsung buys AMD and they finally have an actual budget again.

Could it just be being stuck on 28nm for so long and them miscalculating how far away the shrink is/was?

 

[Silverforce11]

Where has it been confirmed that they are using the Tonga ROPs?

Brand new chip, do you think they are going to use older GCN 1.1 like in Hawaii or use GCN 1.2 as a foundation? Logic.

At the least, it would be Tonga ROPs with memory compression tech. Then they can add any new GCN changes on top, to be 1.3.

 

[96Firebird]

How many times has nvidia released cards only to have amd undercut them and NV then lowers prices?

You could also ask if Nvidia has ever released a Titan at $1000 and then released a slightly slower card at $650 3 months later... The answer would be yes, the same situation happened with the Titan/780. There was no sign of the 290X then, either.

 

[Silverforce11]

You could also ask if Nvidia has ever released a Titan at $1000 and then released a slightly slower card at $650 3 months later... The answer would be yes, the same situation happened with the Titan/780. There was no sign of the 290X then, either.

That's true.

But a question for you, do you think Titan with its superior compute & extra vram (6GB vs 3GB) had a better chance of justifying/selling at $1K compared to Titan X which has 12GB vs 6GB vram as the only advantage?

Titan was well known to retain its resale value, due to its possible use as a cheap Tesla/Quadro.

 

[Justinbaileyman]

Well I will wait a couple more days before I pull the trigger to know for sure who will have the performance crown.right now I am eyeing up the xfx R7 260x 2gb short card or maybe the zotac GTX 750 2gb short card. trying to find some Benchies pitting the 2 against each other.If I find out the 360x is better I waill just wait and grab that..

 

[RussianSensation]

Case in point; their Tonga strategy. They hold back a 4G card on release because of the 280X (must differentiate), we don't get the 4G card until almost a year later in the 'next' generation.

Don't waste your time on Tonga. If you need a card around $200-220, PowerColor R9 290 is the ticket.

The "20% faster than GTX 980" rumors probably referred to the Fury (non-X) with 3584 shaders. That's a fairly substantial cut, only 87.5% as many shaders as the full Fiji chip. It would fall short of the Titan X by around 6% (0.70 x 1.54 x 0.875 = 0.94325). So this may be where the "5% slower than GTX 980 Ti" rumors came from.

1. Right, so if Fiji Pro after-market cards are $549, do you realize what that means? Possibly ~90% performance of the Titan X for nearly half the price 3 months later, a number I threw around months ago. That's pretty good if you ask me. I will say that for single GPU users, possibly paying $100 more and getting 980Ti's 20% overclocking could be worthwhile. Another way of looking at it, is someone who didn't buy a Titan X for $1K may be able to get 70% more performance with Fiji PRO CF for $100 more just 3 months later. WOW.

2. Yes the Fiji PRO is a lot more cut-down this time, but that also means AMD can price it more aggressively and differentiate Fiji XT cards more. However, if those specs are true, Fiji PRO still has the entire back-end intact (ROPs/memory bandwidth->HBM). That means at the same clock speeds, Fiji XT may only be 14% faster best case for hundreds of dollars more. If AMD pulls a GTX970 on the high-end and prices Fiji PRO at $499, it'll be a HUGE seller. That's probably what I would do in AMD's shoes because it would instantly make 980 irrelevant and also for a lot of gamers $500 is their psychological barrier for high-end GPU purchases. This strategy would allow AMD to claim indisputable price/performance on the high-end with Fiji PRO and still get the profit margins and premium performance halo their desire with Fury XT cards.

3. What difference does it make how AMD accomplished the improvements, whether it was from a new architecture, a new node, a new memory type, etc. Since we are comparing this generation, what matters is the end product. NV used a new architecture vs. AMD that keeps building on GCN which is still a very good architecture. It's just a different way to approach GPU design. For example the new Corvette Z06 is beastly but so is the new McLaren 650S, yet their engines are completely different. That's what engineers get paid $$$ for - to accomplish certain objectives using different means.

As long as AMD's next architecture doesn't flop against Pascal, I don't think we should worry that "AMD needed HBM1 to achieve perf/watt improvements." It's not like AMD is sitting still. Eventually they will debut an all new GPU architecture (post-GCN). They used VLIW for a long time, from 2006 to 2011. By December 2016, GCN will also turn 5 years old. What an amazing architecture that survived Fermi, Kepler, Maxwell. You can say AMD is way behind since they needed all-new HBM memory to keep up but another way of looking at it is NV needed 3 brand new architectures to stay in the game. Both companies do things differently.

On paper Fiji cards look like they will use more power than the 980Ti/Titan X but if AMD managed to cram DP performance, that would be mighty impressive since it'll make the card a dual-purpose product for just 40-50W more.

 

[Azix]

a $549 fiji pro won't be too impressive. Well.... I couldn't buy it. Crap ton of money for a GPU. Fiji xt for that with fiji nano for $649 and pro for < $500 would make my day. Still couldnt justify it but closer.

 

[casiofx]

On paper Fiji cards look like they will use more power than the 980Ti/Titan X but if AMD managed to cram DP performance, that would be mighty impressive since it'll make the card a dual-purpose product for just 40-50W more.

Can't AMD do the same thing like original Titan?

Put a toggle to turn on/off the DP.

IINM original Titan doesn't boost with DP on because of higher power consumption?

 

[MrTeal]

Brand new chip, do you think they are going to use older GCN 1.1 like in Hawaii or use GCN 1.2 as a foundation? Logic.

At the least, it would be Tonga ROPs with memory compression tech. Then they can add any new GCN changes on top, to be 1.3.

I don't think that they would use the same ROPs as Hawaii, but that doesn't automatically make them the same as in Tonga. AMD made major changes to the ROPs with Tonga while going to the 256-bit memory bus, which still gave them impressive pixel fillrates. Given that they're going to a much wider bus now, I don't think it's a given that they haven't made further changes to the ROPs from Tonga.

 

[tg2708]

The 4GB HBM is quite interesting, we need to wait and see how well this new tech handles inflated vram requirements. I hope its not a bottleneck for such a promising card.

 

[stahlhart]

Threads merged. We don't need two parallel Fiji/Fury discussions here.
-- stahlhart