Fury Nano is Full Fiji

[richaron]

Fiji seems to be ROP bottlenecked, which hurts its geometry performance. I think the same is true of Tonga, since Tonga is basically 1/2 of Fiji in all stats (aside from the memory interface). 64 ROPs aren't enough to back up 4096 SPs, and I think 32 ROPs aren't enough for 2048 SPs. If full Tonga only offers a minor performance improvement over the cut-down version, then it might not be worth putting out a separate SKU for it. It's worth pointing out that the 7970 only outperformed the 7950 by a few percent when clock speeds and RAM speeds were matched, and those Tahiti-based chips had the same shader/TMU/ROP counts as Tonga XT and Tonga PRO, respectively.

Interesting comparison. And perhaps I'm behind the times in the rumour department, but isn't Tonga XT supposed to have 384bit memory interface? Half again as much theoretical bandwidth, but who knows about performance increase.

 

[tviceman]

The Nano isn't going to be cheap. It's aiming to be the perf/W king, and they will charge a premium for it. The Fury (non X) will be the perf/$ flagship, and the Fury X is the performance flagship.

Wide and slow is the best way to get efficiency; not surprised that it will be a full die.

First off, if Nano is full functional, that is amazing yields would be that good to release 2 desktop sku's without any fused parts on a 600mm2 chip. Awesome stuff there. That being the case, AMD said, during their big 300 series / Fury X reveal, that Fiji was "the word's fastest" and "world's most efficient" GPU. Both claims are irrefutably untrue. I'm not saying the AMD rep is wrong about about Nano's functional units, but I am saying AMD's upcoming performance metrics claims about Nano may be exaggerated.

I have no doubts Fury Nano will be very, very efficient, but their claims of 2x perf/w over 290x are asterisk'd with "up to" claims, and even then, are probably being stretched a bit. Nvidia did the same when both GM107 and GM204 came out when comparing to their previous products. But given that companies (all companies, whether it's Nvidia, Intel, AMD, Qualcomm, Apple, etc.) like to round up their estimations of absolute best case scenarios, I don't see AMD topping GM204's perf/w metrics, especially at 1440p.

EDIT: I also want to point out that there is nothing that I can see preventing Nvidia (or it's partners) from bringing a small GTX 980 to market. The 970 comes in mini-ITX form factors and most 970's end up consuming every bit as much wattage as 980's.

 

[railven]

Hmmmm...How could that power lock this card? 1 8pin/6pin connector? Perhaps there is a defect in the HBM and it is also clocked down, which is what drives the total clocks for the card down?

We're all focusing on the core. Could be the memory? Who knows.

But like I said in the review thread, and now it seems more confirmed, it is the more interesting of the two remaining single chip cards.

 

[tviceman]

Hmmmm...How could that power lock this card? 1 8pin/6pin connector? Perhaps there is a defect in the HBM and it is also clocked down, which is what drives the total clocks for the card down?

We're all focusing on the core. Could be the memory? Who knows.

But like I said in the review thread, and now it seems more confirmed, it is the more interesting of the two remaining single chip cards.

Even if an AMD rep confirmed it, I find it astonishing that an extremely dense 600mm2 chip can yield enough full functioning dies for two desktop SKU's at the same time while leaving only 1 SKU with fused off parts. Perhaps 28nm is yielding out of this world now.

 

[railven]

Even if an AMD rep confirmed it, I find it astonishing that an extremely dense 600mm2 chip can yield enough full functioning dies for two desktop SKU's at the same time while leaving only 1 SKU with fused off parts. Perhaps 28nm is yielding out of this world now.

It is a super old mature process. If I had to guess by inventory of another product, the neutered chips aren't as common as a full chip.

I really want a Fiji based videocard (ahaha).

 

[maddie]

Even if an AMD rep confirmed it, I find it astonishing that an extremely dense 600mm2 chip can yield enough full functioning dies for two desktop SKU's at the same time while leaving only 1 SKU with fused off parts. Perhaps 28nm is yielding out of this world now.

Why do you find the idea so shocking?

Even if we assume that fully functional dies sell evenly as the cut-down die then that is a yield of 1/3 rd each spread between Fury X, Fury Nano and Fury air.

Realistically assuming that the lower priced card [Fury air] will sell more, this skews the needed fully functional die harvest to less than 2/3.
Surely we should think that is possible?

 

[jpiniero]

Even if an AMD rep confirmed it, I find it astonishing that an extremely dense 600mm2 chip can yield enough full functioning dies for two desktop SKU's at the same time while leaving only 1 SKU with fused off parts. Perhaps 28nm is yielding out of this world now.

Apple is probably putting a cut down Nano as the higher end option on the next iMac.

 

[LTC8K6]

Even if an AMD rep confirmed it, I find it astonishing that an extremely dense 600mm2 chip can yield enough full functioning dies for two desktop SKU's at the same time while leaving only 1 SKU with fused off parts. Perhaps 28nm is yielding out of this world now.

Yes, supposedly we couldn't have a full Tonga card because there weren't enough good dies to go around with Apple using them.

In fact, we still can't have a full Tonga card, apparently.

 

[IntelUser2000]

First off, if Nano is full functional, that is amazing yields would be that good to release 2 desktop sku's without any fused parts on a 600mm2 chip.

Ok, there's more than one metric of "working" chip. Do all the dies work at the full size(what you are wondering)? Do all the dies work at the frequency we want? It's not just one metric.

So you can have 4 different SKUs based on the combinations.

1. Full die, full frequency
2. Full die, partial frequency
3. Partial die, full frequency
4. Partial die, partial frequency

 

[maddie]

Yes, supposedly we couldn't have a full Tonga card because there weren't enough good dies to go around with Apple using them.

In fact, we still can't have a full Tonga card, apparently.

This was clarified a little while ago. Can't remember where I read it, but the AMD rep said that the addition of full Tonga would confuse product price/placement between present Tonga and Hawaii. Its absence is a conscious marketing decision by AMD.

 

[jpiniero]

This was clarified a little while ago. Can't remember where I read it, but the AMD rep said that the addition of full Tonga would confuse product price/placement between present Tonga and Hawaii. Its absence is a conscious marketing decision by AMD.

More like the 290 had to take the price point that the 285X would have taken due to the 970, but yeah.

 

[Enigmoid]

Apple is probably putting a cut down Nano as the higher end option on the next iMac.

The imac can't handle the heat of the m295x tonga. Look at some apple forums and search throttling on m295x. No way apple can fit in any more power hungry chip without a chassis redesign. That or release a product that throttles even worse.

 

[monstercameron]

The imac can't handle the heat of the m295x tonga. Look at some apple forums and search throttling on m295x. No way apple can fit in any more power hungry chip without a chassis redesign. That or release a product that throttles even worse.

Sure they can, the fury line is more simple to cool than the prior gen. They might need to redo the cooler but it might not have an effect on the chassis design. Or the could down clock it to full tonga tdp and still have more perf due to increased perf/w

 

[raghu78]

It does not make any sense for R9 Nano to be a fully enabled Fiji chip. I can see a 3072 sp chip at 925 Mhz or a 3328 sp chip at 850 Mhz powering the Nano. By cutting the clocks and with some binning AMD can reduce voltage and power significantly. Yields on a 596 sq mm chip are not going to be good. So definitely I foresee a significantly cut Fiji chip driving the R9 Nano. AMD also needs a chip with better power efficiency which can compete with GTX 980. R9 nano can fill that role.

Right now the big problem is Fury drivers. Perf scaling over R9 390X is pretty wild fluctuating from 15-35% depending on game. There are multiple reviews now indicating that Fury performance is being held back by sub-par drivers. If AMD can solve the drivers issue and improve Fury performance in the games where perf scaling over R9 390X is poor they can improve overall avg performance by 5-10%. That would make all Fury based SKUs more competitive at 1440p which is a key resolution for many gamers. So what AMD needs is better drivers and a R9 Nano which can compete with GTX 980 in perf/watt.

 

[Lepton87]

I think it wouldn't make sense for AMD to release a card with a bigger die that would be slower than the card they already have on the market, namely 390X and because 390X already roughly matches 980 performance I expect Nano to be at least 10% faster than 980 and be competitive with aftermarket 980s. Both 980 and 980Ti have aftermarket variants that are up to 20% faster than the reference models and that is enough to place them in their own performance bracket.

 

[SolMiester]

If they need a CLC for Fury X full Fiji to get what they currently have, I cant see a tiny air cooled card being enough to cool the same chip!

 

[PPB]

^This is what you meant by bashing amd prpducts.

Yeah, lets disregard completely the clocks and voltages of the chip and just keep spreading fud. You could underclock and undervolt 4m vishera to consume 25w, and that is really a bad perf watt IC. How come you cant do the same on a gpu sku? I think 800-750mhz sweet spot with 0.85-0.9v

 

[LTC8K6]

This was clarified a little while ago. Can't remember where I read it, but the AMD rep said that the addition of full Tonga would confuse product price/placement between present Tonga and Hawaii. Its absence is a conscious marketing decision by AMD.

Well, R9-285 was confusing at launch as to it's intended position.

390 and 390X are very close in performance.

285 and 380 cards are GCN 1.2, but 390 series cards are not.

Lots of confusion.

Full Fiji Nano will surely be confusing, especially if I can put a better cooler on it.

Of course, Fiji based cards have an advantage being GCN 1.2

 

[tviceman]

Ok, there's more than one metric of "working" chip. Do all the dies work at the full size(what you are wondering)? Do all the dies work at the frequency we want? It's not just one metric.

So you can have 4 different SKUs based on the combinations.

1. Full die, full frequency
2. Full die, partial frequency
3. Partial die, full frequency
4. Partial die, partial frequency

While this is true, defective transistors are defective and no amount of downclocking can fix that. Which brings me back to my original point: amazed there are enough full functioning dies for two desktop sku's.

 

[Sabrewings]

While this is true, defective transistors are defective and no amount of downclocking can fix that. Which brings me back to my original point: amazed there are enough full functioning dies for two desktop sku's.

Maybe TSMC 28nm yields are so good that they have nearly all working dies. The cut down 980 Ti could be a formality just so they had something to release below the Titan X to combat the Fury X. :hmm:

 

[waffleironhead]

If they need a CLC for Fury X full Fiji to get what they currently have, I cant see a tiny air cooled card being enough to cool the same chip!

Its actually pretty simple. They just limit the clocks to the tdp envelope they are given. They just need to find the sweet spot of performance and power consumption.

 

[tviceman]

Maybe TSMC 28nm yields are so good that they have nearly all working dies. The cut down 980 Ti could be a formality just so they had something to release below the Titan X to combat the Fury X. :hmm:

Anything is possible, but that makes GM204's mobile dGPU's look very, very inefficient, cost-ineffective, and silly in the face of dies 1.5x it's size yielding through the roof.

 

[Sabrewings]

Anything is possible, but that makes GM204's mobile dGPU's look very, very inefficient, cost-ineffective, and silly in the face of dies 1.5x it's size yielding through the roof.

Very true. It's quite unlikely, I'll admit. If not true, then either Nano is not full Fiji or AMD are being quite foolhardy. Maybe the vanilla Fury is the binned chip?

 

[tviceman]

I will gladly accept I am wrong if/when Nano comes and is full fury!

 

[IntelUser2000]

While this is true, defective transistors are defective and no amount of downclocking can fix that. Which brings me back to my original point: amazed there are enough full functioning dies for two desktop sku's.

The yields might be pretty good, 28nm is heading to a 4 year period.

Intel makes even bigger dies, on a much newer 22nm process. Haswell EP/EX is at 660mm2. Knights Landing, the next gen Xeon Phi, will be at ~720mm2!!

I'd doubt the yields on those super large Intel chips on bleeding edge processes are better than the 3+ year old 28nm GPUs that are under <600mm2.

Why do they make them? Because the high prices(thus higher margins) and extremely low volume justifies the risk of having less yields. The Total discrete shipments are ~60 million/year. If we assume AMD takes 1/4 of that and its 15 million, the bleeding edge chips probably take 1-2 million at the most. That's not more than Intel server CPU shipments with much larger dies.

Pure silicon costs are quite negligible in the face of how much the final product sells for at least on the high end. If you have high yields, and you can justify making them, and you find out only some chips can function well at the max frequency, you bin them as something else.