Fury Nano is Full Fiji

[IntelUser2000]

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.

Are you are saying that just because 100% of dies can work on a "BIG" chip, its a "waste" to create a 2/3rd die one?

Obviously even if 100% of the dies work, you can get lot more dies out with a smaller sized die.

 

[SolMiester]

^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

LOL, oh yes, Im having a laugh, your just wishfully thinking!

 

[Alatar]

All the arguments bringing up yields are assuming that the Nano is some sort of a good deal and a price/perf product.

There's no yield issue if the Nano is priced high and niche while the Fury X offers the flagship performance.

The normal Fury would vastly outsell both of those cards.

Personally I've always thought that the Nano would be niche and expensive. I don't understand why it got the hype in the first place. From normal desktop users who don't care about power that is.

 

[3DVagabond]

All the arguments bringing up yields are assuming that the Nano is some sort of a good deal and a price/perf product.

There's no yield issue if the Nano is priced high and niche while the Fury X offers the flagship performance.

The normal Fury would vastly outsell both of those cards.

Personally I've always thought that the Nano would be niche and expensive. I don't understand why it got the hype in the first place. From normal desktop users who don't care about power that is.

I think people expect it to be less expensive than 980 with better performance and better efficiency.

 

[Alatar]

I think people expect it to be less expensive than 980 with better performance and better efficiency.

And those people probably have too high expectations. Especially on the price front.

 

[PPB]

LOL, oh yes, Im having a laugh, your just wishfully thinking!

I know, it happens when you happen to be 100% ignprant about what a clock/voltage/power consumption curve is. It also happens whwn you nwver undervolted in your life.

 

[3DVagabond]

And those people probably have too high expectations. Especially on the price front.

Yeah, who knows?

 

[stahlhart]

LOL, oh yes, Im having a laugh, your just wishfully thinking!

Per the vendor-specific forum rules, you are out of order. Either stop the trolling and contribute constructively to this discussion, or get out.
-- stahlhart

 

[lopri]

I find it hard to believe they would name a fully functional die "Nano."

 

[SimianR]

I find it hard to believe they would name a fully functional die "Nano."

I thought the name was probably referring more to the board size as well as power use, heat, etc.

 

[flopper]

I thought the name was probably referring more to the board size as well as power use, heat, etc.

Me to

 

[richaron]

If AMD is to fully cash in on their investements then they must make use of all advantages they have. Obviously with Fiji/HBM it's the tiny PCB footprint. Even if it's not a top of line performance wise it's still very much a high end chip, so I don't see why Nano must be a "budget" Fiji when if will fill a niche no other card can.

As to the full Fiji idea, I'm sure the HBM mantra holds true with chips as well: "Wide and Slow". Up until now the more viable option for a lower performing product was to sell partially enabled chips with modified clocks (see Fury non-X). But if efficiency is the ultimate goal (like it is in Nano's niche), then a downclocked full Fiji being "wide and slow" may well be the answer.

 

[Sabrewings]

Wide and slow has worked for Nvidia in the mobile segment. It's possible, but still yields have to be good or it won't be cheap as a fully enabled chip.

 

[cbrunny]

Does this suggest that the Fuji Pro will be cut down but have higher clock speeds than the Fuji X?

 

[LTC8K6]

The Nano's PCB footprint is not really tiny, though. A GTX970 fits in that footprint, and likely so would a 980 if someone wanted to build one.

It might even be possible to go with the newer higher density GDDR5 chips and get a surprisingly powerful card in that form factor using GDDR5?

I have no idea what the cost factor is between GDDR5 and HBM though.

 

[flopper]

Does this suggest that the Fuji Pro will be cut down but have higher clock speeds than the Fuji X?

not higher speed.

 

[richaron]

The Nano's PCB footprint is not really tiny, though. A GTX970 fits in that footprint, and likely so would a 980 if someone wanted to build one.

It might even be possible to go with the newer higher density GDDR5 chips and get a surprisingly powerful card in that form factor using GDDR5?

I have no idea what the cost factor is between GDDR5 and HBM though.

This is unreliable (/wrong?) information. Fiji/HBM theoretical PCB footprint is unarguably smaller than any comparable chip with standard GDDR5.

It may be possible to get close in size with a smaller and less powerful chip running GDDR5 on a highly space-optimized PCB. But of course a HBM chip on a similarly space-optimized PCB will still be considerably smaller.

 

[Stuka87]

The Nano's PCB footprint is not really tiny, though. A GTX970 fits in that footprint, and likely so would a 980 if someone wanted to build one.

It might even be possible to go with the newer higher density GDDR5 chips and get a surprisingly powerful card in that form factor using GDDR5?

I have no idea what the cost factor is between GDDR5 and HBM though.

If you went with fewer, higher density GDDR5 chips you would lose a lot of memory bandwidth. And the 970 already has gimped bandwidth.

 

[LTC8K6]

This is unreliable (/wrong?) information. Fiji/HBM theoretical PCB footprint is unarguably smaller than any comparable chip with standard GDDR5.

It may be possible to get close in size with a smaller and less powerful chip running GDDR5 on a highly space-optimized PCB. But of course a HBM chip on a similarly space-optimized PCB will still be considerably smaller.

Theoretically?

Well, we have a mini-ITX sized GTX970 and R9-380.

AFAIK, it's the same size board as the Nano? Around 6-7 inches long?

It seems likely that we could get a GTX980 on that size board as well if someone wanted to bother?

 

[cbrunny]

I really don't understand the Fiji lineup. If nano is a full fuji chip with lower clock speeds and no other limitations, the X makes no sense. Does it?

Couldn't you get a nano chip and sufficiently cool it (probably custom water), oc it to at or above X levels (assuming voltage lock is released at some point)? nano will surely be less expensive than X, and with a smaller footprint anyone making custom water loops surely would choose the nano in this config instead of the full X.

or does it not have enough power? oh wait, nano is only 8 pin to X's 2 x 8 pin.. right? Nevermind that then.

Actually, that kind of irks me. full X is AiO or custom loop only with 2 x 8 pin, nano is full fiji but limited by power so 1 x 8 pin... so what is the point of the pro card being cut down at all? why would they cut it down and lower clock speeds? for all we know fiji doesn't actually need the water cooling at all and can handle itself just fine with air. Now we don't even have the option of finding out. That really frustrates me. How will a full fiju nano compete with the cut down fiji pro? Sheesh the nano will probably win if sufficient clock speeds are there.

Nano can be built on custom pcb boards right? any chance manufacturers will build these boards with 2 x 8 pins so they can be OC'd like crazy?

 

[Enigmoid]

If you went with fewer, higher density GDDR5 chips you would lose a lot of memory bandwidth. And the 970 already has gimped bandwidth.

Each GDDR5 chip has a 32 bit bus width.

A 256 bit bus requires 8 chips of 512 MB (for 4 GB).

http://techreport.com/news/27676/samsung-starts-making-8gb-gddr5-memory-chips

The 980/970 must use at least 8 chips to get full bandwidth. It looks like they are already using 8 chips so there is no room to cut there. Using higher density GDDR5 will allow for 8 GB models however.

 

[LTC8K6]

Those Samsung chips are also 8gbps instead of 7.

Increased bandwidth of the new high-density modules equates to 8 Gbit/s per pin x 170 pins on the BGA package x 32-bits per I/O cycle, or 256 Gbit/s effective bandwidth per chip.

Not sure what that means overall, though.

 

[richaron]

Theoretically?

Well, we have a mini-ITX sized GTX970 and R9-380.

AFAIK, it's the same size board as the Nano? Around 6-7 inches long?

It seems likely that we could get a GTX980 on that size board as well if someone wanted to bother?

Sure, of course there are mini-ITX boards with GDDR5. The point I was trying to make was about the PCB savings of HBM, and I'm not sure there is any disagreement.

Meaning all thing being equal, if "someone" trys minimise the size of a comparable GDDR5 board it will still be larger than "someone" trying to minimise a Fiji/HBM card.

I suspect the confusion is that there is no point "minimizing" the size of the board past a certain point (ie mini-ITX sized dGPU). So even though it may be possible to fit a "comparable" GDDR5 based solution in similar footprint, the HBM card will have more space on the PCB for other circuitry.

 

[Erenhardt]

It is not all about size (LOL)
Just look at your GDDR5 PCB. Notice all these traces going from GPU to memory chips. Imagine, PCB has multiple layers, each with those traces. HBM doesn't need that. There was a rumor that Fiji PCB is only 4 layers while hawaii have 7 of them.

 

[LTC8K6]

I suspect that you can't make an air cooled mini-itx full performance Fiji card, though. Probably cannot cool it, despite having PCB real estate to spare.

I'm not sure there is really a whole lot of board real estate difference, though.

http://cdn.overclock.net/9/9d/900x900px-LL-9d3ac8fc_IMAG08711.jpeg

http://cdn.arstechnica.net/wp-content/uploads/sites/3/2015/06/DSC02004-980x653.jpg