[TT] Pascal rumored to use GDDR5X..

[Silverforce11]

http://www.tweaktown.com/news/47976/nvidia-rumored-use-gddr5x-next-gen-pascal-based-gpus/index.html

Take it with salt.

The latest rumor is that there will be a successor to GDDR5 called GDDR5X, and it will debut with NVIDIA's next-gen GeForce product. GDDR5X will reportedly keep the same 256-bit memory interface we have on the current GTX 980 while ramping up the memory bandwidth to 448GB/sec - higher than the best AMD cards, save the Fiji-based HBM-powered cards. HBM2 is still obviously on the cards, and will include a 4096-bit memory bus at 1GHz with memory bandwidth hovering at 1TB/sec.

In our post about NVIDIA testing its new Pascal GPU internally, I did say "I would love to see a GDDR5-based offering for the cheaper mainstream side, and a HBM2-powered GeForce 1000 series card for the enthusiasts" and it looks like I could be right on the mark if the current rumors are true. It looks like we could expect a GDDR5X-powered, Pascal-based GeForce card while the enthusiast-class cards will enjoy the much faster, but more expensive HBM2 technology.

But there was also that HBM2 delay rumor, which given HBM1 issues thus far is within the realms of possibilities.

Given potential yields, cost with HBM2, it makes sense to use it only for high-end and stick with current cheaper tech for mid-range and below... that's my thought process on this issue.

 

[ShintaiDK]

GDDR5X looks like a good option if we assume HBM2 is in the disaster as its rumoured.

 

[Headfoot]

I wonder how much GDDR5x will cost compared to GDDR5, is there any info on that floating around on the interwebs? I've only seen that slide that Shintai linked above so far

 

[ShintaiDK]

Seems Micron is already ready to ship 10, 12 and 14Ghz GDDR5X. I would guess cost would be very close to GDDR5. Maybe even parity for 10Ghz vs 7Ghz GDDR5.

16FF GP100 with GDDR5X? Starting to look plausible.

 

[AtenRa]

Two drawbacks vs HBM, space and power.

 

[flopper]

old tech reused by Nvidia, what a shocker

 

[thesmokingman]

Two drawbacks vs HBM, space and power.

There are two concepts that will want to be thrown out the window in terms of discussion.

 

[PPB]

GDDR5X looks like a good option if we assume HBM2 is in the disaster as its rumoured.

Seems Micron is already ready to ship 10, 12 and 14Ghz GDDR5X. I would guess cost would be very close to GDDR5. Maybe even parity for 10Ghz vs 7Ghz GDDR5.

16FF GP100 with GDDR5X? Starting to look plausible.

Wow it didnt even take 1 day from a rumor article for the usual suspects to hop into the spinning machine. So now HBM2 is a non-product and GDDR5X, which nobody even heard of before looking at Crucial powerpoint slides after reading this article is the best thing since sliced bread.

To atenra's point, I would add that once that people realizes that HBM isnt just a way to get higher bandwith without going nuts with clocks and voltages, but also a game changer in the form-factor space, 30cm long cards will be a thing of the past.

The money for next gen will be at 980/390x performance, but not at the card size those 2 today sport. Who would buy a 120W card (thinking here about 16nm FF+/Glofo 14nm power benefits) that still has to have at least a 24cm PCB because the space needed for +4GB of GDDR5/X memory?

 

[lilltesaito]

old tech reused by Nvidia, what a shocker

Sound like it is newer tech to me. Also seem like a great way to bring about new cards next year is HBM2 isnt ready.

Smart idea to be ready and have a back up plan to keep tech moving forward.

drawbacks vs HBM, space and power.

Is this even really an issue for Nvidia? With the smaller node size and them already having really good power usage.

Way I see this is the lower-mid range cards will be getting this and they will charge an arm and leg for them. After while they will bring the big guns out and move up the price of the higher end cards. Just like they did with the 980. Would not shock me to see a 980ti class card costing $800+ next year.

 

[thesmokingman]

Is this even really an issue for Nvidia?

Exactly see, as if DDR power consumption was ever at a head? It will now be as it never existed!

 

[AtenRa]

Is this even really an issue for Nvidia?

Perhaps not for the desktop but im sure in Laptops a Pascal GDDR5X will loose big time vs Radeon HBM at BoM, thermals, power consumption etc etc. .

 

[ShintaiDK]

How many HBM Fijis are in laptops?

 

[AtenRa]

How many HBM Fijis are in laptops?

What Fiji has to do with next Gen 14/16nm GPUs ???

 

[crisium]

Makes sense if HBM2 is delayed/slow production and/or very expensive.

Could be a problem for AMD if there is an HBM2 delay. Going with HBM1 again, even if just for the first round of cards (7970/680 "mid-range") will really sour people if they are at 4GB and Nvidia is on 8GB GDDR5X. Alternatively if AMD used GDDR5/GDDR5X they will lose much of the power efficiency gains that Fiji had and unless Arctic Islands is a massive improvement in PPW then AMD will have no answer to the top Nvidia card.

But hopefully HBM2 arrives plentifully.

 

[thesmokingman]

How many HBM Fijis are in laptops?

Ah, we're not talking about Fiji here. Ya?

 

[Pinstripe]

Who would buy a 120W card (thinking here about 16nm FF+/Glofo 14nm power benefits) that still has to have at least a 24cm PCB because the space needed for +4GB of GDDR5/X memory?

Uhm, I would buy that?

 

[tential]

Two drawbacks vs HBM, space and power.

Not too worried about power given fiji vs maxwell. Seems nvidia is ok on that front.

Space wise, this is perfect for nano. It reigns supreme in fps per inch (lol) and its successor has the potential to do the same.

However, without more mitx style cards like nano being mainstream, we won't get cases specifically designed for these builds.

Sucks, but nvidia is the gpu industry apple.

 

[Pinstripe]

HBM2 for the 4K gaming solutions, GDDR5X for the 1080p gamer. Makes perfect sense.

 

[ShintaiDK]

The money for next gen will be at 980/390x performance, but not at the card size those 2 today sport. Who would buy a 120W card (thinking here about 16nm FF+/Glofo 14nm power benefits) that still has to have at least a 24cm PCB because the space needed for +4GB of GDDR5/X memory?

GDDR5(X) cards can be significantly shorter, and already plenty of examples. You have been on too much Nano koolaid.

A 120-150W Pascal card with GDDR5X could easily look like this (GTX970).

 

[AtenRa]

Not too worried about power given fiji vs maxwell. Seems nvidia is ok on that front.

Really ?? Nano at only 175W TDP is only 10-15% slower than 250W TDP GTX980Ti in latest games.

Take Nano,
Shrink it to 16nm FF
add DX-12 games
Include Laptops in the equation

I dont think Pascal + GDDR5X will have fun competing against it.

 

[ShintaiDK]

Really ?? Nano at only 175W TDP is only 10-15% slower than 250W TDP GTX980Ti in latest games.

Take Nano,
Shrink it to 16nm FF
add DX-12 games
Include Laptops in the equation

I dont think Pascal + GDDR5X will have fun competing against it.

As always you forget the Fury and Fury X with 275W TDP. You lack any numbers for GM200 to do a direct compare.

 

[PPB]

GDDR5(X) cards can be significantly shorter, and already plenty of examples. You have been on too much Nano koolaid.

A 120-150W Pascal card with GDDR5X could easily look like this (GTX970).

So you are condemning GP104 to 4GB of memory. Nice

EDIT: Only 2 OEMs have that size of PCB in the market right now, seems the cost structure isnt as nice as some posters may think if only 2 itx 970 were released. Also, those 2 PCBs are shorter but wider than reference PCB, they wont fit in any case with a width lower than 14cm (without mobo tray backside, with mobo tray backside think about 16cm) without the need of a riser card (which adds to cost, which negates any other expansion card, etc)

 

[AtenRa]

You lack any numbers for GM200 to do a direct compare.

What numbers do we lack from GM200 ???

 

[ShintaiDK]

What numbers do we lack from GM200 ???

You got 3 Fiji cards. You conveniently pick the lowest. A binned and down clocked version. Then compare it to 980TI instead of the Fury/Fury X. It was too obvious wasn't it.

 

[ShintaiDK]

So you are condemning GP104 to 4GB of memory. Nice

GDDR5 got higher densities. How do you think pro cards can get 24/32GB on 384/512bit? Or Titan X with 12GB? This isn't HBM1 that is limited to 1GB per stack.

The current GDDR5X chip that used for implementation is 8Gbit. That means a GP104 would have 8GB with an option for 16GB.