Rumor: GeForce GTX 1080 8GB gddr5x based on Pascal GPU in May
Well, at least I have a nice juicy rumor just in time for the weekend. Asian based Benchlife claims that Nvidia will release a Pascal architecture based GeForce GTX 1080 with 8GB gddr5x in May. It would make sense as HBM2 will be available in volume only later in the year.
There's a precise date being mentioned as well, May 27th, four days prior to the Computex exhibition. Obviously Nvidia is not going to comment on this rumor.
The GP104 based GTX 1080 would get a 256-bit memory controller tied to 8GB of gddr5x memory. The usage of gddr5x over regular gddr5 is rather speculative as it isn't even in volume production that we know of? Micron (the only one that will fab these gddr5x ICs) announced in a recent blog post that they intend to commence high-volume production of GDDR5X chips in mid-2016, or sometime in the summer. But who knows, Nvidia have gotten their hands on an early batch for a couple of thousand cards right ? GDDR5X is your standard GDDR5 memory however, opposed to delivering 32 byte/access to the memory cells, this is doubled up towards 64 byte/access. And that in theory could double up graphics card memory bandwidth.
Early indications according to a presentation show numbers with the memory capable of doing up-to 10 to 12 Gbps, and in the future 16 Gbps. So your high-end graphics cards these days hover at say 400 GB/s. With GDDR5X that could increase to 800~1000 GB/sec and thus these are very significant improvements, actually they are competitive enough with HBM.
Pascal GPUs are fabbed in the Taiwan Semiconductor Manufacturing Company (TSMC) (recently hit by an earthquake) based on the new 16nm FinFET process. 16nm should result in significant power savings. The card would only need one 8-pin connector for power needs. The card would get a DVI, HDMI and two display ports. From the looks of things, and as expected .. this is not 'Big Pascal' GP100. That GPU would get 17 billion transistors made on the 16 nm FinFET+ node at TSMC and likely the HBM2 memory.
So to recap, in the latest prognosis Nvidia are rumored to launch the first high-end GP104 GPUs with models GTX 1080 and GTX 1070 (should Nvidia use this naming scheme). These chips would come in a 37.5×37.5mm BGA package, and get 2152 pins.
Member
Posts: 60
Joined: 2014-01-24
But how would it work in DX12 compaired to AMD
Senior Member
Posts: 990
Joined: 2010-08-24
Hah first?
Pretty much what I expected. Except the 256bit bus... not sure this is a good idea, although I don't mind GDDR5X for the cards at all, 8GB should be more than sufficient.
At least it meets the time frame I expected for first launches, let's see how things turn out.
Let's just hope this time there won't be any 224-bit bus

But seriously though, a 256-bit bus could be enough if GDDR5X is aimed towards speed. I mean a 15GHz GDDR5 memory suddenly doubles what we currently have.
And it's not that hard to achieve even if it sounds like a lot, remember - GDDR5 reaches around 7GHz _effective_ data rate since it's a QDR-type memory. If they somehow managed to make it octa-pumped instead of quad-pumped a 5GHz memory today becomes a 10GHz memory tomorrow.
That's a possibility.
Junior Member
Posts: 15
Joined: 2012-12-31
I really hope this is true.
Especially since NVIDIA seems to have started nerfing the 7xx series with (lack of) driver support. In some benchmarks the 780ti seems to perform worse than the 970, which worries me... I thought I'd be fine skipping the 9xx series, but I'm slightly disappointed by how my 780ti's are holding up. Too late to jump on the 9xx bandwagon now, so can't wait for something new to come out.
Senior Member
Posts: 111
Joined: 2015-01-22
Ebay is about to get flooded with used 980's and 970's...

Senior Member
Posts: 14643
Joined: 2014-07-21
Hah first?
Pretty much what I expected. Except the 256bit bus... not sure this is a good idea, although I don't mind GDDR5X for the cards at all, 8GB should be more than sufficient.
At least it meets the time frame I expected for first launches, let's see how things turn out.