Upcoming GPUs (2016)

bob, I know you like to pick on me, but here's where I'm coming from.

GPU	Memory Interface	Video RAM
GTX 460	192-bit / 256-bit	1GB / 1GB
GTX 470	320-bit	1.2GB or 2.5GB
GTX 480	384-bit	1.5GB
GTX 560	256-bit	1GB
GTX 560 Ti	256-bit / 320-bit	1GB / 1.2GB
GTX 570	320-bit	1.2GB
GTX 580	384-bit	1.5GB or 3GB
GTX 660	192-bit	2GB
GTX 660 Ti	192-bit	2GB
GTX 670	256-bit	2GB or 4GB
GTX 680	256-bit	2GB or 4GB
GTX 760	256-bit	2GB or 4GB
GTX 770	256-bit	2GB or 4GB
GTX 780	384-bit	3GB
GTX 780 Ti	384-bit	3GB
GTX Titan	384-bit	6GB
GTX Titan Black	384-bit	6GB
GTX 960	128-bit	2GB
GTX 970	256-bit	4GB
GTX 980	256-bit	4GB
GTX 980 Ti	384-bit	6GB
GTX Titan-X	384-bit	6GB
GTX 1060	192-bit	6GB
GTX 1070	256-bit	8GB
GTX 1080	256-bit	8GB

So, while I'm sure there is plenty of math involved, I think we can safely use a cheat sheet and say...

GDDR5/X 128-bit/256-bit = 1/2/4/8/16/32GB
GDDR5/X 192-bit/384-bit = 1.5/3/6/12/24GB

The only GPUs which deviate from this are the older 192-bit x60 GPUs, which only came with 1GB or 2GB each. I also leave out the 512-bit memory interface because AMD is the only one who has used it in recent times, but it follows the same guidelines where it's just twice of the 256-bit interface.

Quote:

---

Originally Posted by mbox_bob

HBM1 is 1024 bits wide.

---

I was only going off of what AMD prints on their website, and what the sites which reviewed and benchmarked both the R9 Nano and R9 Fury X were stating, but I know very little about HBM beyond that.

Quote:

---

Originally Posted by AMD

Ground-breaking performance requires big bandwidth. AMD’s latest innovation delivers more than 3x the bandwidth per watt over GDDR5 along with a 4096-bit memory interface for incredible new advances in power and efficiency.

---

:).... I was just saying.

I completely agree with your simplification; it was just that proposed 48GB card!!. You're going to start rumours.

/e re: HBM
In HBM1, each stack interface was 1024bits wide with 2Gb chips (256MB), 4 chips per stack (1GB per stack), the GPU memory controller interface was 4096bit, so, 4 stacks are needed.

HBM2 looks to be coming in 1024bits (still), with 8Gb chips in 4 or 8 chip stacks (rumours from Samsung), i.e. 4GB per stack (currently), potentially 8GB in a stack! With a 4096bit GPU interface, and it will pave the way for 16GB or potentially 32GB cards.

The math gets harder with HBM as you need to add in the number of chips per stack.

http://www.overclock.net/t/1604953/v...unch-in-august

VR World is reporting that they've had hands-on time with the GP102-based GeForce GTX Titan card, with a few details on NVIDIA's new monster.

NVIDIA's next-gen Titan X successor will feature both 8+8 and 8+6-pin PCIe power connectors, with the PCIe power connectors being found at the front of the card, and not on top. The 8+8-pin power connectors will provide up to 375W TDP, while the 8+6-pin version will provide 300W. Performance-wise, it should be a monster with around 50% more horsepower over the already lightning quick GTX 1080.

The new Pascal-based Titan X successor will reportedly be 12 inches long, while the new Titan X successor will arrive in 12/16GB versions (as we've previously reported) rocking HBM2 memory. VR World's sources say that the new GPU is actually now bound by the CPU, with NVIDIA's engineers reportedly saying that even Intel's new Core i7-6950X isn't powerful enough to deliver the performance the new Titan cards need in enthusiast level scenarios.

We should expect NVIDIA to unveil its new Titan P at Gamescom, which takes place in Cologne, Germany between August 17-21, 2016.

Quote:

---

Originally Posted by Pazgaz

while the new Titan X successor will arrive in 12/16GB versions (as we've previously reported) rocking HBM2 memory.

---

Makes sense based on what they say about the 4096bit interface or 3072bit interface. Presumably the lessor interface will be because part of the memory controller is disabled (whether due to bad etching or just because).

Also, if you look at the rx480, some of the cards have all 8 mem chips with 4 disabled while still having the same mem interface as far as I know.

Quote:

---

Originally Posted by mbox_bob

I completely agree with your simplification; it was just that proposed 48GB card!!. You're going to start rumours.

---

Well, we've already got GPUs with 24GB, so 48GB probably isn't all that far off, but I don't think I'd try to game on them (or even want to afford them).

Quote:

---

Originally Posted by Pazgaz

Also, if you look at the rx480, some of the cards have all 8 mem chips with 4 disabled while still having the same mem interface as far as I know.

---

I didn't say you can't have GPUs with two different amounts of VRAM on it, I said 12GB and 16GB were unlikely and I explained why. If you look at the table I created above you can see several different GPUs throughout the years which have had two different amounts of VRAM offered on it.

The RX 480 has a 256-bit interface, so it's capable of having any of the following amounts of VRAM on it:

GDDR5/X 128-bit/256-bit = 1/2/4/8/16/32GB

The new Titan X has been revealed:

http://videocardz.com/62498/nvidia-a...titan-x-pascal

12GB of G5X, and it's $1200 USD; available August 2nd (direct from www.nvidia.com).

Quote:

---

Originally Posted by MiRai

The new Titan X has been revealed:

http://videocardz.com/62498/nvidia-a...titan-x-pascal

12GB of G5X, and it's $1200 USD; available August 2nd.

---

Wow! that seemed fast!?!? I wonder how long until 1080ti!

Quote:

---

Originally Posted by Lyonheart

Wow! that seemed fast!?!? I wonder how long until 1080ti!

---

If we go by the typical nVidia release schedule as of late, then I'd expect them to be released in October or November... just in time for the holidays. :)

People are already saying this is a "small" Titan X (whatever that means) and that there will be another one later on (2017) with more CUDA cores and more RAM.
As for the prices, that's what you get with zero competition.