VEGA Preview slides leaked
- Thread starter ParallaxScroll2
- Start date
ParallaxScroll2
New member
bittermann
Well-known member
Well...at least it's got HBM2!
That said, in this close up picture at least a few more things can be discerned for the version being held by Raja, it's using 2 stacks of HBM 2.0 rather than the 4 seen in Fiji but gets the same bandwidth since HBM 2.0 also runs at twice the clock speed...
It's 500 Mhz base clock, Quad pumped, 4 stack HBM 1.0 for Fiji vs 1000 Mhz base clock, Quad pumped, 2 stack HBM 2.0 for Vega and the other big difference is that this configuration allows cards with 8 or 16 GB of ram onboard, so each stack could be 4 GB or 8 GB.
Fiji needed 4 stacks to even get a total of 4 GB of ram, so each HBM stack on Vega has increased it's density dramatically even if they're only 4GB stacks each for a potential " gamer " version, and goes to the maximum of the HBM 2.0 spec with 16 GB versions ( 8GB per stack ), assuming this latter one is exclusive to the Compute / workstation market willing to pay the price.
The last thing I noticed was the size of the GPU but this is just eyeballing it......Seems it isn't far from 1 inch per side ( so 25.4 mm) but slightly wider than it is taller ( slightly rectangular ), meaning a total size in the ~500mm^ range so it's a big boy and using the 14nm process that is smaller than TSMC's 16nm process ( a known fact ), it looks like Vega is larger than the GP 102 found in the Titan - X ( 471mm^ for this with a 12 billion transistor budget ).
The preview indicates a single precision math output performance that could be as high as 12.5 teraflop while the GP 102 in the titan - X maxes out at 11 teraflop.......
GP 104 found in the GTX1080 just isn't there at 9 teraflop, and a hardware castrated version of the GP102 in the titan -X, but used in the eventual release of a GTX1080 TI card, wouldn't be there either......It'll obviously have less single precision math than the Titan-X with less hardware onboard.
bittermann
Well-known member
I think the 1080 overclocked could get upwards of 11 teraflops but don't quote me on that. Someone esle on [H] was doing the math. Whatever we end up getting it seems it's a much bigger architectural change than what most thought it would be.
Much better picture follows :
It is definitely rectangular, and the picture large enough for anyone willing to try and measure this out.....Just a thought...
The wider end could be pretty close to an inch on the nose, so a ~500mm die size seems about right.
HyperNovae
New member
In the first shot you figure rajas thumb from his middle joint to the end of the thumb is about an inch so looks about right.