I figured it out, they aren't counted. I used a chart on this article: http://wccftech.com/nvidia-gp100-gpu-tesla-p100-benchmarks/ The one comparing the specs of the Tesla Cards specs. If you multiply the SMs times the Cuda Cores Per SM, you get the cuda core count total. There is also the FP64 Cuda Cores Per SM, which when multiplied by SMs equals the FP64 Cuda Core Count Total.
The 8000 part maybe, although considering AMD has been jumping like 1000ish cores per generation, Idk if 8000 is reasonable.
Dude all I know is that there is not a lot of the chips available, as in its hard to get and expensive. Only an AMD or Nvidia would know costs or yields outside of the manufactures.
uh, usually a node shrink means more cores. The original Titan X Die size was barely small enough that the chip producers could make it. Also, Intel was able to go from 18 cores on 22nm to 24 cores on 14nm because the cores are smaller and more can be fit into the same sized package.
The new Tesla is only 16GB so I doubt they will go higher than that with this generation of cards.
Site seems to be loading slow... seems a lot of interest right now... I'll read it and post comments in a minute
EDIT: Read the post. It seems to be overly cocky, proving that Nvidia knows it has a decent advantage over AMD as of the time of this release. In all honesty if I did more engineering or video editing or any of the things I originally aspired to do, I would 100% get this card for the raw power.
I understand that a lot of people look at the cost, but honestly I would consider this a "prosumer" option. It's really for a professional that may also want to play games and perform other functions. The consumers who buy this card either want to be at the top of their tier or just have the money to blow.
twice the density should mean twice the core count if the die size is the same.
it sounds like bs. silicon wafers arent expensive and you get a lot of chips per wafer. sounds like a really rich complaining about how they cant afford the 30 million dollar yacht and only the 15 million dollar one.
So 1080ti doesn't have to be Volta to close the gap with the new Titan X? They could always just go, say, 2800+ cuda cores for the 1080ti with insane clock speeds, and keep the new Titan X at lower clock speeds just like their past Titans.
Yeah but designing architecture that supports 8000 cores is another thing
If they had stock they would already be selling Tesla P100's. Stock isn't good enough for them to even start selling those until Q1 2017, even though they announced the card months ago.
I bet they'll go for a 52 SM card, maybe 3328 a tad lower.
I just have to wonder what the Titan Z price was in Australia at launch. I mean shit it seems like prices near double over there, and it was ridiculous in America ffs. If it was more than a small used car I wouldn't be surprised rofl
A yield is how many working ASICs can be cut from a printed wafer. Those wafers have impurities and imperfections scattered across them, and not every full print is salvageable. In a new node these imperfections can be plentiful, and can even be caused by the printing itself as the design of the ASIC matures.
The reason that large Pascal's yields are so low is because the ASIC die size is huge, and so fewer per wafer can be printed. Since there are now fewer prints, each print has a higher chance of hitting one of these imperfections and being useless. This extends across the industry.
Why HBM 2's yields are low would be down to the accuracy of the printing method at the lower process size. These things take time to calibrate and configure, and so the first few hundred thousand prints may almost entirely be scrapped as the process matures.
The price of the individual wafer is cheap, but you must buy them in bulk and they then need to be processed for production for the target node and design. The cost comes from having to buy 1,000 - 10,000 wafers at a time, printing them all, cutting and fusing all the individual chips then taking them back for testing and evaluation. Huge amounts of money invested for a small return. If that yield is bad, you spend more money and get another small return until you can finally get the whole fab process dialed in and working to the point where production of actual products can begin.
TL;DR - Yields are how many working chips make it off the wafers. Bad yields = little to no product.
if 1080 was listed as $599 (but actually costs around $900), 1080ti should be cost around $799 it'll be for $1200, and titan x if it's $1200 we can expect it to be $1500