NVIDIA Pascal GP104 Die Photo
Last week Nvidia announced the Tesla P100 data-center GPU and if you looked closely at the photo's you could already clearly see that big Pascal 15B transistor GPU being used. A new photo this time showing the GP104 GPU (successor to the GM204 = GTX 970/980).
According to the leak the photo shown below is the GP104 (GM204 successor) intended for Nvidia's pending high-end products, and it measures roughly 290-300mm². At the right side you can see Samsung K4G80325FB - 1.5V 8Gb 8Gbps (8000MHz) GDDR5. Word on the street is that this GPU would hold 2560 shader processors.
The rectangular die of the GP104 was measured at 15.35 mm x 19.18 mm which should house (very speculative) a transistor-count of 7.4~7.9 billion. Interesting to see is that this chip is tied towards Samsung 8-gigabit GDDR5 memory chips. These ICs run an effective speed of 8 GHz (GDDR5 and thus not GDDR5X). At 256-bit you'ld be looking at 256 GB/s of bandwidth. The shader processor core count of the GP104 could be closer to 2,560, than the 4,096 from an older report.
Earlier on it was speculated that graphics cards based on these GPUs would be called X70 and X80, the specs do not meet up even slightly though.
Judge rules that Samsung and Qualcomm do not violate Nvidia patents - 10/12/2015 09:10 AM
According to a judge ruling from the International Trade Commission, Samsung and Qualcomm do not violate Nvidia patents. Nvidia released this news themselves on their blog. Nvidia will appeal this r...
AMD and Nvidia prep for next-gen DirectX 12 - 03/24/2015 11:19 AM
Several new sets of slides from AMD and Nvidia as well surfaced on the webs through slideshare, the decks show what the companies have been up to in regards to DirectX 12, and they do look pretty exc...
Nvidia PhysX Source Code Now Available Free - 03/06/2015 09:03 AM
NVIDIA today put more than a decade of research, development and investment in gaming physics into the hands of game developers – by offering free source code for NVIDIA PhysX on GitHub. Thi...
New Nvidia PhysX FleX features - 12/11/2014 10:20 AM
PhysX FleX is a particle based simulation technique for real-time visual effects. So far we showed examples for rigid body stacking, particle piles, soft bodies and fluids. In this video we'll be sho...
Ubisoft and NVIDIA Partner Up for Assassin's Creed Unity Far Cry 4 And More - 06/05/2014 09:03 AM
Ubisoft and NVIDIA today announced the next chapter in their strategic partnership bringing amazing PC gaming experiences to life in Ubisoft's highly anticipated upcoming titles including Assassin's...
Senior Member
Posts: 13716
Joined: 2004-05-16
Guess i'll be waiting for the 1080/x80 Ti then. Hopefully AMD has a better product or comparable one so the price isn't super high.
Yeah, but by default. Workstation GPUs automatically cost more just because they can. They have ECC and other features not used on Consumer cards.
Trust me, there's no way Tesla cards actually cost 12k$ to make. They probably cost a fraction of that.
They obviously cost a fraction of that, but the point is yields at the start of a new node, especially 600mm2 chip, are going to be really low. You're going to lose over 75% of a wafer. High margins will offset that, which is why they are reserving GP100's to high margin products at the start of production.
Senior Member
Posts: 954
Joined: 2010-08-24
Wow I never thought of that, it makes perfect sense. They can still make great profit this way.
Senior Member
Posts: 3355
Joined: 2012-02-02
Guess i'll be waiting for the 1080/x80 Ti then. Hopefully AMD has a better product or comparable one so the price isn't super high.
I kinda get the feeling big Pascal GP100 will be only for Tesla GPUs. Those big chips are just too expensive to sell for consumer gear. The DX-1 box costs around $130,000 with 8 of those GPUs. If you cut off around $30K for the other server components, your looking around $12,500 per GPU! I don't think we'll see anything like this chip for consumers until 16nm process is mature. I bet the chip yields are horrible right now, hence why they cost a fortune.
I bet the GP104 chip will be a different design entirely. Well still based off the GP100 but with some changes. They will probably kill all the Double Precision (DP) units, games don't really make use of that hardware. That silicon space would be much better utilized for Single Precision (SP) shader cores. Currently half of the silicon space on the GP100 is used for DP units, using that space for SP units would be a much better design for a gaming GPU.
Yeah, but by default. Workstation GPUs automatically cost more just because they can. They have ECC and other features not used on Consumer cards.
Trust me, there's no way Tesla cards actually cost 12k$ to make. They probably cost a fraction of that.