Loading recent posts...

Mar 25, 2014

NVIDIA Intros Pascal GPU, 3D Stacked Memory Chips, and NVLINK Technology

Even though NVIDIA hasn't even launched the real Maxwell GPU yet (that will only happen in late 2014, and the GTX 750 / 750 Ti don't count), the company has already introduced the Pascal GPU, scheduled for 2016.

Well, technically the company hasn't actually released the chip. More like it revealed its presence in the roadmap. The real highlight of the ongoing GTC, taking place in between March 24-27, 2014, in San Jose, California, was the NVLINK technology. What NVLINK does is unify the GPU and CPU cache in a system, allowing the two distinct processing units to divide tasks between each other much faster than right now. As allowed by the PCI Express 3.0 interface, this means that performance for the card, over PCI Express, can rise by 5 to 12 times. In turn, this enables supercomputing performance increases by a factor of 5 to 100. This is a huge leap in performance, on both counts. The Pascal GPU, which will succeed Maxwell, will supposedly have a bandwidth of thousands of bits on a single chip as well, compared to the 512 maximum of right now.

That raised another issue: how will you get enough memory to use that interface, and how will all the chips fit and communicate quickly, without having to wait in line as it were? The answer was 3D packaging for the VRAM chips. So you get a memory interface of thousands of bits, and dozens of chips stacked on top of one another, with holes in them through which silicon interconnects pass. The Pascal sounds like the superman of GPU technology, and it may very well be. In fact, NVIDIA spoke at length about the use of the chip module (which is as big as two credit cards) and how it does best in “machine learning.” Which is to say, computers, or supercomputers, that can learn from past experiences how to recognize things when asked. So far they've been limited to recognizing that human faces and cats show up most often on the Internet.

Of course, that conclusion was drawn after doing the equivalent of watching YouTube for 2 days, so it's not all that reliable, but still. The Pascal is smaller than a PCI Express slot, so with some hardware platform customization it could be installed in an HPC (high-performance computing) installation in high density. And with NVLINK working both CPU-to-GPU and GPU-to-GPU, performance will scale quite well there. Click for larger images.

NVIDIA Pascal GPU saves Moore's Law
Image credits to NVIDIA


Post a Comment

Twitter Delicious Facebook Digg Stumbleupon Favorites More

Design by Free WordPress Themes | Bloggerized by Lasantha - Premium Blogger Themes | coupon codes