March 27, 2014, 12:55 AM — Nvidia kicked off its GPU Technology Conference with a few bangs, introducing a new GPU architecture and a monster of a card that will further the company's push into supercomputing.
At the show's opening keynote, CEO Jen-Hsun Huang introduced the next generation of GPU architecture codenamed Pascal, due in 2016. Like Intel and AMD, Nvidia strives for a new architecture every two years.
Pascal is going to be a huge leap if it keeps its promises. One of the major bottlenecks to GPU performance has been memory bandwidth, and Nvidia is addressing that with 3D "stacked" memory with the Pascal generation.
Pascal's primary innovation will be the integration of stacked "3D" memory situated on the same substrate with the GPU, providing substantially higher bandwidth than traditional DRAMs mounted on the same circuit board. Huang said Pascal's 3D memory will offer two to four times the memory bandwidth but on a card one-third the size of current add-in boards.
3D-layered DRAM chips can be integrated directly onto the GPU die instead of soldering chips onto the card, like they do now with GDDR5 RAM modules used in video cards today. Think of this on-die integration the way CPUs absorbed things like memory co-processors and other previously external chips. It means less chips on the board and faster performance.
The 3D integration will be of interest to the HPC crowd, who are using Nvidia processors in greater numbers to boost their supercomputers. After all, most HPC tasks boil down to giant math calculations, and a GPU is pretty much a math co-processor on pro wrestling levels of steroids. Now they will have vastly improved bandwidth between memory and the GPU which will mean for faster processing all around.
Gamers will love Pascal's NVLink technology, a high-bandwidth alternative to PCI Express 3.0 that Nvidia claims will be much more power-efficient. Huang claimed NVLink will be anywhere from 5 to 20 times faster than PCI Express 3.0. Ironically, it was created for use in supercomputing clusters and is meant to enhance GPU-to-GPU communication and many analysts at the event have commented that in many ways NVLink resembles PCI Express, just with a few tweaks.
Not all the talk was on the future. Currently, the company is introducing new cards based on the Maxwell design, and it introduced a doozy of a Maxwell card. Huang announced a dual-GPU Titan card, called Titan Z. This is a dual GK110 card with full implementations of the GPU. Nvidia whittles down the cores for lower-end cards, while leaving more for high end.
In this case, Titan Z has two chips with 2,880 cores each, along with 12GB of RAM, for up to 8TFLOPS of theoretical FLOPS performance. The price? A mere $2,999. This brought all kinds of sarcasm from Internet users, but they failed to realize this card isn't for them.