Nvidia has lifted the curtain on a monster new GPU, dubbed the GeForce GTX Titan Z, a heavy metal graphics card capable of 8 TeraFLOPS of compute performance.
The GTX Titan Z packs two Kepler-class GK110 GPUs with 12GB of memory and a whopping 5,760 CUDA cores, making it a "supercomputer you can fit under your desk," according to Nvidia CEO Jen-Hsun Huang. Nvidia didn't offer a release date for the follow-up to 2013's originalGeForce GTX Titan and the GeForce GTX Titan Black refresh launched last month.
A single,Titan Z-accelerated server with three GTX Titan Z GPUs sporting 17,280 combined cores could run the workloads performed on Google Brain, a massive, neural network using 1,000 dual-CPU socket servers.
What's more, the Titan Z server would only require 2 kilowatts of power to perform on a par with Google Brain, which would suck down 600 kilowatts just to keep up, Nvidia's chief said.
"Unlike traditional dual-GPU cards, Titan Z's twin GPUs are tuned to run at the same clock speed, and with dynamic power balancing. So neither GPU creates a performance bottleneck," Nvidia said in a blog post.
Huang, opening the GPU Technology Conference here, also offered a peek at a future processor architecture code named Pascal that will be the first GPU to use stacked, 3D chip packaging and to incorporate a new PCI Express-based interconnect technology called NVLink.
The Nvidia chief spent some time towards the end of his keynote talking about Tegra, the company's mobile chipset for smartphones, tablets, and embedded systems. Tegra moves fully into Nvidia's CUDA framework with the Tegra K1 "mobile super chip" introduced in January at the Consumer Electronics Show.
"This chip unifies the architecture of CUDA and Kepler across everything we do. It's a supercomputer on a mobile chip. What runs on the Oak Ridge Titan supercomputer will run on this chip, only much, much more slowly," Huang said.
Nvidia will be rolling out a development kit called the Jetson TK1 using next-gen Tegra GPUs, offering 192 CUDA cores, 326 gigaFLOPS of compute performance, support for the CUDA 6.0 developer tool suite, and the company's VisionWorks software in a package.
"It's the first mobile supercomputer for embedded systems," Huang said. To demonstrate the power of the Jetson TK1, Nvidia hosted Audi on stage for a demonstration of the automaker's self-driving car using a Tegra-powered advanced driver assistance system (ADAS).
The Jetson TK1 Development Kit is available for pre-order starting today from Nvidia, Microcenter, and Newegg in the United States, from Avionic Design, SECO, and Zotac in Europe, and from Ryoyo Electro Corporation in Japan.
Nvidia also introduced a new GPU rendering appliance called the Iray VCA, which "dramatically accelerates ray tracing" for photorealistic rendering. It also announced the availability of Nvidia GRID technology on VMware's Horizon DaaS Platform for cloud-based delivery of 3D graphics to virtualised desktops.