Nvidia Releases Next-Generation Tesla GPU Accelerators

The new K80 offers twice the performance and memory bandwidth of the current K40, and is being leveraged by HP, Dell and other system makers.

data center

Nvidia is rolling out the Tesla K80, next generation of its GPU accelerators that offers almost twice the performance and memory bandwidth of its predecessor, which was released only a year ago.

The K80, announced Nov. 17 at the SC14 supercomputing show in New Orleans, becomes the flagship product for Nvidia's Tesla Accelerated Computing Platform, which offers hardware, software and an accompanying ecosystem for accelerated applications in the data center.

High-performance computing (HPC) organizations are increasingly turning to accelerators—both GPUs from Nvidia and Advanced Micro Devices and x86 co-processors from Intel—to boost the performance of their systems while keeping a lid on power consumption and operating costs. Sixty-two of the 500 fastest supercomputers in the world—according to the Top500 list released in June—used accelerators or co-processors, up from 53 on the November 2013 list. Forty-four of those systems used Nvidia GPUs, with 17 running Intel's x86-based Xeon Phi co-processors. Two used ATI Radeon GPUs from AMD.

Those numbers will change when the latest list is released at the supercomputer show this week.

According to Sumit Gupta, general manager of Tesla accelerated computing at Nvidia, the K80—with 4,992 CUDA parallel processing cores—runs complex and computationally-intensive scientific applications 10 times faster than CPUs. Running such applications on systems with CPUs alone, without accelerators, at the same rate of performance would require supercomputers that are significantly larger, more expensive and less power efficient, Gupta told eWEEK.

"CPUs just don't give the performance" with the same energy savings as GPUs, he said.

The introduction of the K80 comes just days after the Department of Energy announced that it was awarding IBM, Nvidia and Mellanox Technologies $325 million to build two new supercomputers at the agency's Oak Ridge and Lawrence Livermore national labs that will be five to seven times more powerful than the systems currently running at the facilities.

Those systems, which will be delivered sometime in 2017 or 2018, will feature Nvidia's upcoming Volta accelerators, which are two generations ahead. The Volta GPUs will include a technology developed with IBM called NVLink, a high-speed interconnect first introduced in March that will enable CPUs and GPUs to exchange data five to 12 times faster than they can today.

The Tesla K80 comes with Nvidia's GPU Boost technology, which increases the performance of the GPU when applications need it. The accelerator offers two GPUs per board to double the throughput for applications that are designed to take advantage of multiple GPUs, 24GB of GDDR5 memory—twice the memory of the Tesla K40—480 GB/s of data throughput and dynamic parallelism, which enables the GPU to dynamically spin out new threads.

The K80 supports x86, Power and ARM-based chips, and more than 280 HPC and data center applications. The new accelerator "extends what now has been a very large and broad ecosystem for the platform," Gupta said.

The K80 is available immediately, and a broad array of server OEMs are putting the GPU accelerators into their systems, including Hewlett-Packard—which is using them in its high-end Apollo supercomputers—Dell, Cray, Asus, and Supermicro. Other system makers leveraging the K80 are Cirrascale, Gigabyte, Inspur, Penguin, Quanta, Sugon and Tyan.

The GPUs also are available via Nvidia resellers.