News
News Categories

NVIDIA announces a pair of new Maxwell-based Tesla graphics accelerators

By Wong Chung Wee - on 10 Nov 2015, 12:00pm

NVIDIA announces a pair of new Maxwell-based Tesla graphics accelerators

(Image source: NVIDIA)

NVIDIA has announced a pair of new Maxwell-based Tesla graphics accelerators. The Tesla M40 and the Tesla M4 are compute accelerators best suited for hyperscale computing in the aspects of machine learning and visual processing. Exabytes of user-driven content are created daily across popular web services that include YouTube and Twitch. According to data shared by NVIDIA, Baidu handles 6 billion search queries daily and 10% of them are voice and image inputs. Baidu expects voice search and search-by-image requests to grow quickly. In the same breath, there’s also the need to meet the user request as they consume video content over web services like Facebook, Periscope and YouTube. Therefore, in order to harness this deluge of data, akin to a tsunami, in real-time, companies have to turn to hyperscale computing with special graphics compute capabilities.

(Image source: NVIDIA)

Hoping to ride on the forefront of this data tsunami, NVIDIA has introduced two Maxwell-based Tesla graphics accelerators. First, the single-GPU Tesla M40 that boasts of 3,072 CUDA cores, with a peak performance of 7.1 TFLOPs (single precision floating-point operations). It features 12GB of GDDR5 VRAM. We can safely assume the memory modules operate at 6GHz, and with its stated 288GB/s memory bandwidth, the card should feature a 384-bit wide memory bus. Its TDP is rated at 250W. The card’s main application is its deep learning capabilities; on the Caffe framework, its learning duration is 8 times faster than a non-GPU accelerated, Intel Xeon-based server. In a hyperscale computing environment, the NVIDIA Tesla M40 will be deployed in servers that “scale with data” in order to develop trained machine learning models from the exabytes of data, which the servers parse daily.

(Image source: NVIDIA)

In order to make use of the trained models and scale them across servers that running dedicated services for users, there’s the Tesla M4 graphics accelerator. This low-profile, energy-efficient accelerator boasts of 1,024 CUDA cores, with a peak performance of 2.2 TFLOPs (single precision floating-point operations). The card features 4GB of GDDR5 VRAM, and has a memory bandwidth of 88GB/s. Its TPP is rated from 50- to 75W.

(Image source: NVIDIA)

To complete the Tesla ecosystem, NVIDIA has also announced the new Hyperscale suite and its key features. First, the GPU Rest engine offers real-time accelerated services for developers; the app developers don’t have to explicitly take care of scaling their applications across servers as GPU Rest engine is able to do so behind the scenes. Next, the CUDA developers have developed a new GPU-accelerated plug-in for FFmpeg. Therefore, users of this popular multimedia solution can make use of the Tesla-enabled GPU systems to work on their AV content. Last of all, there’s the Image Compute engine, which is built on the GPU Rest engine, and this compute algorithm allows for image sizing on-the-fly. An application for this feature will be how a server farm, of Tesla M4 accelerators, is able to meet throughput requirements of a large user base that is consuming content through different devices.

The Hyperscale suite is supported on Apache Mesos, so that hyperscale resources can be managed and scheduled with the open-source datacenter manager. The exact launch details of this new Tesla graphics compute ecosystem aren’t made known yet; however, some simulation work of the Tesla-based system can be viewed at the upcoming SC15, the International Conference for HPC, Networking, Storage and Analysis, which will be held next week in Austin, Texas, United States of America.

(Source: NVIDIA, SC15)

Join HWZ's Telegram channel here and catch all the latest tech news!
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.