Home » Gadgets » Project Brainwave: Microsoft accelerates AI technology with cloud FPGAs

Project Brainwave: Microsoft accelerates AI technology with cloud FPGAs

Wednesday, August 23rd, 2017 | Gadgets

The most widespread opportunity to speed up machine learning is the use of graphics cards. However, some manufacturers are now using hard-wired ASICs such as Google with their TPUs to apply and train models. Microsoft wants to use the project name Brainwave future for such fields of application on FPGAs in its Azure cloud.

The announcement states that this is to enable "real-time AI", whereby corresponding data are processed as quickly as possible by the system. Microsoft uses three components: a powerful cloud architecture, an engine for neural networks synthesized in the FPGAs, and a compiler and runtime environment to roll out trained models.
Connected directly to the network
The FPGAs are accommodated as PCIe cards in the servers, but are connected directly to the network of the data center via their own interfaces, so that the CPU of the server can be bypassed. Likewise, the FPGAs available in this way can easily be combined into pools. Google is pursuing a similar strategy for connecting its TPUs.
In order to implement the neural networks in the FPGAs, Microsoft relies on the so-called soft-processor concept. This DPU (Deep Neural Network Processing Unit) is more flexible compared to ASICs, since different data types can be chosen at the time of DPU synthesis. Furthermore, thanks to the FPGAs, innovations in the field of machine learning can be performed within the data centers within only a few weeks.

Wide support for software and models
Brainwave can be used to date for machine learning applications that have been implemented with Microsoft's cognitive toolkit or with Google's tensorflow. Support for other frameworks should follow. The models are then first transformed into a graph representation and then compiled for the brainwave infrastructure.
Microsoft also points out that the FPGA system can be used not only for the easily parallelizable Convolutional Neural Networks (CNN), but also especially for the acceleration of recurrent neural networks (RNN) such as LSTMs or those with GRUs. RNNs contain a kind of return channel and are used, for example, to process natural speech.
Almost 40 TOPS thanks to Altera FPGAs
As FPGA hardware, Microsoft uses the Stratix 10 of the Intel subsidiary Altera. According to manufacturers, the chips offer a computing speed of 10 teraflops with single precision (FP32). According to Microsoft tests, the Brainwave system achieves nearly 40 teraflops with reduced precision of 8-bit (FP8), which is common in machine-learning, when running a GRU model.
For comparison: The second generation of Google's TPU, which can also be used via the Google cloud, offers 180 Teraops (INT8) per board and therefore approximately 45 Teraops per ASIC.

Related

Chris Cornell Dead Wishes

Chris Cornell Dead Wishes Chris Cornell Dead Wishes Vlogs Rihanna, Kanye West, Paul

Der Messerattentäter of Hamburg apparently also considered an attack with a truck on

All healthy: Nico Rosberg for the second

16 September 2017 – 10:39 am Daughter Naila and mother Vivian are doing

Rihanna Rude Boy Mp3

Rihanna Rude Boy Mp3 Rihanna Rude Boy Mp3 Vlogs Bobby Shmurda – Hot