A couple of weeks ago at the Ignite conference in Atlanta, Microsoft announced an upgrade to the underlying software defined network (SDN) infrastructure which powers the Azure public cloud – an upgrade which makes Azure the fastest and most Artificial Intelligence-optimized network platform amongst cloud providers.
Some time ago in what was known as ‘Project Catapult’ (no relation to Catapult Systems!) Microsoft carried out research of the adaptation of field-programmable gate arrays (FPGAs). The aim of the project was to enhance performance of the Bing search engine.
FGPAs are chips (i.e. hardware) which, in this case, were added to Altera PCIe cards and programed to allow certain aspects of the search engines ranking algorithm to take place on the chip rather than via software which is processed on regular CPUs – in short, hardware-accelerated implementations of software-defined algorithms.
These chips not only provide the efficiency of hardware but also the agility of software since they can be programmed on-the-fly to manipulate their functionality. The results saw an increase in performance of Bing response times of around 40x.
Azure network experts took the results of this research and applied it to try and eradicate the bottleneck previously inherent in the cloud infrastructure – the underlying network. By adopting FPGAs on the SDN speeds of 25Gbps were achieved, up from 10Gbps. In additional, latency times are also reduced by up to 10%.
Combined with newly available upgraded GPU nodes in the Azure Portal, nodes which are capable of running neuro networking-based applications, cognitive computing, and not least AI, the Azure public cloud has all the makings of being able to operate as the fastest supercomputer on the planet.
As Doug Berger of Microsoft Research stated at Ignite “this new architecture that we’ve built effectively embeds an FPGA-based AI supercomputer into our global hyperscale cloud. We get awesome speed, scale and efficiency. It will change what’s possible for AI.”
Burger also went on to say that Microsoft is using a special type of neural network called a “convolutional neural net,” which has the ability to recognize the actual content of a collection of images.
The addition of FPGAs effectively supercharges a server resulting in a significant boost for tasks related to accelerated AI, and at hyperscale too. Without doubt, these advances made by Microsoft give the Azure platform a distinct advantage in the cloud provider arena.