We’re Using The Same Stuff That Powers The Azure and Bing Data Centers. Should You?
“FPGAs have been installed across every Azure cloud server, creating “the world’s first AI supercomputer” according to Mark Russinovich, CTO, Azure in this live presentation. "Configurable Clouds will change the world [with] the ability to reprogram a datacenter's hardware protocols: networking, storage, security", Doug Burger of Microsoft recently said at FPL2016.:
Why is the world's largest software company messing around with programmable hardware? Why are they using FPGAs when they could do the same thing in software? In their own words, it's to "get the performance of hardware with the programmability of software".
Without having to geek out and read/watch all of Microsoft's presentation, here's the key take-aways
- For their applications (Azure, Bing, Office365), they need infrastructure that excels at "scale, low-latency and throughput"
- To accomplish this, they have an FPGA on every server sitting between the server and the network as part of their Catapult architecture
- The FPGA is used for search acceleration (Bing), software-defined-network acceleration and with plans for storage infrastructure acceleration.
- The benefits are "extremely high efficiency... getting really low latency... amazing amounts of data throughput and operations per second."
- Microsoft admits that if they were to use software, they would be "burning cores of expensive processors... introducing latency into the network... and these cores can't be used for what they should be doing which is general purpose processing".
Great minds think alike. it's no coincidence that their motivation matches Attala's. While you might not be thinking about FPGA technology explicitly for your data center, don't be surprised to hear more and more about FPGAs powering your infrastructure in the years to come.
Outside of Microsoft, Attala is the industry's first vendor to offer a broad infrastructure solution using this technology. What this means for you as a customer is a storage infrastructure with low-latency, predictable performance, CPU-less cost-efficiency, agility and flexibility.