Google’s Cloud Tensor Processing Units at the moment are accessible in public beta for anybody to strive, offering clients of the tech titan’s cloud platform with specialised hardware that massively accelerates the coaching and execution of AI fashions.
The Cloud TPUs, which Google first introduced final 12 months, work by offering clients with specialised circuits solely for the aim of accelerating AI computation. Google examined utilizing 64 of them to coach ResNet-50 (a neural community for figuring out photos that additionally serves as a benchmarking device for AI coaching velocity) in solely 30 minutes.
This new hardware may assist entice clients to Google’s cloud platform with the promise of sooner machine studying computation and execution. Accelerating the coaching of latest AI methods generally is a important assist, since information scientists can then use the outcomes of these experiments to make enhancements for future mannequin iterations.
Google is utilizing its superior AI capabilities to draw new blood to its cloud platform and away from market leaders Amazon Web Services and Microsoft Azure. Businesses are more and more seeking to diversify their use of public cloud platforms, and Google’s new AI hardware may assist the corporate capitalize on that pattern.
Companies had already lined as much as check the Cloud TPUs whereas they had been in non-public alpha, together with Lyft, which is utilizing the hardware to coach the AI fashions powering its self-driving vehicles.
It’s been an extended street for the corporate to get right here. Google introduced the unique Tensor Processing Units (which solely supplied inference capabilities) in 2016 and promised that clients would be capable to run customized fashions on them, along with offering a velocity increase for different companies’ workloads via the cloud machine studying APIs. But enterprises had been by no means capable of run their very own customized workloads on high of an unique TPU.
Google isn’t the one one to push AI acceleration via specialised hardware. Microsoft is utilizing a fleet of field-programmable gate arrays (FPGAs) to hurry up its in-house machine studying operations and supply clients of its Azure cloud platform with accelerated networking. In the long run, Microsoft is engaged on offering clients with a solution to run their machine studying fashions on high of the FPGAs, identical to the corporate’s proprietary code.
Amazon, in the meantime, is offering its clients with compute situations which have their very own devoted FPGA. The firm can be engaged on creating a specialised AI chip that may speed up its Alexa gadgets’ machine studying computation, in response to a report launched by The Information in the present day.
Actually getting AI acceleration from TPUs received’t be low cost. Google is at present charging $6.50 per TPU per hour, although that pricing might shift as soon as the hardware is mostly accessible. Right now, Google remains to be throttling the Cloud TPU quotas which are accessible to its clients, however anybody can request entry to the brand new chips.
Once folks get entry to the Cloud TPUs, Google has a number of optimized reference fashions accessible that may allow them to begin kicking the tires and utilizing the hardware to speed up AI computation.
This article sources info from VentureBeat