Nowhere is this more pertinent than in the analysis of huge volumes of data, in applications such as big data and, increasingly, machine learning. In search of high performance in these applications, the leading cloud providers are investing in increasingly esoteric processor technologies, even to the extent of custom-building their own designs, as both Google and IBM are known to have done, while Microsoft and Amazon are thought to be doing the same — perhaps also Apple and Facebook.
Google's Tensor Processing Unit (TPU) is a case in point. Google first released details of the custom design during its I/O developer conference in May, revealing in an accompanying blog post that it was already using TPUs in several production systems:
TPUs already power many applications at Google, including RankBrain, used to improve the relevancy of search results and Street View, to improve the accuracy and quality of our maps and navigation. AlphaGo was powered by TPUs in the matches against Go world champion, Lee Sedol, enabling it to 'think' much faster and look farther ahead between moves.
The TPU's role is to help Google deliver machine learning at production scale. Rather than a general-purpose CPU or GPU, each of which can perform a variety of different tasks, the TPU is a custom ASIC built to perform a specific set of tasks highly efficiently — in this case, applying machine learning inferences to data, which it does exceedingly well, according to Google:
We’ve been running TPUs inside our data centers for more than a year, and have found them to deliver an order of magnitude better-optimized performance per watt for machine learning. This is roughly equivalent to fast-forwarding technology about seven years into the future (three generations of Moore’s Law).
Its name derives from the word 'tensor', a multi-dimensional matrix that represents the learning stored in a neural network. The TPU is optimized to work with the Google-designed TensorFlow machine learning library. Google says it took just 22 days from receiving tested silicon to bring the chips live in its datacenters, which suggests an iterative approach to the design lifecycle.
AI accelerators — DPUs?
The TPU is an example of a new breed of AI accelerators that are more akin to the highly optimized digital signal processors (DSPs) developed to accelerate audio and video streaming in the 1990s than they are to more general-purpose processors. If this new breed needs its own overarching acronym alongside CPU and GPU, perhaps it should be DPU (for Data Processing Unit).
Rising demand for accelerators that can deliver more efficient data analysis and machine intelligence is fueling a wave of innovation in microprocessor hardware. Another example to add to my putative DPU classification is the Micron Automata, which as I wrote earlier this year, is a programmable accelerator for high-performance pattern matching:
Built to analyze huge datasets in parallel, Automata has a reconfigurable processing architecture that Micron believes has applications across graph analysis, pattern matching, and data analytics. Whereas conventional parallelism consists of a single instruction applied to many chunks of data, Automata consists of a DRAM-like fabric of tiny interconnected processing elements that can focus a vast number of instructions at a targeted problem as data is streamed across the chip.
IBM's esoteric processors
IBM is pushing the frontiers of processor research with its initiatives. One example is TrueNorth, the result of a partnership with Cornell University which has had $100m in public funding from the US government's DARPA military research agency. It uses a 'neuromorphic' architecture designed to mimic the way the human brain works. But TrueNorth is seen as experimental and several experts doubt it offers any practical benefits over current technologies.
Even more esoteric is IBM's excursion into quantum computing, which it recently put online as an on-demand service for experimental applications. Quantum processors are expected to have an ability to analyse large amounts of unstructured data far more efficiently than conventional processors, especially when analyzing natural phenomena that themselves exhibit quantum mechanics. But that is still all in theory, as Dr Stefan Fillip, quantum researcher at IBM Research in Zurich, admits:
Making this publicly available will help to answer the one big question of what it is for and develop the field of quantum computing.
Nvidia and Intel
Back in the real world, the more established technologies of GPUs are already playing their role in AI datacenters. Building on NVidia's leadership in this field, its Tesla P100, shipping later this year, is expected to be a highly efficient device for training neural networks — the process by which the 'tensors' that record the learning are created in the first place. As Kurt Marko has discussed, some database makers are also optimizing their software to exploit the capabilities of GPUs to accelerate big data operations:
GPU accelerated databases are superficially analogous to in-memory DBs, however the availability of many more GPU cores, much faster GDDR5 memory and instruction sets optimized for combining, multiplying, summing and filtering complex data sets makes GPUs particularly adept at accelerating many database operations.
Chip giant Intel for its part is catering to the demand for more customizable processor capabilities. Following on from its acquisition of Altera last year, Intel later this year will ship Xeon processors with built-in FPGAs. The purpose is to offer programmable hardware acceleration to address a variety of needs from compression and encryption to analytics and AI.
Who'd have thought cloud computing would end up driving more diversity and innovation in processor design than we've seen since the days when every self-respecting computer manufacturer had its own chip foundry? I agree with Kurt Marko's conclusion:
The days of homogeneous server farms with racks and racks of largely identical systems are over.
Suddenly, it's become important for enterprise IT buyers to pay attention to processor architectures in ways they probably haven't since the 1990s. Whether they're considering how to equip their own data centers or, more likely, evaluating the capabilities of third-party providers, the hardware at the heart of that infrastructure can no longer be taken for granted — especially when evaluating deep learning capabilities.