Data-Hungry Algorithms and the Thirst for AI
At Tabor Communications’ Leverage Big Data + EnterpriseHPC Summit in Florida last week, esteemed HPC professional Jay Boisseau, chief HPC technology strategist at Dell EMC, engaged the audience with his presentation, “Big Computing, Big Data, Big Trends, Big Results.”
Trends around big computing and big data are converging in powerful ways, including the Internet of Things (IoT), artificial intelligence (AI) and deep learning. Innovating and competing is now about big, scalable computing and big, fast data analytics – and “those with the tools and talent will reap the big rewards,” Boisseau expressed.
Prior to joining Dell EMC (then Dell Inc.) in 2014, Boisseau made his mark as the founding director of the Texas Advanced Computing Center (TACC). Under his leadership the site became a center of HPC innovation, a legacy that continues today under Director Dan Stanzione.
“I’m an HPC person who’s fascinated by the possibilities of augmenting intelligence with deep learning techniques; I’ve drunk the ‘deep learning Kool-Aid,’” Boisseau told the crowd of advanced computing professionals.
AI as a field goes back to the 50s, Boisseau noted, but the current proliferation of deep learning using deep neural networks has been made possible by three advances: “One is that we actually have big data; these deep learning algorithms are data hungry. Whereas we sometimes lament the growth of our data sizes, these deep neural networks are useless on small data. Use other techniques if you have small data, but if you have massive data and you want to draw insights that you’re not even sure how to formulate the hypothesis ahead of time, these neural network based methods can be really really powerful.
“Parallelizing the deep learning algorithms was another one of the advances, and having sufficiently powerful processors is another one,” Boisseau said.
AI, big data, cloud and deep learning are all intertwined and they are driving rapid expansion of the market for HPC-class hardware. Boisseau mines for correlations with the aid of Google Trends; the fun-to-play-with Google tool elucidates the contemporaneous rise of big data, deep learning, and IoT. Boisseau goes a step a further showing how Nvidia stock floats up on these tech trends.
The narrow point here is that deep learning/big data is an engine for GPU sales; the larger point is that these multiple related trends are driving silicon specialization and impacting market dynamics. As Boisseau points out, we’re only at the beginning of this trend cluster and we’re seeing silicon developed specifically for AI workloads as hardware vendors compete to establish themselves as the incumbent in this emerging field.
Another deep learning champion Nvidia CEO Jen Hsun Huang refers to machine learning as HPC’s consumer first killer app. When Nvidia’s CUDA-based ecosystem for HPC application acceleration launched in 2006, it kick started an era of heterogeneity in HPC (we’ll give the IBM-Sony Cell BE processor some cred here too even if the processor design was an evolutionary dead end). Fast forward to 2013-2014 and the emerging deep learning community found a friend in GPUs. With Nvidia, they could get their foot in the DL door with an economical gaming board and work their way up the chain to server-class Tesla GPUs, for max bandwidth and FLOPS.
Optimizations for single-precision (32-bit) processing, and support for half-precision (16-bit) on Nvidia’s newer GPUs, translates into faster computation for most AI workloads, which unlike many traditional HPC applications do not require full 64-bit precision. Intel is incorporating variable precision compute into its next-gen Phi product, the Knights Mill processor (due out this year).
Boisseau observed that starting about two decades ago HPC began the swing towards commodity architectures, with the invention of commodity-grade Beowulf clusters by Thomas Sterling in 1994. Benefiting from PC-based economies of scale, these x86 server-based Linux clusters became the dominant architecture in HPC. In turn, this spurred the movement toward broader enterprise adoption of HPC.
Although Xeon-flavored x86 is something of a de facto standard in HPC (with > 90 percent share), the pendulum appears headed back toward greater specialization and greater “disaggregation of technology,” to use a phrase offered by industry analyst Addison Snell (CEO, Intersect360 Research). Examples include IBM’s OpenPower systems; GPU-accelerated computing (and Power+GPU); ARM (now in server variants with HPC optimizations); AMD’s Zen/Ryzen CPU; and Intel’s Xeon Phi line (also its Altera FPGAs and imminent Xeon Skylake).
A major driver of all this: a gathering profusion of data.
“In short, HPC may be getting diverse again, but much of the forcing function is big data,” Boisseau observed. “Very simply, we used to have no digital data, then a trickle, but the ubiquity of computers, mobile devices, sensors, instruments and user/producers has produced an avalanche of data.”
Buzz terminology aside, big data is a fact of life now, “a forever reality” and those who can use big data effectively (or just “data” if the “big” tag drops off), will be in a position to out-compete, Boisseau added.
When data is your pinnacle directive and prime advantage, opportunity accrues to whoever holds the data, and that would be the hyperscalers, said Boisseau. Google, Facebook, Amazon, et al. are investing heavily in AI, amassing AI-friendly hardware like GPUs but also innovating ahead with even more efficient AI hardware (e.g., Tensor Processing Units at Google, FPGAs at Microsoft). On the tool side are about a dozen popular frameworks; TensorFlow (Google), mxnet (Amazon), and CNTK (Microsoft) among them.
Tech giants are advancing quickly too with AI strategies, Boisseau noted. Intel has made a quick succession of acquisitions (Nervana, Movidius, Saffron, MobilEye); IBM’s got its acquisition-enhanced Watson; Apple bought Turi.
“You [also] have companies like GraphCore, Wave Computing, and KnuPath that are designing special silicon with lower precision and higher performance,” said Boisseau. “There was a fourth one, Nervana, and Intel liked that company so much they bought it. So there were at least four companies making silicon dedicated to deep learning. I’m really eager to see if Nvidia – and I don’t have inside knowledge on this – further optimizes their technology for deep learning and removes some of the circuitry that’s still heritage graphics oriented as well as how the special silicon providers do competing against Intel and Nvidia as well as how Intel’s Nervana shapes up.”
Adding to the cloud/hyperscaler mix is the quickly expanding world of IoT, which is driving big data. The Internet of Things is enabling companies to operate more efficiently; it’s facilitating smart buildings, smart manufacturing, and smart products, said Boisseau. But as the spate of high-profile DDoS attacks attest, there’s a troubling security gap. The biggest challenge for IoT is “security, security, security,” Boisseau emphasized.
Another top-level point Boisseau made is that over half of HPC systems are now sold to industry, notably across manufacturing, financial services, life sciences, energy, EDA, weather and digital content creation. “Big computing is now as fundamental to many industries as it is in research,” Boisseau said. Half of the high performance computing TAM (total addressable market), estimated at nearly $30 billion, is now in enterprise/industry, and there’s still a lot of untapped potential, in Boisseau’s opinion.
Market projections for AI are even steeper. Research houses are predicting that AI will grow to tens of billions of dollars a year (IDC predicts a surge past $4 billion in 2020; IBM expects market to be $2 trillion over next decade; Tractica plots $3.5 trillion in revenue by 2025).
Boisseau is confident that the world needs big data AND deep learning, citing the following reasons/scenarios:
- Innovation requires ever more capability: to design, engineer, manufacture, distribute, market and produce new/better products and services.
- Modeling and simulation enable design, in accordance with physics/natural laws, and virtual engineering, manufacturing, testing.
- Machine learning and deep learning enable discovery and innovation
- When laws of nature don’t apply (social media, sentiment, etc.) or are non-linear/difficult to simulate accurately over time (e.g. weather forecasting).
- That may be quicker and/or less costly depending on simulation scale, complexity versus data completeness.
“When we understand the laws of nature, when we understand the equations, it gives us an ability to model and simulate highly accurately,” said Boisseau. “But for crash simulations, we still don’t want to drive a car that’s designed with data analysis; we need modeling and simulation to truly understand structural dynamics and fluid flow and even then data analysis can be used in the interpretation.
“There will be times where data mining over all those crash simulations adds to the modeling and simulation accuracy. So modeling and simulation will always remain important, at least as long as the universe is governed by visible laws, especially in virtual engineering and manufacturing testing, but machine learning and deep learning enable discovery in other ways, especially when the laws of nature don’t apply.”
“If you’ve adopted HPC great, but deep learning is next,” Boisseau told the audience. “It might not be next year for some of you, it might be two years, five years, but I suspect it’s sooner than you think.”