Slashdot Mirror
IBM Claims Big Breakthrough in Deep Learning (fortune.com)
The race to make computers smarter and more human-like continued this week with IBM claiming it has developed technology that dramatically cuts the time it takes to crunch massive amounts of data and then come up with useful insights. From a report: Deep learning, the technique used by IBM, is a subset of artificial intelligence (AI) that mimics how the human brain works. IBM's stated goal is to reduce the time it takes for deep learning systems to digest data from days to hours. The improvements could help radiologists get faster, more accurate reads of anomalies and masses on medical images, according to Hillery Hunter, an IBM Fellow and director of systems acceleration and memory at IBM Research. Until now, deep learning has largely run on single server because of the complexity of moving huge amounts of data between different computers. The problem is in keeping data synchronized between lots of different servers and processors In it announcement early Tuesday, IBM says it has come up with software that can divvy those tasks among 64 servers running up to 256 processors total, and still reap huge benefits in speed. The company is making that technology available to customers using IBM Power System servers and to other techies who want to test it.
Seriously. Does IBM actually make products anymore? "Deep Learning"?? Really? IBM, can you tell me where I can buy a Deep Learning? How about a Watson? How about a Cognitive Computing System? Can I buy a Big Data, please? From a technology standpoint, IBM has completely jumped the shark with all of this platform-y, non-productized, framework-y bullshit that requires millions in services hours to implement one-off solutions.
IBM used to make real contributions from their research division into actual software products. Postfix, anyone? RISC technology with AIX and the RS-6k was revolutionary. Their virtualization innovations became the foundation of the AS/400. But no. They jettisoned all of that.
They only do two things now: 1) Research for marketing releases to keep their stock price stable, and 2) Add cash-cow products to their portfolio through acquisition, call them "cognitive"/"big data"/"deep xxxxx", and offshore dev and tech support to a country which charges the lowest wages in the world.
They really have just went down the tubes. It is no wonder that they have declining revenues for so many quarters that I lost count.
Anything labeled "tensor" just means that it does matrix multiplications according to the principles of linear algebra. The reason why 3D game GPUs are popular for this is because gaming has used this technique since at least the Quake engine was invented. Basically everything that happens in a game rendering engine is just setting up a matrix and then multiplying a bunch of them together in a particular order.
What you see with things like the TPU is more dedicated hardware that is optimized for matrix and linear operations. These would be more optimized for AI workloads and have busses and memory pipelines designed for these problem sets rather than things like textures or shaders in a GPU. Mathematically speaking they are all doing the same thing, though.
So, essentially any piece of software that relies on these techniques can benefit from dedicated hardware acceleration and many products offer several backends that will support various hardware accelerator platforms. The recent article on a USB-based one comes to mind. Google's TensorFlow can run on a native CPU, a GPU, or dedicated CUDA-based hardware cards, also. Part of the advantage of using TensorFlow is having an abstraction over this hardware, actually. IBM's effort here mirrors other's in an attempt to distribute huge data workloads across many machines efficiently. There is a big problem in AI right now with idle cores waiting for data to load and sync.
The data workloads start at multiple TBs and sometimes can't even be hosted on a single machine. This is about distributing workloads to multiple machines that may then have dedicated hardware accelerators attached.