Slashdot Mirror
Google's Tensor Processing Unit Could Advance Moore's Law 7 Years Into The Future (pcworld.com)
An anonymous reader writes from a report via PCWorld: Google says its Tensor Processing Unit (TPU) advances machine learning capability by a factor of three generations. "TPUs deliver an order of magnitude higher performance per watt than all commercially available GPUs and FPGA," said Google CEO Sundar Pichai during the company's I/O developer conference on Wednesday. The chips powered the AlphaGo computer that beat Lee Sedol, world champion of the game called Go. "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)," said Google's blog post. "TPU is tailored to machine learning applications, allowing the chip to be more tolerant of reduced computational precision, which means it requires fewer transistors per operation. Because of this, we can squeeze more operations per second into the silicon, use more sophisticated and powerful machine learning models, and apply these models more quickly, so users get more intelligent results more rapidly." The chip is called the Tensor Processing Unit because it underpins TensorFlow, the software engine that powers its deep learning services under an open-source license.
Moore's law relates to the number of components in an integrated circuit.
I really doubt these things put more transistors onto a piece of silicon.
It's because today was Google I/O, Google's developer conference, so a lot of projects were announced. Still, Slashdot could have combined all these into story with links to details rather than spam us with a ton of Google stories. However, this TPU project might be the most interesting thing to come out of the conference. Not because the chips are novel (it's just the same principles as GPUs but taken to a further extreme), but because it sounds like Google's getting into low level chip manufacture. We'll have to wait and see if Google can deliver more FLOPS per dollar/watt than the leading co-processor manufacturer, NVIDIA.
Specialized processing chips have been several 'generations' ahead in terms of processing per dollar for many decades. In the 90's at least, DSPs were doing audio/video processing much cheaper by performing many machine-level steps simultaneously in one 'cycle' with less power than a general processor, by leaving out the features and cost of a general processor. And all you had to do to use them was test them on a hardware emulator, flash them, then pop them into production test run until you were good enough to deploy. Depending on the chip, they could run on a trickle of power, without active cooling, and match a much most costly general chip for pennies.
I mean, it's how we got cell phones, and LOTS of other things, including most things in a computer that aren't the CPU.
But isn't Moore's law more about transistors per unit cost, rather than performance per cost? Seems like a fundamental misunderstanding in the headline... which seems about as common as specialized chips in modern technology.
Ryan Fenton
It's easy to disparage the efforts of somebody trying big things when they don't go as planned isn't it. But they are trying and the next step they take could be that 'big thing'. What have you done lately?
Tensor Processing Units are not new. SGI used to offer that for their Octane, aimed pretty heavily at the satellite image analysis crowd.
What's amazing (or maybe not, based on their history) is that Google doesn't know what Moore's law says.
"order of magnitude better...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)"
Uh, no. Moore's law says nothing at all about performance. It speaks to the number of transistors. It was Dave House who predicted a doubling in performance every 18 months (Moore predicted doubling transistor counts every 2 years). Both were based on the size and/or speed of transistors, neither took changes in processor architecture into consideration.
"National Security is the chief cause of national insecurity." - Celine's First Law
Most people who aren't chip architects don't really care one way or another about transistor density, other than that it was a convenient proxy for performance, the frequent doubling of which ordinary people do care about. Now that transistor density has largely hit a physics wall, perhaps we need a new term for the projected trajectory of performance that would have continued had physics allowed transistors to be infinitely small, which engineers are attempting to satisfy by coming up with novel architectures instead.
Google I/O is happening at the moment.
Was that not something introduces about 20 years ago by Silicon Graphics? Or am I getting old
http://manx.classiccmp.org/mir...
Nope. People do care about density, which [...]
haha, you're cute.
Go ask 50 random people on the street if they care about transistor density and report back to us.