Slashdot Mirror
Google's Custom Machine Learning Chips Are 15-30x Faster Than GPUs and CPUs (pcworld.com)
Four years ago, Google was faced with a conundrum: if all its users hit its voice recognition services for three minutes a day, the company would need to double the number of data centers just to handle all of the requests to the machine learning system powering those services, reads a PCWorld article, which talks about how Tensor Processing Unit (TPU), a chip that is designed to accelerate the inference stage of deep neural networks came into being. The article shares an update: Google published a paper on Wednesday laying out the performance gains the company saw over comparable CPUs and GPUs, both in terms of raw power and the performance per watt of power consumed. A TPU was on average 15 to 30 times faster at the machine learning inference tasks tested than a comparable server-class Intel Haswell CPU or Nvidia K80 GPU. Importantly, the performance per watt of the TPU was 25 to 80 times better than what Google found with the CPU and GPU.
Man is this a "duh" moment. Purpose built ASICs are extremely fast and low power for what they accomplish. That's why we use them. Look at a small desktop network switch: Little tiny processor that can pass 16gb/sec of traffic around. try and put 8 NICs in a computer and have it switch traffic and you'll be amazed at how much power you need. The reason the switch is small is it is purpose built: It's ASIC does nothing but switch Ethernet packets.
Same deal with some thing on a CPU. You find that decoding an AVC video stream takes next to no CPU power on modern CPUs, yet decoding an MPEG-2 video takes some. Why? Because they have a small bit of dedicated logic for AVC decoding (usually some other formats too). It is low power because it is dedicated.
Always the question in designing a system is flexibility and unit cost vs fixed function and up front cost. A CPU is great because it can do anything, and you can just buy them straight out, tons of companies have them available for purchase right now. However they take a lot of silicon and power to perform a given task. An ASIC takes a bunch of up front money to design and do a manufacturing run, but is very small and efficient, however it can't be reconfigured to do anything else and needs a full respin. In the middle there is something like an FPGA. Which one is right for a application just depends on the balance of a lot of factors.
You're thinking of a TPM. This is a TPU.
How so?
The TPU is a "purpose built" chip, but that purpose is very broad. It is optimized for massively parallel low-precision matrix operations, which is useful not only for neural nets, but also simulation of physical processes like CFD, weather prediction, climate models, computational chemistry, etc. It can do everything a GPU can do except the rasterization and texture mapping, but it can do it faster and with much less power.
It is optimized for massively parallel low-precision matrix operations, which is useful not only for neural nets, but also simulation of physical processes like CFD, weather prediction, climate models, computational chemistry, etc.
Maybe, but I doubt it. It's far too low precision, for one thing: 8-bits doesn't get you very far in any of those fields (you typically want at least 32-bit FLOPS for those, and quite often 64-bit precision is required, as numerical errors accumulate exponentially in a chaotic system), and they're really not even big matrices (just 256x256). Really the only place this kind of thing would excel is signal processing, which is basically what they're using them for.
"None can love freedom heartily, but good men; the rest love not freedom, but license." --John Milton
They are 15-30 times faster, not 15-30%.
Every little order of magnitude really helps :)
When all you have is a hammer, every problem starts to look like a thumb.