Google's TPUs could end up costing it a billion-plus, thanks to this patent challenge Allegations that Google's Tensor Processing Units (TPUs) were developed using stolen designs are being put to the test as a jury trial brought against the search giant by Singular Computing kicks off this week. Google is accused of infringing patents held by Singular and developed by computer scientist Joseph Bates, an …
And there are no email traces whatsoever of communications between Google's Left Hand and Google's Right Hand, or won't be once the "errors" in the message archiving and retrieval software are done with things. (Never happened before. Nope. Purely an "accident." *sigh*)
...and even if they didn't, independently coming up with an "invention" you've never seen or heard of before doesn't stop it from being a patent infringement if it's works in exactly the same way as an existing patented device. More so if the company in question has actual provable knowledge of the patents beforehand and should have, at the very least, passed on those patents to their design team and told them to avoid and or work around them. Passing the patents on might even have helped speed up a different design process not only because sometimes just knowing a thing is possible can be half the battle, but they could look at a working design and more easily invent alternatives[*]
On the other hand, it's the US Patent Office who granted the patents so they may be worthless anyway.
*Yes, I'm aware that's not always the case, but sometime it is :-)
In 2005, Bates founded Singular Computing to commercialize various computing architectures. According to Singular's website, the biz "develops and licenses hardware and software technologies for high-performance energy-efficient computing, both large scale and embedded."
That to me sounds like patent trolling. Just wrap any brain fart into a patent lingo, submit where committee has no clue and waves everything through. Wait for someone to release a product remotely related and is making money. Sue.
patent trolls typically patent something that's not necessarily real tech, but looks good on paper and in a courtroom. get in first with your "design".
if you built one you were actually "inventing" or at least that's the story they're telling
IANAL - however this is what patents are supposed to cover, an actual invention you might have had along the way in your development work
First patent I was involved in was used for demos and was worked on for a year or so. never sold, though, but it definitely worked. Tech changed around it and it was just not enough to make its own niche in the market. (it was related to wifi and has my name on it). At this time there were a LOT of things getting quickly obsoleted and/or replaced with newer/better
Yep the definition of patent troll in my mind is someone who files a patent with little or no intention of producing anything with it other than lawsuits. Or when someone who files a patent and maybe hopes to do something useful with it is unable to for whatever reason and sells it to a company that has no intention of producing anything with it other than lawsuits.
Sounds like this guy not only wanted to do something useful with it, he actually build a prototype. But "doing something useful" with it involved a deep pocketed partner, which he hoped Google was. That didn't work out, and it sounds like he's only suing because he believes Google is using his patent protected technology. We'll see have to see what comes up at trial (or if Google essentially admits they did this by settling with him)
It absolutely is not. That's but one of many reforms patent law should have, but probably won't in our lifetime because apparently congress has more important things to do like talking about the border but not actually doing anything, talking about the budget but not actually doing anything, etc.
An argument that can work for copyright violation defence. Not so much for a regular patent. This is not the fight against Oracle.
Elephants in the room: USPTO allows patents they should not, but that won't help Google/Alphabet. Is in fact much current AI any use and does it do much more than warm the planet (which we don't want), but that won't help this case either.
I remember that if it can be shown that the infringement was deliberate, damages are tripled...? Or something like that.
Also, if Google tries to argue that their method is in fact different than that described on the patent, I suppose it might be argued that they started from the patent and changed a few details to make it different (but that's just my common sense argument, so it's probably absolutely the opposite of what's goes in a patent trial)
RCA chose to purchase a patent for a 3-gun color TV picture tube to a man named Geer who constructed a prototype color picture tube that looked like a 3 legged octopus, though if I remember correctly his first conceptual prototype (showing the 'pyramids') was constructed with sugar cubes and mayonnaise. There was enough similarity that the RCA patent could have been challenged and they wanted to hedge their bets by buying the competing design.
https://en.wikipedia.org/wiki/Geer_tube <-- a true hacker
The ideas behind the TPU or similar have been around for decades but its only been relatively recently that the technology has been available to realize them on a large scale. Its likely that Singular brought their ideas to Google hoping to sell them but after review they were not considered unique enough to proceed with licensing or an acquisition. Its really unlikely that Singular's material would have been ripped off as the company alleges -- this is asking for trouble and Google probably took proactive steps to prevent them from being sued in the future.
I've been in this situation with another large technology company and they went to considerable lengths to prevent IP leakage contaminating their product line (even though the outside product was functionally superior...).
Hardware implementations of piecewise linear function for neural networks are not new at all: H. Amin; K.M. Curtis; B.R. Hayes-Gill;IEE Proceedings - Circuits, Devices and Systems, Volume 144, Issue 6, December 1997, p. 313 – 317
An efficient piecewise linear approximation of a nonlinear function (PLAN) is proposed. This uses a simple digital gate design to perform a direct transformation from X to Y, where X is the input and Y is the approximated sigmoidal output. This PLAN is then used within the outputs of an artificial neural network to perform the nonlinear approximation. The comparison of this technique with two other sigmoidal approximation techniques for digital circuits is presented and the results show that the fast and compact digital circuit proposed produces the closest approximation to the sigmoid function. The hardware implementation of PLAN has been verified by a VHDL simulation with Mentor Graphics running under the UNIX operating system.
One of many such reports. To evaluate Bates claimed infringement you would have to know
- Algorithms are not patentable, even parallelized ones, thank Hydra of Lerna for that.
- What was original (if anything) about the *implementation* of Bate's prototype compared to what existed more than a decade earlier.
- Did Google's TPU implementation infringe on *that* originality or just create their own common sense implementation of the same material that Bates prototype was based upon?
Implementing an analogue circuit to approximate the Sigmoid function and implementing that in silicon in an efficient and effective manner is very interesting.
Analogue circuits can be very fiddly and easily influenced by noise. Figuring out how to form them on silicon would requite some tricks to prevent them interfering with each other and prevent the rest of the processor from interfering with them. Very easy for analogue circuits to experience resonance with other similar circuits that would degrade their utility.
Creating something that can work in high performance and massively parallel silicon is a significant achievement.
This technology is very valuable in the current machine learning bottleneck.
Such technology is definitely patentable and worthy of protection.
It is also easy to steal.
It will be interesting to see if Google gets away with it.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
