The future of AI is ... analog? Upstart bags $100M to push GPU-like brains on less juice

AI chip startup EnCharge claims its analog artificial intelligence accelerators could rival desktop GPUs while using just a fraction of the power. Impressive — on paper, at least. Now comes the hard part: Proving it in the real world. The outfit boasts it has developed a novel in-memory compute architecture for AI inferencing …

[Inventor of the Marmite Laser]

Dear Lord. I remember 741 and CA3140 op amps.

Back in the days when you couldn't mix up electronic components and spilt grains of sugar, component legs went through boards, there was proper solder with lead and tin in it and power transistors were almost all 2N3055 and 2N2905.in T03 cans.

[Inkey]

Guitar amps was as close as i got with those the hum and the smell, and they had to warm up for a short time ....

[Martin Gregorie]

What does this mean for would-be AI bitbarn build-out?

Seems like only yesterday I was wondering whether AI bit-barn building would need to be restricted due to national electric grid and generation capacity (or lack of same), and now suddenly the current generation of GPU based AI servers are about to be scrapped as power-hungry junk,

Result!

[Inkey]

Antikythera mechanism

I seem to remenber a veratasium video on youtube (waited ages for it to get there after promoed)

The kicker was that the study of this and other ingeious devices like the tide predicter that definately had a hand in giving the royal navey the... er upper hand, but that led to modern studies of using logic in analog ways ... I wonder if this start up was one of the group of researchers that took part in that video... recomend watching it ... was epic

[HuBo]

Cool tech

Seems to me this analog route does make the most sense when it comes to biomimetic AI, neuromorphic, or not -- as long as it can be trained (say by error back-propagation). There's healthy competition in this space, like Blumind's AMPL, that also claims 20x improvements in energy use, and Polyn's NASP, among others. Great to see EnCharge securing this $100M Series B Funding imho, especially seeing how their tech can be made available as 28nm Macros (iiuc)!

More broadly, I would expect this type of tech to remain useful through the transition of "AI" from its current trouser-meltingly exciting euphuistic pleonasm implementations, where any reasonable stochastic next-word recall output requires the prior rote ingestion of 82 TBs of pirated Wikipedia contents, to something maybe a bit more parsimonious and animal-like, but named after Turkmen suffering (or Jepa), rather than an actual animal (or Llama), for example. The sooner we move beyond the Fine Young Cannibals "drive me crazy" idol worship of increasingly bodacious models of language, and into the worship of models of models instead, the better imho!

[mevets]

Re: Cool tech

Original message:

---------

Great analysis, I hope you invest heavily....

I also hope that nobody else follows the obvious grift behind this.

Autonomous Illucidation is a hilariously wonderful technology.

It has reduced even noted legal firm to babbling incompetents.

Somehow, apparently the jaggy logic of quantization is the flaw in the hallucination industry?

It can only be fixed by a can of op-amps and capacitors?

Well done.

Keep the grift alive.

---------

On re-read -- yay me stopped at the first period -- now recognizes some sort of travesty generator was at work.

Well done!

[mevets]

Fine line.

I believe there is a fine line between AI-grift, and every other AI-thing.

I suppose a reduction equation would read: AI * -> AI grift.

But the maths are beyond me.

[steelpillow]

Capacitors

An early application of on-chip analogue capacitors was the charge coupled device (CCD) used in digital cameras: every photon pings up an electron, collect 'em up and you have the light level. Pass it to the output via a series of such "charge buckets".

Trouble with compute is, that the various laws of electricity and magnetism, assembled by Michael Faraday and his ilk, are not Turing complete. Hence the need for hi-tech rethinks of the technology. Memristors proved a false dawn, maybe resurrecting the CCD is the true path - fingers crossed.

[Anonymous Coward]

Re: Capacitors

The CCD image sensor was an accident. They were working on storing data and found that light affected the charge in the cells.

I'm not picking fault with your analysis, just mentioning this to show we've come full circle.

[Bebu sa Ware]

Full circle

When I was† a wee lad and first learnt of computers from a library book that text spend quite a few chapters on analogue computers and the problems that they were used to solve and later chapters on digital computers with a conclusion that hybrid (analogue + digital) computers were the future - the author clearly not a frontrunner in the Nostradamus stakes.

But I guess if you prophesy too far into the future you and your prophecies will be long forgotten before their fruition.

My take on this "new" technology is that it's pretty much a late 1960s hybrid computer element miniaturized to nano scale with massive on chip replication.

I could see that an element that could solve a range of differential equations replicated millions of time could run the same model with a million different boundary and initial conditions in parallel might be very useful.

Even back then (1960s) early AI were building neuristors and basic neural networks from such basic electronics. The move from physical devices to modelling the same devices and networks in software probably happened during the 1970s and the rest is history and the lamentable state we find ourselves in today.

I cannot help but think if you were to make a large number of adjacent compound analogue-digital neuron like elements small enough then quantum mechanical interactions might provides some unexpected surprises.:)

_{† I instinctively feel that this should be "were" but cannot see any reason why.}

[LogicGate]

Re: Full circle

One of my favorite books is Heinlein's "The Moon is a Harsh Mistress".

However, one thing that used to irk me was the description of the Computer, Mike" developing his digital "face" in realtime:

"We waited in silence. Then screen showed neutral gray with a hint of scan lines. Went black again, then a faint light filled middle and congealed into cloudy areas light and dark, ellipsoid. Not a face, but

suggestion of face that one sees in cloud patterns covering Terra. It cleared a little and reminded me of pictures alleged to be ectoplasm. A ghost of a face.

Suddenly firmed and we saw "Adam Selene."

Was a still picture of a mature man. No background, just a face as if trimmed out of a print. Yet was, to me, "Adam Selene." Could not he anybody else. Then he smiled, moving lips and jaw and touching tongue to lips, a quick gesture--and I was frightened."

Being used to polygon based 3d graphics, this seemed wrong to me. It did not bother me too much, because the book was obviously written during the infancy of computing.

Looking back, I must say that Heinlein's description reads scarily close to an AI / Neural network being trained. The text improves with age (although the same can not be said for all of Heinleins work).

[xanadu42]

Re: Full circle

If simple analogue electronics can perform better than it's very complex digital equivalent at a similar task then it suggests we must re-evaluate where we are heading

...

When I was turned 9 (in the very early 1970's) my parent's birthday present was a "100 in one" electronics kit based on my (way too obvious to them) interest in electricity...

Before I received this electronics kit I had done many stupid and dangerous things (in retrospect) a few months before with "240V AC mains" like digging out the wall-mounted socket to see how it worked and connecting my sister's battery-powered food-mixer toy into the Mains socket (thinking that it would go faster!) after destroying said "food-mixer" and getting the wires to the battery compartment inserted into the Mains socket!

Needless to say a big bang from the food-processor and sudden darkness (didn't know about fuses) OOPS!

I also looked inside the radios and TV's my parents had (for which I was not found out - I did tell them a decade or two later)

My parents immediately knew the cause (which obviously I tried to deny to no avail) and I was in deep-sh*t for months - but I got that "100 in one" electronics kit...

Learned all about NPN and PNP Transistors, Diodes, Resistors, Capacitors and the difference between a battery and a cell...

My Dad even helped me when I built the "Crystal Radio" part of the kit by putting up a 30-odd-foot piece of wood with wire attached so so that the "Crystal Radio" worked with a good audio volume!

I absolutely loved this analogue electronics...

Around 17 years of age a friend of mine showed me an "electronic computer" - around an A4 sized circuit board with a small number if "Integrated Circuits", four seven-segment LEDs, 256 bytes of "RAM" and 10(?) switches

Didn't understand what was meant by "Integrated Circuit" or "RAM" but my friend showed me how, by manipulating the switches in a convoluted pattern he could get the display to say "boob" after much flashing of the LEDs

I was Hooked and "the digital world" became my career...

But I lament the day where everything analogue became digital...

The "Transistor Radio", a simple and obvious extension of the "Crystal Radio" - Digital Radios are so boring in comparison - a "black blob" and "nothing obvious" about how it works...

Pretty much the same for all current electronics - no avenue for curiosity as it is ALL "black blobs"

Analogue world we live in so analogue WILL rule :)

[Andrew Scott]

Re: Full circle

blew a fuse once grounding the chassis of a 120 ac radio to a radiator. thought chassis ground was the same thing as my crystal radio ground. scar;y stuff in retrospect.

[JeffRT]

Isn't digital inefficient to start with?

Just thinking out loud here, and theoretically at that- a digital bit seems to be a waste of space from an information capacity perspective. You need 8 separate bits to store or compute an 8-bit number, but if the information is encoded within that same voltage range, then it would be significantly more efficient. After all, our brains have been biologically optimized for maximum compute per unit of storage for least amount of energy. That said, I realize that the digital works well for lots of things, but canting this analog approach towards AI seems natural to me and it seems to me that there is a lot of room for efficiency to improve.

[steelpillow]

Re: Isn't digital inefficient to start with?

Often, it's more to do with the manufacturing technology. Analogue can be efficient but it also tends to be bulky. It is also not very flexible: too much has to be hardwired or switched in and out, and a more complex equation may need a bigger processor circuit not just longer runtime.

Ternary -1, 0, +1 logic is theoretically more efficient that binary 0, 1 but is an arse to manufacture.

[druck]

Analogue might be low power...

...but it's also far slower than digital logic, just like biological brains.

[Richard 12]

Re: Analogue might be low power...

But if you can have more of them for the same power budget, it could still be faster.