[HN Gopher] New analog chip capable of outperforming top-end GPU...
___________________________________________________________________
New analog chip capable of outperforming top-end GPUs by as much as
1000x
Study: https://www.nature.com/articles/s41928-025-01477-0
Author : mrbluecoat
Score : 44 points
Date : 2025-11-01 04:09 UTC (18 hours ago)
(HTM) web link (www.livescience.com)
(TXT) w3m dump (www.livescience.com)
| alexnewman wrote:
| What's this good for?
| andrewstuart wrote:
| Fear
| falseprofit wrote:
| Computing, they propose
| teruakohatu wrote:
| Faster than an H100 for solving 128x128 matrices. But it's not
| clear to me how they tested this, code is only available on
| request.
|
| > We have described a high-precision and scalable analogue matrix
| equation solver. The solver involves low-precision matrix
| operations, which are suited well to RRAM-based computing. The
| matrix operations were implemented with a foundry-developed 40-nm
| 1T1R RRAM array with 3-bit resolution. Bit-slicing was used to
| guarantee the high preci- sion. Scalability was addressed through
| the BlockAMC algorithm, which was experimentally demonstrated. A
| 16 x 16 matrix inversion problem was solved with the BlockAMC
| algorithm with 24-bit fixed-point preci- sion. The analogue
| solver was also applied to the detection process in massive MIMO
| systems and showed identical BER performance within only three
| iterative cycles compared with digital counterparts for 128 x 8
| systems with 256-QAM modulation.
| alyxya wrote:
| This looks like one of many ideas for more efficient compute
| chips for machine learning. I'm waiting for the day some chip
| gets mass produced and works at scale for some large model and
| with sufficient reliability, but until then, I don't think
| there's anything particularly newsworthy here. I do think it'll
| eventually happen at some point maybe within a decade, but surely
| some alternative computing paradigm to the GPU will succeed. The
| analog chip in the article only seems to be a research prototype
| for now.
| drnick1 wrote:
| Seems a bit too good to be true.
| gnarlouse wrote:
| Huge if true, room temperature semiconductor if false
| makapuf wrote:
| Semi or supra conductor ?
| generuso wrote:
| The idea was always appealing, but the implementation has always
| remained challenging.
|
| For over a decade, "Mythic AI" was making accelerator chips with
| analog multipliers based on research by Laura Fick and coworkers.
| They raised $165M and produced actual hardware, but at the end of
| 2022 have almost gone bankrupt and since then there has been very
| little heard from them.
|
| Much earlier, the legendary chip designers Federico Faggin and
| Carver Mead founded Synaptics with an idea to make neuromorphic
| chips which would be fast and power efficient by harnessing
| analog computation. Carver Mead published a book on that in 1989:
| "Analog VLSI and Neural Systems", but making working chips turned
| to be too hard, and Synaptics successfully pivoted to touchpads
| and later many other types of hardware.
|
| Of course, the concept can be traced to an even older and still
| more legendary Frank Rosenblatt's "Perceptron" -- the original
| machine learning system from 1950s. It implemented the weights of
| the neural network as variable resistors that were adjusted by
| little motors during training. Multiplication was simply input
| voltage times conductivity of the resistor producing the current
| -- which is what all the newer system are also trying to use.
| rasz wrote:
| I know of only one real world successful product using analog
| computation in place of expensive high end micro. It was the
| first proper (no dedicated special mousepads) Optical Mouse
| designed and build by HP->Agilent->Avago and released by
| Microsoft in 1999 as IntelliMouse Optical.
| https://gizmodo.com/20-years-ago-microsoft-changed-how-we-mo...
| Afaik Microsoft bought 1 year explosivity for the sensor. Avago
| HDNS-2000 chip did all the heavy lifting in analog domain.
|
| Travis Blalock Oral History
| https://www.youtube.com/watch?v=wmqa9XJED-Q https://archive.com
| puterhistory.org/resources/access/text/20...:
|
| "each array element had nearest neighbor connectivity so you
| would calculate nine correlations, an autocorrelation and eight
| cross-correlations, with each of your eight nearest neighbors,
| the diagonals and the perpendicular, and then you could
| interpolate in correlation space where the best fit was."
|
| "And the reason we did difference squared instead of
| multiplication is because in the analog domain I could
| implement a difference-squared circuit with six transistors and
| so I was like "Okay, six transistors. I can't do multiplication
| that cheaply so sold, difference squared, that's how we're
| going to do it."
|
| "little chip running in the 0.8 micron CMOS could do the
| equivalent operations per second to 1-1/2 giga operations per
| second and it was doing this for under 200 milliwatts, nothing
| you could have approached at that time in the digital domain."
|
| Extra Oral History with inventor of the sensor Gary Gordon:
| https://www.youtube.com/watch?v=TxxoWhCzIeU
| smartbit wrote:
| The idea of analog neural networks is appealing. I bought
| _Analog VLSI and Neural Systems_ in 1989 and still have it as a
| trophy on my bookshelves. My gut feeling says _one day_ analog
| neural networks will be a thing, if only for the reason of
| considerable lower power consumption.
|
| I'm not saying that life is analog, DNA is two bits. IMHO life
| is a mix of Analog & Digital.
| ConteMascetti71 wrote:
| Using all analog signal, why non analogue multiplying cells
| (operation amplifier)!
| Archit3ch wrote:
| Now put it in a guitar pedal!
| smitty1e wrote:
| Wo Fat has you covered: "Analog Man" =>
| https://open.spotify.com/track/6KcM6et6Pn6UIna1o8Vl07?si=qFu...
| xeonmc wrote:
| Wonderful, can't wait to run Crysis with this chip.
___________________________________________________________________
(page generated 2025-11-01 23:01 UTC)