[HN Gopher] A New Kind of Computer (April 2025)
       ___________________________________________________________________
        
       A New Kind of Computer (April 2025)
        
       Author : gkolli
       Score  : 27 points
       Date   : 2025-06-24 03:27 UTC (3 days ago)
        
 (HTM) web link (lightmatter.co)
 (TXT) w3m dump (lightmatter.co)
        
       | croemer wrote:
       | I stopped reading after "Soon, you will not be able to afford
       | your computer. Consumer GPUs are already prohibitively
       | expensive."
        
         | kevin_thibedeau wrote:
         | This is always a hilarious take. If you inflation adjust a 386
         | PC from the early 90s when 486's were on the market you'd find
         | they range in excess of $3000 and the 486s are in the $5000
         | zone. Computers are incredibly cheap now. What isn't cheap is
         | the bleeding edge. A place fewer and fewer people have to be
         | at, which leads to lower demand and higher prices to
         | compensate.
        
           | ge96 wrote:
           | It is crazy you can buy a used laptop for $15 and do
           | something meaningful with like writing code (meaningful as in
           | make money)
           | 
           | I used to have this weird obsession of doing this, buying old
           | chromebooks putting linux on them, with 4GB of RAM it was
           | still useful but I realize nowadays for "ideal" computing it
           | seems 16GB is a min for RAM
        
             | ge96 wrote:
             | It's like the black Mac from 2007, I know its tech is
             | outdated but I want it
        
         | TedDallas wrote:
         | It was kind or that way in early days of high end personal
         | computing. I remember seeing an ad in the early 90s for a 486
         | laptop that was $6,000. Historically prices have always gone
         | down. You just have to wait. SoTA is always going to go for a
         | premium.
        
         | ghusto wrote:
         | That irked me too. "_Bleeding edge" consumer GPUs are ...",
         | sure, but you wait 6 months and you have it at a fraction of
         | the cost.
         | 
         | It's like saying "cars are already prohibitively expensive"
         | whilst looking a Ferraris.
        
         | Animats wrote:
         | That's related more to NVidia's discovery that they could get
         | away with huge margins, and the China GPU projects for graphics
         | being years behind.[1]
         | 
         | [1] https://www.msn.com/en-in/money/news/china-s-first-gaming-
         | gp...
        
       | Anduia wrote:
       | > Critically, this processor achieves accuracies approaching
       | those of conventional 32-bit floating-point digital systems "out-
       | of-the-box," without relying on advanced methods such as fine-
       | tuning or quantization-aware training.
       | 
       | Hmm... what? So it is not accurate?
        
         | btilly wrote:
         | It's an analog system. Which means that accuracy is naturally
         | limited.
         | 
         | However a single analog math operation requires the same energy
         | as a single bit flip in a digital computer. And it takes a lot
         | of bit flips to do a single floating point operation. So a
         | digital calculation can be approximated with far less energy
         | and hardware. And neural nets don't need digital precision to
         | produce useful results.
        
           | B1FF_PSUVM wrote:
           | > neural nets don't need digital precision to produce useful
           | results.
           | 
           | The point - as shown by the original implementation...
        
         | bee_rider wrote:
         | It seems weirdly backwards. They don't do techniques like
         | quantization aware tuning to increase the accuracy of the
         | coprocessor, right? (I mean that's nonsense). They use those
         | techniques, to allow them to use less accurate coprocessors, I
         | thought.
         | 
         | I think they are just saying the coprocessor is pretty
         | accurate, so they don't need to use these advanced techniques.
        
       | btilly wrote:
       | This paradigm for computing was already covered three years ago
       | by Veratasium in https://www.youtube.com/watch?v=GVsUOuSjvcg.
       | 
       | Maybe not the specific photonic system that they are describing.
       | Which I'm sure has some significant improvements over what
       | existed then. But the idea of using analog approximations of
       | existing neural net AI models, to allow us to run AI models far
       | more cheaply, with far less energy.
       | 
       | Whether or not this system is the one that wins out, I'm very
       | sure that AI run on an analog system will have a very important
       | role to play in the future. It will allow technologies like
       | guiding autonomous robots with AI models running on hardware
       | inside of the robot.
        
       | boznz wrote:
       | Weirdly complex to read yet light on key technical details. My
       | TLDR (as an old clueless electronics engineer) was the compute
       | part is photonic/analog, lasers and waveguides, yet we still
       | require 50 billion transistors performing the (I guess non-
       | compute) parts such as ADC, I/O, memory etc. The bottom line is
       | 65 TOPS for <80W - The processing (optical) part consuming 1.65W
       | and the 'helper electronics' consuming the rest so scaling the
       | (optical) processing should not have the thermal bottlenecks of a
       | solely transistor based processor. Also parallelism of the
       | optical part though using different wavelengths of light as
       | threads may be possible. Nothing about problems, costs, or can
       | the helper electronics eventually use photonics.
       | 
       | I remember a TV Program in the UK from the 70's (tomorrows world
       | I think) that talked about this so I am guessing silicon was just
       | more cost effective until now. Still taking it at face value I
       | would say it is quite an exciting technology.
        
       | quantadev wrote:
       | In 25 years we'll have #GlassModels. A "chip", which is a passive
       | device (just a complex lens) made only of glass or graphene,
       | which can do an "AI Inference" simply by shining the "input
       | tokens" thru it. (i.e. arrays of photons). In other words, the
       | "numeric value" at one MLP "neuron input" will be the amplitude
       | of the light (number of simultaneous photons).
       | 
       | All addition, multiplication, and tanh functions will be done by
       | photon superposition/interference effects, and it will consume
       | zero power (since it's only a complex "lens").
       | 
       | It will probably do parallel computations where each photon
       | frequency range will not interfere with other ranges, allowing
       | multiple "inferences" to be "Shining Thru" simultaneously.
       | 
       | This design will completely solve the energy crisis and each
       | inference will take the same time as it takes light to travel a
       | centimeter. i.e. essentially instantaneous.
        
       | Animats wrote:
       | Interesting. Questions, the Nature paper being expensively
       | paywalled:
       | 
       | - Is the analog computation actually done with light? What's the
       | actual compute element like? Do they have an analog photonic
       | multiplier? Those exist, and have been scaling up for a while.[1]
       | The announcement isn't clear on how much compute is photonic.
       | There are still a lot of digital components involved. Is it worth
       | it to go D/A, generate light, do some photonic operations, go
       | A/D, and put the bits back into memory? That's been the classic
       | problem with photonic computing. Memory is really hard, and
       | without memory, pretty soon you have to go back to a domain where
       | you can store results. Pure photonic systems do exist, such as
       | fiber optic cable amplifiers, but they are memoryless.
       | 
       | - If all this works, is loss of repeatability going to be a
       | problem?
       | 
       | [1] https://ieeexplore.ieee.org/document/10484797
        
       ___________________________________________________________________
       (page generated 2025-06-27 23:00 UTC)