[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1750900679805.jpg (872 KB, 1080x1918)
872 KB
872 KB JPG
Aporogise
>>
>>107077132
>Researchers say
>their [...] chip MAY be capable of speeds 1,000 (sic) faster
>>
>>107077132
>Analog
>For a virtual brain
Makes sense.
>>
>>107077132
the west has fallen again
>>
File: 1699050763594858.gif (8 KB, 200x260)
8 KB
8 KB GIF
>analog cpu
nigga wut
>>
>>107077132
>RAM chip is faster than GPUs
What do they mean by this? Are they comparing to the VRAM on the cards? Because the VRAM speed isn't something they're really trying to innovate on.
>>
>owen hughes
INB4 retards start blaming China for the spam, all this "analog CPU invented by China" are clickbaits using an normal paper and most of the time not even linking it, because the fact is irrelevant, the goal is the retards clicking the news

https://www.nature.com/articles/s41377-025-01753-w (it's open access)
>>
>>107077132
>>107077169
/g/ is so fucking stupid...
for the nth time, go read about analog "AI" chips, retards. these are not comparable to GPUs, because these can only be used for inference, and are a lot more inflexible, which means they are used and programmed for specific uses.
stop falling for these popsci articles, retards.
>>
>>107077203
> inference
That's actually nice.
>>
File: 1762043710001.jpg (237 KB, 960x978)
237 KB
237 KB JPG
>>107077279
ITS NOT NICE
STOP MAKING ORDINARY PEOPLE HAVE AFFORDABLE INFERENCE MACHINE
IT'S... UHHHH DANGEROUS FOR OUR DEMOCRACY!
IT'S ALSO LE BAD FOR THE ECONOMY OKAY
THINK OF THE CUDA PROFIT MARGIN
OH VEYYYYY THIS IS ANNUDA SHOAH
>>
>>107077132
>>107077201
that's not the paper. the article links to https://www.nature.com/articles/s41928-025-01477-0

>>107077203
(samefag)
though have a disclaimer from me: I'm no expert on this shit (or anything close, really), and I hadn't even read the article or the paper (https://www.nature.com/articles/s41928-025-01477-0) before posting >>107077203 .there could be some advances in this research but I have no clue what they would be. the article talks about RRAM, so I guess that's the real "novelty" here.

I've seen shit like this published multiple times, and apparently the tech is nothing new, like, at all. it's what chips like google Coral do.

>>107077279
apparently the tech is not meant to use with LLMs.
>>
>>107077361
>correct link
Thanks, my bad.
>>
>>107077203
>these can only be used for inference
So? What's wrong with specialized tools?
>>
File: 1755222201773533.jpg (863 KB, 1179x1242)
863 KB
863 KB JPG
>>107077353
YOU CAN'T MOCK THE CHINESE CENTURY BRO, THIS TIME IT'S GONNA HAPPEN
>>
>>107077405
nothing, but like I said before, these are (AFAIU) not meant for LLMs.
DYOR, though. share if you think I'm wrong here. again, I'm no expert, I've learned from things I've read and videos I've watched on yt
>>
>>107077132
The journo class should apologize via seppuku
>>
>>107077416
GDP and all that metrics are created by the Anglosphere to make them look good.
It is now past its expiry date.

The inability and/or unwillingness of Westoids to create/adapt new metrics shows that it is indeed doomed.
>>
>>107077447
>It’s nawt meant fo LLMs
So? GPUs aren’t technically meant for LLMs either.
>>
File: 1762045401196.jpg (133 KB, 507x425)
133 KB
133 KB JPG
does this mean we're going to get cheap gpu or not??
>>
>>107077174
memory is the primary bottleneck for all processors, while logic scales quiet well, memory does not, the fastest ram in 1998 responds about as fast as the fastest memory today. Its why cpus and gpus are getting progressively larger caches over time, inorder to mitigate
>>
>>107077554
Idk man, DYOR.
reading the paper, I get the idea that these basically calculate an approximation for x in matrix equations of type Ax = b. their approach seems to be to mix 2 analog approaches: a low precision and a high precision one.
you can see throughput and efficiency compared to GPUs in pic related (c and d). this.. doesn't look much better than GPUs, but I didn't read the whole thing yet, and I'm not knowledgeable enough to tell. plus at least they get similar performance, so that's something, I suppose.
>>
File: 41928_2025_1477_Fig5_HTML.png (412 KB, 1573x1206)
412 KB
412 KB PNG
>>107077666
>you can see throughput and efficiency compared to GPUs in pic related (c and d). this
>>
>>107077361
>using light for chips

Yeah the chinks stole this from Europe.
There have been Dutch teams working on this for years.
And suddenly it pops up in China...
We were so dumb to let them come to our universities, infiltrate and exfiltrate information
>>
>>107077416
>GDP is when LEDs at night
based retard
>>
>>107077169
Vacuum tubes
>>
>>107077132
https://youtu.be/GVsUOuSjvcg?t=898
>>
>>107077556
not
>>
>>107077416
>gdp is measured by the amount of lights a city has
lol face culture
>>
>>107077132
so it's a mining asic but for..?
>>
>>107079299
Model inference.
>>
>>107077169
>>107077132
I'm actually working on one such chip, amazing tech, super low power consumption, super fast, but it's purpose is super limited. The closest thing I can describe it as is a GPU core. It can do matrix multiplication very fast at low power, that's about it.
>>
>>107077132
>researcher says
>may be capable
Should i be implessed?
>>
>>107077132
chinese centuries begin now
>>
>>107077757
jokes on them we've realized it's a retarded time and money sink because there are still enormous unsolved problems (coupling, sources) that make it necessary to have a whole lab to operate the chip
good luck to chinks lmao
>>
>>107077132
>Resistive RAM
Imagine the power drain
>>
File: file.png (919 KB, 800x600)
919 KB
919 KB PNG
>>107077416
because GDP is a retarded metric
>>
>>107080037
It uses less power though.
>>
>>107077169
I think Kurzweil had it in his book on the future.
>>
File: IMG_2564.jpg (74 KB, 700x693)
74 KB
74 KB JPG
>>107077145
Just trust us, white men
>>
>>107080071
You just posted this because Elon Musk brought it up just now on Joe Rogan. Nice original thinking.
>>
>>107079377
Watch out, Sarah Connor is coming for you
>>
File: maybe-liam-gallagher.gif (93 KB, 640x480)
93 KB
93 KB GIF
>>107077132
>>
>>107077132
>>
Omg cant they say 1000 times faster than. for crying out loud wtf.
>>
>>107081912
Oh they did… where are my glasses
>>
>>107081920
No wait of speeds 1000 faster yep
>>
>>107081849
Not him but I distinctly remember telling that joke to someone in 2019. Somebody should tell Elon to stop stealing shit from me.
>>
>china makes claim
Yaaaaawn
>>
>>107077132
>Aurora Borealis in Xi's Kitchen
May I see it?
>No.
>>
>>107077132
Totally real and I believe it. Now to sit back and watch based Chinese lunge ahead a 1000 times faster than the shit decrepit literally who with the retard orange. Thank you Mr. China very cool.
>>
>>107079377
How many beers did it take you?
>>
>>107079377
Isn’t that what optical CPUs are good for?
>>
>>107080071
The thing poctured, is actually called "circular money" and it's almost universally a fraud.
GDP is indeed a meme, but for different reasons than that.
>>
>>107081959
read the thread, retard
>>
>>107077203
>because these can only be used for inference
these first gen ones. they'll eventually become the basis for a synthetic human brain. they're pretty close in theory of operation than LLMs running on GPUs are (they're not even close)
not talking about the chink tech, talking about neuromorphic hardware in general, analog spiking neural networks. the chinese didn't invent SSN hardware lmao
>>
>>107082066
>may be
If I wanted to waste my time I'd just goon. Glad to see you're mad as fuck. Keep it up.
>>
>>107080570
>resistive
>causes voltage drops
>where power go
>>
>>107082103
they're most likely memristors
>A memristor (/ˈmɛmrJstər/; a portmanteau of memory resistor) is a non-linear two-terminal electrical component relating electric charge and magnetic flux linkage. It was described and named in 1971 by Leon Chua, completing a theoretical quartet of fundamental electrical components which also comprises the resistor, capacitor and inductor.[1]
huh, look at that, dude was chinese american lol
>>
>>107081946
this
>>
File: may-i-723052600.gif (830 KB, 498x377)
830 KB
830 KB GIF
>>107077132
>4chan says that
>a unknown "news" site says that
>china says that
>researchers says that
>a thing has happened!

Ok, I'm sure that all checks out and is 100% truthful and accurate and not a wild game of telephone played by people who have ulterior motives........ But uhhh .... may we see it?
>>
File: 555.jpg (60 KB, 646x464)
60 KB
60 KB JPG
>>107082212
>>
>>107077169
new word for asics.
>>
>>107082175
Memeristors don't exist. They can be emulated by a much larger active circuit that has no practical applications, but they don't exist as individual rlements
>>
>>107077353
It's not nice to not be able to do any training.
For LLMs that's irrelvant, because you can't really train on anything consumer grade anyway, but it is a drawback for image generation, and other modalities like audio. It means you're locked in to whatever those with training capable hardware deign to give you.
I would still like a good value inference only chip though. A lot of this is all artifical market games anyway. Intel's Arc GPU roadmap shows that. They could give us more VRAM for cheap if they really wanted, but why would they?
>>
>>107079605
>enormous unsolved problems (coupling, sources) that make it necessary to have a whole lab to operate the chip
pls elaborate on this
>>
>>107082389
integrated photonics is using tiny waveguides, basically small fibers on a chip
to get light inside these you either
>make a light source (laser) on the chip
which is usually only low-power crappy laser light at specific wavelengths
>make light externally and couple it into the waveguides
which is very difficult to align (has to be done manually, machines can't do it) and has a lot of light loss, so you need a whole lab around ot make it manageable
And then you need to get light out, which is also another problem
>>
>>107081971
I work on the compiler not the hardware. The hardware part seems damn near magical to me.
>>
>>107081973
IDK what they are, how they are made or how they work, I just work on the software part.
>>
To the anons that might be interested, lots of countries are working on this, not just China, there are also US companies, Indian companies and many more working on building this tech.

Required watch: https://www.youtube.com/watch?v=lNypq1XuZRo
https://www.youtube.com/watch?v=LMuqWQcuy_0
>>
>>107082527
asianometry one looks interesting.
also gonna leave this here for people who want a primer on analog computing:
https://youtu.be/GVsUOuSjvcg
>>
>>107082568
Veritasium? Love to see it, thanks senpai.
>>
>>107077132
China checkmated the USA years ago. Americans just haven't realized it yet how fucked they are.
>>
>>107081849
This meme is older than you.
>>
>>107077169
you know how voltage in a chip determines if transistor lets current flow and therefore its 0 or a 1?
yea, now instead of it being 2 states it can be infinitely multiple states where the value of voltage determines the value of calculation
like an amplifier

I highly suggest you read up on ternary computers, quantum computers and thermodynamics computing
>>
>>107082029
>The thing poctured, is actually called "circular money" and it's almost universally a fraud.
I'm glad you added "almost" because it's currently responsible for almost the entirety of year-over-year US GDP growth.
>>
>>107077416
They probably have a higher GDP per capita because they don't waste it all on fucking RGB.
>>
>>107083044
>having the electric grid to cool this level of compute commercially
Not on this planet.
>>
Man, the wumaos going all out on this right now.
>>
>analog chips for low-power inference
cool, here's a company in Texas making these 3 years ago https://www.youtube.com/watch?v=GVsUOuSjvcg
actual hardware, not just a paper
>>
>>107083132
And they weren't adopted because they couldn't figure out how to make analog components less susceptible to noise and combining them with digital parts significantly diminished their benefits.
>here's a company in Texas making these 3 years ago
Both US and USSR were making these in the 70s
>>
>>107083407
Those are the kind of things you discover when you start building actual hardware instead of writing a paper. Especially the noise looks much less threatening until you actually measure the real thing.
>>
>>107083460
Well, they didn't really need to discover anything as everything I've just mentioned was already well-understood over half a century ago.
>>
>>107077203
so it's an ASIC
>>
>>107077132
Utter bollocks.
>>
>>107083107
you can't waste gdp retard. wasting is THE way to get gdp.
>>
>>107084279
Sure, okay, tardlard. I'm sure that makes perfect sense in that snot between your ears.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.