Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

phantafield.com

Post reply on HN
Print view View original

31–40 of 40 posts

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#31

Post by throwaway89201 »

minkowsky wrote: post #29
throwaway89201 wrote: post #28

The entire design looks very interesting, but from the outside and without domain expertise I find it very hard to assess if anything about this is actually real or just a large and well-executed product of an AI psychosis. The signs that this project is real are hard to verify: - Angel investment by FinFET inventor Chenming Hu [1] seems a big vote of confidence, but there is no independent confirmation of this anywh…

I also have another side project, thepantheon.ai. I think it's ok to have multiple talents. To make the world a much more interesting place. We talk to fabs. I am not allowed to expose any conversation. We don't need to prove to average Joe what we have. It's indeed teamwork to bring it into production, even with huge help from AI. It also takes expertise to make sure AI is correct. As a founder, it's more of a merit…

> I think it's ok to have multiple talents.

Of course that's okay, but do recognize that most blockchain projects in general and DAOs and NFTs specifically have been considered frauds or at the very least pipe dreams by many on this site from the beginning. And wider in tech society from the moment the hype was gone.

> We don't need to prove to average Joe what we have.

Of course you don't need to prove anyone anything, but it really can't hurt and there doesn't seem to be much effort in addressing the main issues. Not having anyone write about you also just seems like bad marketing: as you clearly are not in stealth mode, an extremely verbose public website stands oddly against no external coverage.

> It's indeed teamwork to bring it into production

So is or isn't the NanoGalaxy an actually physical, working device ready for a demo? (which isn't necessarily production, but somewhat close)

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#32

Post by minkowsky »

throwaway89201 wrote: post #31
minkowsky wrote: post #29

Earlier quoted context omitted.

I also have another side project, thepantheon.ai. I think it's ok to have multiple talents. To make the world a much more interesting place. We talk to fabs. I am not allowed to expose any conversation. We don't need to prove to average Joe what we have. It's indeed teamwork to bring it into production, even with huge help from AI. It also takes expertise to make sure AI is correct. As a founder, it's more of a merit…

> I think it's ok to have multiple talents. Of course that's okay, but do recognize that most blockchain projects in general and DAOs and NFTs specifically have been considered frauds or at the very least pipe dreams by many on this site from the beginning. And wider in tech society from the moment the hype was gone. > We don't need to prove to average Joe what we have. Of course you don't need to prove anyone anythi…

It's a valid prejudice. But smart contract is a beautiful technology.

I didn’t spend time on PR. Now I changed my mind. It’s a noisy world and good thing needs broadcasting.

Yes. We offer demo if you are a potential customer.

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#33

Post by binyu »

minkowsky wrote: post #26
binyu wrote: post #20

Earlier quoted context omitted.

> they scale out in 2D, we scale up in 3D. This actually helps a lot, thanks. > Instead of spreading SRAM across a wafer, we stack DRAM on top of the logic Is this done with current manufacturing technologies? Does it require a special process? > no streaming, no off-chip memory at all. ~1 kW, not 23 kW Is this for an individual compute unit? Compared to Cerebras, what's the ratio of power used vs compute output?

I think you are asking for the Energy/token. Cerebras is 12.8J, Sophon is 25.8mJ. Three orders of difference.

so Sophon is less efficient than Cerebras?

Edit: is that Joule vs micro-Joule? I need better glasses

> Cerebras is 12.8J, Sophon is 25.8mJ

Are your figures hypothetical or do you have a working prototype?

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#35

Post by robocat »

minkowsky wrote: post #22
vessenes wrote: post #12

Minkowsky, cool design! Question - the ASIC designers I've worked with over the years have been fairly adamant that integrating memory on package interspersed with logic is very difficult; the general statements run like "those designs always look great on paper, but never tape out properly". Have you done any hardware tests of this plan? Is this still considered quality advice? Second q, why start with 28nm? Is the…

Due to the thermal budget, most of the silicon design is constrained to a 2D layout. So the Memory is competing with logic for layout. Now we stack logic in the backend between metals. We fabricated 2T0C DRAM arrays with a 3D monolithic structure. That's a must-do. Why 28nm? Because it's cheap, widely available, and already gives us enough performance to beat Nvidia Vera Rubin. We have a road map, scaling it down. ht…

> 2T0C

2 transistor zero capacitors

See last paragraph (3.) of another comment for context: https://news.ycombinator.com/item?id=48713803

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#36

Post by robocat »

minkowsky wrote: post #6
codingpanic wrote: post #3

I've been wondering how long before RAM is fabbed on die to get around supply issues. This is one of the first I've read of so far. How long before Apple releases a CPU with ram on die?

Author here. The supply angle is exactly the motivation — HBM is the hardest part to get and ~26% of an AI rack's BOM. First, separate three things people lump together. Apple already does memory on package (M-series unified memory = LPDDR5X dies next to the SoC). The near-term industry path is bonded stacking (AMD 3D V-cache, HBM4's logic base die). What we're doing is monolithic — growing the memory on top of finis…

> TMDs

= Transition Metal Dichalcogenides

> BEOL

= Back End Of Line. The later stages of semiconductor manufacturing (after the standard CMOS logic transistors) e.g. adding the metal wiring and interconnect layers. Think end of a manufacturing line.

The core concept is to layer multiple non-standard non-silicon memory transistors above the metal layers.

That sounds like a stunning invention, since I think that alone implies better memory density than current SRAM (ignoring the extra complexity of stacking it above a logic layer).

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#37

Post by elisbce »

throwaway89201 wrote: post #28

The entire design looks very interesting, but from the outside and without domain expertise I find it very hard to assess if anything about this is actually real or just a large and well-executed product of an AI psychosis. The signs that this project is real are hard to verify: - Angel investment by FinFET inventor Chenming Hu [1] seems a big vote of confidence, but there is no independent confirmation of this anywh…

What a great bullshit detector signal it is if someone's involved in the DAOs and NFT stuff. Don't waste your time on a fake AI-generated one-man shell company. The company had like 3 investments since 2017, first $2.5k from school startup grant in 2017, and two more from incubator programs in 2023 and 2024, for "augmented reality". And now you are the most innovative chip designer in AI. Nowhere on Chenming Hu's Linkedin said anything about being an official advisor for this one-man company. I can take a photo with Messi, does that make me a top soccer player?

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#39

Post by mikewarot »

There are quite a few clever optimizations interplaying here. The use of bit-serial arithmetic, for instance, makes it possible to handle both FP16 and FP8 values without much overhead.

I wonder if the 2T0C DRAM could be used to directly feed a LUT, for FPGA applications. If so, that could greatly reduce their die size.

Re: Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM

#40

Post by minkowsky »

elisbce wrote: post #37
throwaway89201 wrote: post #28

The entire design looks very interesting, but from the outside and without domain expertise I find it very hard to assess if anything about this is actually real or just a large and well-executed product of an AI psychosis. The signs that this project is real are hard to verify: - Angel investment by FinFET inventor Chenming Hu [1] seems a big vote of confidence, but there is no independent confirmation of this anywh…

What a great bullshit detector signal it is if someone's involved in the DAOs and NFT stuff. Don't waste your time on a fake AI-generated one-man shell company. The company had like 3 investments since 2017, first $2.5k from school startup grant in 2017, and two more from incubator programs in 2023 and 2024, for "augmented reality". And now you are the most innovative chip designer in AI. Nowhere on Chenming Hu's Lin…

There is a thread on why HN is so biased against crypto. I don't feel the necessity to argue here. Prejudice is what prejudice is.

https://news.ycombinator.com/item?id=31302494

Also, PitchBook is hallucinating about every company. Don't trust their data.

Post reply on HN