
Finally an NVIDIA Competitor for AI? Maybe, RAIN's Brain chips!!!
What this covers
This is a single-presenter walkthrough from the 1littlecoder channel about Rain Neuromorphic, a California startup that has raised roughly $30 million, with Sam Altman and Daniel Gross among its investors. The presenter starts from the market. Most deep learning work, from training to fine-tuning to inference, runs on Nvidia GPUs, largely because of the proprietary CUDA) software layer, and there is no strong equivalent competitor. OpenAI buys most of its compute through Microsoft, which ties it to Nvidia. A reported letter of intent from Altman to buy Rain's chips is read as a sign that OpenAI wants alternatives. The presenter also mentions Nvidia's roughly 800% stock gain in a year as an incentive for investors.
The middle of the video covers the technology. Neuromorphic computing uses analog systems instead of binary digital switching. The presenter's argument is that the brain works with continuous processes, which is why artificial neural networks) use fractional weights. Rain's design layers memristors on silicon wafers, a technology Intel and HP have also explored. The presenter also mentions Rain's digital in-memory computing paradigm, which aims to cut data movement. Rain's pitch is to co-design hardware with fine-tuning algorithms so models can adapt after deployment, and eventually to fit ChatGPT-sized models on thumbnail-sized chips.
The video turns contentious on the question of whether the limit is algorithms or hardware. CTO Jack Kendall is cited as doubting that backpropagation will scale and suggesting it may only imitate the brains of simple insects. Another camp says the computing architecture itself is the bottleneck. Yann LeCun is cited as arguing that current architectures cannot reach AGI). The presenter raises doubts too. Rain shares little openly beyond one paper, and porting PyTorch or JAX code to custom hardware could slow adoption. The video also considers that Rain, by controlling the whole stack, might simply become another concentrated chip monopoly.
Rain Neuromorphic is developing neuromorphic computing chips using analog systems to create AI accelerators that match brain efficiency and versatility, positioning themselves as a potential alternative to Nvidia by controlling the entire AI hardware-software stack.
- Rain uses analog neuromorphic computing instead of digital systems to emulate brain computation with better power efficiency
- The company is co-designing hardware, algorithms, and software across the entire stack from fabrication to inference
- Rain's stated goal is brain-parity computation: achieving orders of magnitude cost reduction and ability to run ChatGPT-sized models on thumbnail-sized chips
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The human brain does not operate with discrete digital values but uses continuous analog processes, which is why artificial neural networks use floating-point weights (such as 0.23, 0.24) rather than binary values.
“but our brain doesn't work like that and in fact that is exactly why when you do artificial neural network you change the weights in a floating Point order so you do like. 23 then you do0 24 maybe like you have got a different learning rate maybe you have got a different algorithm but everything that you store inside an artificial neural network those are the weights and and those weights are floating points”
Sam Altman sent a letter of intent to buy AI chips as an alternative to Nvidia, and this company mentioned was Rain Neuromorphic, suggesting OpenAI is actively seeking chip alternatives.
“remember the company in which Sam Altman sent the letter of intent to buy AI chips that could be a competitor of Nvidia we're going to talk about this company in this video this company is called Rain neuromorphic Inc”
Another group of scientists believes that current computation architecture—not the algorithmic approach—is the fundamental bottleneck preventing complex neural computations similar to how the brain operates, necessitating a new computing architecture altogether.
“while that is one side of the thought there is another group of people who believe that the current architecture a current computation architecture is actually the bottleneck for us to run really really really complex computations just like how our brains do and that is why they believe that there is a need for a new Computing architecture itself”
Rain's CTO Jack Kendall discusses in a podcast how backpropagation would not scale and why Moore's Law may not apply, questioning the viability of current deep learning architectures.
“their CTO Jack Kendall actually talks about what is the current disadvantage like why back publication would not scale up mood law and all the other things like he discusses in this particular podcast on the channel like the AG show”
Rain's CTO suggests that certain fundamental concepts of deep learning (backpropagation) may actually be emulating what happens in primitive brains of small insects, positioning neuromorphic computing as a return to biological plausibility.
“he also says that there are certain fundamental concepts of deep learning that might be actually what is happening in the Primitive brain like small insects and all these things”
Most current deep learning work, including model training, fine-tuning, and inference, happens on Nvidia GPUs rather than alternatives like Apple Metal, AMD processors, or CPUs, thanks to Nvidia's proprietary CUDA kernel and software.
“if you see most of the current deep Learning Works happen on Nvidia gpus I mean even if you have certain options like metal on Apple silicon or um AMD processors or even on CPU but most of the model training most of the model fine-tuning most of the model inference actually happens on Nvidia GPU thanks to their proprietary kernel or software called cuda”
OpenAI purchases most compute from Microsoft (which is partnered with Nvidia), meaning OpenAI's current AI infrastructure depends on Nvidia GPUs through the Microsoft partnership.
“opena already buys a lot of compute from Microsoft that's where like their partnership has been open AI would definitely either buy compute from Nvidia or because Microsoft is partnered with Nvidia they would get it”
The speaker expresses uncertainty about the openness of Rain's architecture, noting that while they found one paper on the website, there is little openly-shared information about how Rain's architecture works, likely because the semiconductor industry emphasizes patents over openness unlike the software industry.
“I'm not sure how much of their architecture would be open because I could not see a single paper from their website like um not even like okay there is there is one paper here uh but I don't see like a lot of information openly being shared about how what is their architecture and all these things but I guess like the semiconductor industry has been one of the one of the industries where patents matter a lot”
Rain's chip design is built by adding a circuit called memristor on top of silicon wafers, with memristor being a relatively new technology that companies like Intel and HP have previously explored for neuromorphic computing and artificial synapses.
“this rain chip is designed by adding a circuit called memor on top of silicon Wafers memer is something new it there there are like Intel that has tried like this kind of neuromorphic computing before HP is another example that has got like artificial synapsis”
Neuromorphic computing emulates brain function using analog systems rather than traditional digital systems, which operate on binary zeros and ones with voltage switching.
“neuromorphic Computing is a way to emulate brain not by using digital systems by using analog systems so if you have studied computer science digital systems operate with zeros and ones you switch it on it is one you switch it off it is zero how do you switch it on you give either plus 5 Volt or something like that or you do not give it zero that is switch it off this is how digital systems work”
There are currently no strong or powerful equivalent competitors to Nvidia in AI accelerator hardware.
“we do not have a very strong a powerful and equivalent competitor for NVIDIA this is the state of the art today”
The speaker notes that Nvidia's stock has appreciated by 800% in approximately one year, representing absurdly high returns that create an investor incentive to replicate such success.
“I think in the last one year the stocks have appreciated by 800% or something like that absurd numbers”
Yann LeCun argues that with current neural network architecture it is impossible to achieve artificial general intelligence (AGI) or superintelligence better than humans.
“there are a lot of scientists like Yan liun who still says that with the current architecture it is impossible for us to achieve AGI which stands for artificial general intelligence if your ultimatum or the Nar is to achieve a or achieve intelligence or super intelligence better than human beings then the current neural network architecture will not give you that”
Rain's approach resembles Nvidia's strategy of creating an accelerator chip with specialized memory architecture (like graphics memory) but using completely different architecture and a CUDA-equivalent version to run seamlessly across their hardware resources.
“this is going to be like completely revolutionary in what what it does and in fact you can see that it almost feels like they're almost trying to create a chip just like Nvidia has done also an accelerator chip that has like you know the graphics memory but totally different architecture almost feels also like they also have like one Cuda version that they have got so that it runs seamlessly across their Hardware resources”
AI models often fail upon deployment due to inevitable mismatch between training and deployment environments, and Rain addresses this through co-designing fine-tuning algorithms with hardware to enable efficient real-time training and adaptation.
“AI models often fail uh upon deployment due to inevitable mismatch in training and deployment environments that is exactly what I spoke about fine tuning solves this problem but requires devices to support high performance training rain AI is cod designing fine-tuning algorithms with Hardware to facilitate efficient realtime training”
The speaker notes the possibility that Rain could become another Nvidia, building a successful but highly concentrated chip monopoly, and speculates that many Silicon Valley investors would not necessarily mind this outcome given Nvidia's exceptional stock performance (800% appreciation in one year).
“the only catch here might be that we are actually you know building another Nvidia which I don't think uh uh a lot of Silicon Valley investors would mind because Nvidia is a Cash cover and you can see that I think in the last one year the stocks have appreciated by 800% or something like that absurd numbers”
Rain does not claim to create merely an alternative to Nvidia GPUs but instead pushes further by attempting to mimic the human brain through neuromorphic computing.
“the company doesn't even claim that they want to create an alternative to Nvidia GPU just not only just says that but they are actually pushing froners further saying that they want to mimic human brain which is exactly what the entire artificial neural network deep learning and all these things have been and they are trying to do it through something called neuromorphic Computing”
The speaker concludes that Rain's neuromorphic approach represents an attempt to use analog systems that are not as power-hungry as existing GPU-based systems and do not have current neural network algorithmic bottlenecks, all achieved through full-stack control of the hardware-software ecosystem.
“at the end of the day what Raina is trying to do is trying to use neuromorphic Computing to have analog systems that can do Computing for you that are not um as bad as the existing system not as power hungry as the existing systems even as you go for like Nvidia gpus that are that do not have a lot of bottlenecks like the current neural network algorithms have and they want to overcome all these things by having full control over the existing stack”
Rain's strategy is to control the entire AI ecosystem by maintaining semiconductor-level access, owning proprietary algorithms, and controlling the complete training, inference, and fine-tuning process, rather than just selling chips.
“at the end of the day what they trying to do is they're trying to have a hardware like a chip that is trying to create like a brain maybe the first version would definitely be nothing like a brain I don't know but not just the chip but they want to control the ecosystem they want to have uh you know the semiconductor level access they want to have their own algorithm they want to have their own um training um inference and uh fine tuning process so they basically want to have all these things along with the chip”
Rain Neuromorphic Inc is a California-based company that has raised approximately $30 million in funding, with Sam Altman as one of the investors alongside others including Daniel Gross.
“rain. I is uh of course a California based company and they have raised about like $30 million I guess and um Sam Altman is of course one of the investors in that but there are lot of other names like Daniel gross and all the others are there”
Rain's extreme vision is to eventually place models the size of ChatGPT into chips the size of a thumbnail, representing radical miniaturization and efficiency gains.
“we will ultimately be able to put models of the size of chat GPT into the chips of size of thumbnail”
Rain's core claim is that while the human brain performs hard computations, it does so with much greater power efficiency than existing digital computers, and therefore does not require massive data centers to run models like ChatGPT.
“while our brain can do all these hard computation that is not possible for like existing computers to do that is like uh very hard but also brain is doing it a much lesser power efficiency why do we need like huge data centers to run models like ch GPD and all the other things and that is where rain. comes in”
Rain's CEO has mentioned in interviews published approximately three years ago (during the COVID-19 era) that the company was closely working with OpenAI, suggesting long-standing engagement between the companies.
“I I would strongly encourage you to watch three videos that I link it in the YouTube description where their CEO and CTO CEO here actually talks about um what rain AI or rain neuromorphic is actually doing I don't know if they've renamed the company or not but this video was published almost 3 years back during the time of Co U not a lot of chat GPD at this point but even in this video he actually specifically mentions that they were closely working with opena”
When Rain Neuromorphic obtained $25 million in funding, Gordon Wilson (CEO and co-founder) stated that while most AI chips on the market are digital, Rain's technology uses analog systems to decipher incremental information such as sound waves.
“when rain. AI actually got the $25 million funding they they were in the news now they are in the news for a different reason but back then they were in the news because they wanted to create like a brain chip a chip that is like a brain and Gordon Wilson the CEO and co-founder of rain said that while most AI chips on the market today are digital is company's Technologies analog so analog chips can decipher incremental information such as sound waves”
Rain is pioneering a digital in-memory computing (DC) paradigm designed to address inefficiencies in AI processing by minimizing data movement and data storage, which unlike traditional in-memory computing designs can scale to high production volumes and support both training and inference.
“rain. aai or rain AI is pioneering the digital inmemory Computing DC Paradigm to address the inefficiencies to refine AI processing data movement and data storage unlike the traditional inmemory Computing designs rain AI proprietary DC CES are scalable to high production and support training and inference”
A key technical challenge with Rain's approach is that porting existing algorithms written in PyTorch or JAX to their custom hardware would require significant porting effort rather than seamless execution, which could limit adoption.
“I'm not again sure how the existing algorithms like for example an algorithm written on pych or Jax would run on their Hardware because that would require some kind of porting”
Rain's CEO states that the company's benchmark for success is mimicking the human brain: making computations as efficient as the human brain and as versatile as the human brain.
“he actually says that they're NSTAR like every company has a benchmark to compare every company has a ner and their Benchmark to compare is actually brain they want to make uh their ships computation as efficient as brain like human brain and they want to make computations as um was and um versatile as human brain”
The speaker acknowledges that Rain's neuromorphic goal of mimicking the brain has been the stated aspiration of deep learning for a very long time, but expresses optimism that Rain is collaborating on hardware at all levels from fabrication to algorithms.
“I know like you know this might sound very discouraging because this is what has deep learning has been for a very long time but I'm very happy to see somebody actually collaborating on the hardware and every level of Hardware like starting from fabrication to chip design to uh you know like the risk architecture to uh the software kernel and even like at the algorithm level”
Rain is actively recruiting across the entire AI hardware stack including fabrication, architecture, semiconductors, RISC-V, and ML/AI architecture design roles, indicating comprehensive technical ambitions.
“the kind of roles that they're recruiting you can see that they're covering the entire uh system like fabrication architecture um semiconductors risk RC risk V and also you have got uh you know this ML and AI architecture designs all these things are there”
Rain's CEO has stated in interviews that the company has closely worked with OpenAI, though the exact nature and scope of their partnership agreement is unclear.
“I've seen a couple of interviews from their CEO they've said like they've closely worked with opena even couple of years back I'm not sure what is the kind of agreement that they have with opena but this seems like very interesting”
Rain's proprietary quantization algorithms allow their accelerator to maintain FP32 (floating-point 32) accuracy on their completely different hardware compared to floating-point 32 storage in traditional artificial neural networks, achieving no loss in accuracy.
“when combined this this with rain AI proprietary quantization algorithms the accelerator maintains fp32 accuracy which means with a totally different Hardware with their proprietary quantization algorithm you get to see no loss in accuracy when you compare it with floating Point 32 which is like one of the highest forms of storing data in your artificial neural network”
Artificial neural networks originated from the fundamental goal to create computing systems that emulate human brain function, which is why the field borrowed neuroscience terminology like neurons, hidden layers, and synapses.
“the field artificial neural network Ann started with this fundamental understanding that they want to create computations or Computing systems that can emulate human brain that is exactly why it was named artificial neural network”
Rain's job descriptions state that their products achieve an order of magnitude improvement over the status quo through co-designing every layer of the AI stack from circuits to algorithms.
“our products achieve that means they already have the product an order of magnitude improvement over the status quo by co-designing every layer of a AI stack from circuits to algorithms”
Rain's mission statement is to make AI radically cheaper, and the company claims to be world leaders in algorithm-hardware co-design for AI, with a roadmap beginning with products 100% better than GPUs that will ultimately deliver orders of magnitude cost efficiency improvements.
“our mission is to make eii radically cheaper we are the world leaders in algorithm hardare codesigned for artificial intelligence our road map begins in products 100% better than gpus will ultimately deliver products that are orders of magnitude more cost efficient cost effective than what is available today”
Daniel Gross is a prominent investor appearing in nearly every significant AI startup, indicating he represents a systematic venture strategy in the AI space.
“Daniel gross is definitely a very popular name in almost every single eii startup that you have got”
Deep learning has been in continuous rise for many years until recently, when interest has dramatically intensified due to Transformer architecture and large language models like ChatGPT and Stable Diffusion.
“deep learning has been on the rise for so many years until recently people have gone Bonkers and crazy because of Rise of Transformers architecture large language models things like CH GPT and Sable diffusion people are absolutely blown up by the things that they can create”