From my testing, the problem Cerebras has is they need to give people a taste to get new customers in, but since it requires so much hardware to get that speed and they don't have OpenAI-type money (< $3 billion compared to over $100 billion) and it's just incredibly hard regardless to manufacture that many of these types of crazy giant wafer nodes, there is just no way for them to keep up with the demand.<p>So what they do is they have these cheap plans that let anyone get a little taste of the incredible speed, but they aren't reliable at all, because they have to share limited resources dedicated to sales. I believe that is really the main purpose of like the $50/month or whatever plans -- part of the sales funnel. The cheap plans compete directly with resources they need for their "real" paying enterprise customers.<p>The "real" plans start at like $1500 per month, and I assume those have reasonable service level reliability.<p>But actually I believe those $1500/month plans are totally worth it for some applications.
Lots of folks here say Cerebras serves outdated models.<p>Here’s my view: lets assume the above is true and the average serve with a delay of 6-9 months. Given that, its fair to expect them to start serving DS flash v4, glm 5.3, mimo 2.6 pro or kimi 3.1 pro in a matter of months. DS flash is already fast to the point I’ve became a bottleneck. But my point is: Cerebras should soon release better models that are not that far from frontier ones, at a fraction of cost and with multiples of the speed.<p>Just be patient and plan for the future.
Cerebras is good on paper but all I can reasonably access is "GPT OSS 120b, Qwen 3.8 27B" which aren't very good for my use.<p>Openrouter on nitro gets me there. Although the TPS is 5X less still.<p>I have tried GPT OSS 120b on cerebras and it flies. But it's mostly wrong.
> GPT OSS 120b<p>It's extremely frustrating that these fast models are very old at this point. Chat jimmy launched in february, and nothing has come of it, not even GA let alone improvements with intelligence.<p>AI is currently moving so fast, that might as well be a lifetime.<p>We get SOTA from anthropic/openai every few weeks on models that seem to focus on one useful thing (e.g. opus) and then 5 useless things like booking restuarants. We're ignoring really useful applications for some reason.<p>It took jev to shake things up. I remember seeing reactions that were basically "people want this? the big labs probably have some version of this lying around"... and it's like yea, maybe these labs need to fix their priorities a little bit.
I don’t think I really understand the motivation to move as fast as possible. Who cares if you’re behind by 6 months? Objectively you’re still ahead of where you would be without the 6 months old tech.
I suspect they’re ignoring it because there isn’t money in it.<p>Jev is super cool, and it’s super cool that it can handle novel categorization. Most organizations aren’t spending the big bucks on novel classification, though. They’re doing the same thing a billion times, which means Jev has to compete with having a custom model built and doing your own inference.<p>It’s very hard to build something generic like Jev, but it’s fairly manageable to build something more specific that handles a single kind of task. Especially at the scale where OpenAI would care about your attention.<p>I don’t really understand neural networks and I managed to train one to do hand gesture recognition on an old Leap Motion. It’s a pretty limited domain, but it trains on the CPU of my decade old laptop in under 3 minutes. The accuracy is quite good too, especially when you discount errors from the Leap (like misidentifying which finger is extended).
When talking about Jev, people love to bring up the whole "bitter lesson" idea that generalization will win out. Maybe it will in the end, but I tried most of the recent Jev-likes on my own not-very-complex use case and they weren't competitive with domain-specific classification at all.
General classifiers are actually incredible for two things:<p>1. inference on arbitrary context / world knowledge with fast, structured output<p>2. I almost don't want to say it because I don't know if people realize this yet... but that makes traces from any kind of non-AI process or system "learnable", you can easily model it by by just asking it to predict the next outcome during learning and compare against what happened.<p>which also means it could also learn new strategies through rollouts in a way that generalizes across semantic space which means...<p>so these models are actually quite promisingly useful for existing companies because they <i>do</i> have general traces of what happened in their company and want to discover ways they can tweak their business processes, and then there is durable demand for models you can embed within a product or UX flow without exposing your user to slop + then train or customize and build real IP or custom capabilities<p>general classification itself is quite interesting IMO for applications that need to make judgment calls or select one of many UX actions, and a much better fit for this than LLMs. So if you think about it, this is the type of thing businesses might be calling a lot - every trace (eg user session, package shipment, sales motion, marketing campaign, job outreach) + proactively throughout dynamic experiences to decide whether to eg end a call or which of multiple UIs to manipulate<p>this is just an unfamiliar product experience to most people because it hasn't really been bootstrappable with a strong general baseline until now. But now it's posible and I think it will be very col.
I think you’re misunderstanding me; I’m not saying places will use an LLM instead (whether cloud-based or on-premise).<p>I’m talking about them training their own models with a much more specific purpose.<p>You brought up choosing a menu option in a UI. Jev is really good at something like “find the menu option that best fits the users prompt”. That’s general, and requires a big complicated model.<p>Companies are usually more interested in something less variable. Ie the menu option they’re looking for is fixed. “Take an arbitrary website and find the Login button”. That’s a _way_ simpler model, and well within the abilities of many large companies to train.<p>You only need to pay Jev if there are so many variables the model basically has to be general.<p>Eg classifiers are a separate thing from LLMs and are literally only capable of structured output. It will output one of the X options it was trained to recognize.<p>I think some of this is that the Venn diagram of people who like AI and the people who like machine learning isn’t a circle. There’s a lot of “intelligences” out there that aren’t an LLM or decision model.
what might be useful to you (e.g. coding) might be completely useless to 99% of the people
Chatjimmy got taalas aquired
This - I use Cerebras supported models here and there, the speed is amazing and I can see the potential with having almost instant inference, but the models are so outdated at this point.<p>Cerebras if you're reading this, get your act together - maybe that will also help your stock price. Give consumers more options, not just to OpenAI.
GPT-5.6 Sol Ultrafast is running on Cerebras hardware.
That has never been made available even to limited customers as far as reporting goes. Currently available ultrafast on GPT-6 Astra and GPT-6.1 Sol are all using Nvidia for good reason [0].<p>The fact that Cerebras, which has a willing and major customer like OpenAI begging at their heels to deploy on their hardware, now posts about a 2.8T open weight model says a lot.<p>[0] <a href="https://blogs.nvidia.com/blog/gpus-openai-gpt-6-astra-ultrafast/" rel="nofollow">https://blogs.nvidia.com/blog/gpus-openai-gpt-6-astra-ultraf...</a>
It is not available to "developers".
I was a subscriber of their now-unavailable coding plan. The 1000 TPS was very unique at the time, but nowadays their qwen 3.8 gets roflstomped by Mercury 3 at both cost and performance at similar TPS speed. There's not much reason to use them unless you're an enterprise with a custom deal.
Qwen3.8 27B is a dense model with expensive K/V caching. It's <i>fantastic</i> if you have 32-48GB of VRAM and want to run 1 or 2 agent sessions locally. But if serving batch inference for many users, it's expensive.<p>DeepSeek V4 Flash 0731 is dirt cheap at scale, less than 10% of the cost of Qwen3.8 27B. This is because it's an MoE and it uses much less VRAM for KV cache. And it's generally a better model. V4.1 is also cheap at scale, though I've heard mixed reports on the optimized K/V caching.<p>You should not be running Qwen3.8 27B on an API normally. Maybe Cerberas's hardware is bad at batching or supporting more efficient K/V caching schemes, or something like that?
If you have 48GB of VRAM, you should be able to get Qwen3.8-Flash-Next running and from my use, it's basically the tipping point. 27B is great for a bit of work, but I have QFN running non stop (when I can gather the correct ideas for it to implement).
Having dealt with them professionally my view is they’re a company with truly remarkable technology and a bafflingly inadequate team. They need to hire better. Almost stopped me using their stuff wholesale. Groq are somehow worse.
Ah, this prompted a few questions from me. Perhaps someone with more hardware background could chime in.<p>They don't really show it here, but supposedly a Cerebras' chip is sub-divided into a 2D mesh of what they call PEs (cores). So data flow generally flows in two axis, and maintains consistent timings to its 4 neighboring cores. They show the layers flowing from left-to-right being streamed by PEs, but I wonder how true that ends up being: maybe the communications are easier to stream via diagonals, or some other shape that has better topology for data flow.<p>Then, even from die shots of CPUs and GPUs we often see that cores are tiled as squares, and those don't have the same cross-core connection and streaming requirement. Now, I assume that some parts of what we call a core can be moved around more freely while maintaining whatever power/heating/timing requirements we have, so perhaps it's easier to 'force' a square by moving those loose parts around. But why do we confine those to squares? Is super-core layout just something not worth investing time into, like how interconnects are timed, non-convex tilings, and distances to I/O?<p>And for streaming architectures like Cerebras, why aren't other types of tilings considered, something like a hexagon grid? This could have 3 inputs / 3 outputs, which could allow more complex data flow (?). Similarly, we could have 8 connections with longer diagonals in a square tiling, etc.<p>I guess, what are the tradeoffs in place such that so many chip designs seem to reuse square tilings both locally and globally?
The reason 2D grids work well is because you can split tensor axes along whatever dim you want across PEs (batch, sequence, hidden, etc.). And your per PE code is relatively simple. It's of the type : get data from neighbor, compute my subtensor, send to neighbor. And you can implement broadcasts and reductions along any axis.
A few years old now, but their co-founder whiteboards how their chips work <a href="https://www.youtube.com/watch?v=7GV_OdqzmIU" rel="nofollow">https://www.youtube.com/watch?v=7GV_OdqzmIU</a>
The EDA tools that are used to design chips do not support diagonal wires so that kind of rules out hexagonal grids. With only horizontal and vertical wires you end up with rectangular/square grids or rings.
Do you need diagonal wires for a topological hexagonal tiling?<p>Wouldn't a rectangular brick layout be topologically a hexagonal tiling
If that is so, would it be entirely to decrease search space for layout algorithms? I believe there isn't anything in the fabrication process that couldn't model diagonals.<p>I could see that square packing is more simple, but by shifting elements per row I don't see why it couldn't propagate at odd angles.
Oh the tools support it. The process doesn't.<p>Diagonal wires worked everywhere at 180nm and on a few fabs at 90nm if you rotated a whole metal layer (i.e. all metal-3 wires are diagonal to the metal-2 wires) and basically nowhere beyond 90nm.<p>OPC killed it. Too hard to OPC the diagonal stuff. And even if you rotate an entire metal layer, the contacts between a rotated and un-rotated layer would require enormous metal enclosures.
Citing full precision benchmarks then just going with 4 bit quants
> The total number of concurrent requests represents the number of simultaneous requests moving through the pipeline at any point in time. Unlike a GPU batch, those requests don’t wait for one another to finish, they stream through the pipeline continuously.<p>What do they mean by “unlike a GPU batch”? It’s not a hardware limitation of GPUs right? It’s just the scaffolding to do this in software isn’t available right
It's describing a different type of parallelism - pipeline parallelism. The network is laid out spatially on the wafer in CB's case. So weights are stationary and activations flow. A GPU is going to process the network one layer at a time. GPUs also use other types of parallelism (tensor and expert mainly in inference). Pipeline parallelism doesn't make much sense on a GPU for inference.
From reading xpct's comment in this thread it seems like they use effectively a very deep and wide pipeline by tiling the cores without central clocking or dispatch. In theory if a GPU is centrally organized it could simulate this behaviour, but I imagine that creates transfer overhead. Just speculating though.
59 wafers for kimi?!?! That sounds like some kind of architecture problem to me. I suppose if they're after a high tokens per second and this allows them to run a lot of agents at once and at high speed it might be ok.
This is amazing, and I'd love to see K3.1
> For ultrafast latency, only activations need to be passed between the wafers, and this can be done via RDMA over Ethernet.<p>I believe it for decode but the article says nothing about how they make prefill fast on that interconnect story :/.
Cerebras is the Transmeta of the AI wave.
I am invested, because I see this hardware as inherently.. neat. But did anyone do a comparison that is a little more.. human ( not purely benchmark based ). You know what I mean. There are things that make me pick chatgpt vs anthropic sub. Not all of it is absolute capability.
They've shown the calculation for number of concurrent requests limited by kv cache size. I guess for each new request loaded into the queue, the kv cache for that context will need to be loaded from a supporting storage server, that will need to have a lot of ram to store the active caches of multiple requests?
Less "storage server" and more "attention server".<p>A server that both holds KV cache in its own memory, and uses its own compute to perform attention operations. That way, you don't have to transfer cache entries server to server for every attention sweep - only the "requests" and "responses". Attention bandwidth stays server-local.
When will we see some speed improvements? Opus 5.5 is smart enough now, speed is what we need.<p>(though the agents also spend a lot of time in testing and building, which cannot be improved in speed a lot, I suppose)
I tried Cerebras this week for my project and their 27b model at 2000tks would be perfect for it. Unfortunately it doesn't do the 2k tokens per second which is what I am looking for.
We can push 1k tps on our speed lane with the right request parameters with deep seek v4.1 flash. Is 2k a hard requirement?
are you evaluating using OpenRouter.ai or other routers/gateways?<p>they have never been able to keep up with Groq/Cerebras-level inference speeds for some reason.
What's your use case?
Sorry, can't say just yet but nothing as dramatic as what was guessed below :)
I bet it is large-scale data crunching for labeling or categorization.
And the fact that you will never be benefited from the batching they’ve explained either, you’ll be paying for every token sequential or parallel unlike nvidia or amd gpus
what is a project that requires 2k tokens per second?
It's not launched yet and 2k isn't a hard requirement but faster inference of relatively simple prompts make a big difference for us. Cerebras also failed because it has very small context window.
Well, a weekend project that requires 345'600'000 tokens — you only get so many seconds in two days ;)
The issue is the models they offer are extremely outdated.
At 20 kW average consumption per WSE-3T wafer this gives roughly 1.18 MW for the whole system. Assuming 1500 tokens per second and 190 concurrent requests this gives 4.1 J/token. Which is good I suppose for the size of the model?
For a fairly weak comparison, modern Bitcoin mining chips are around 13J/TH, so around 0.32TH per token, or 320 billion double-sha256 mining operations per token served by this system. Peanuts are also about 4200 Joules, or almost exactly 1000 tokens per nut assuming you could efficiently turn it into electricity to run this chip. Just as a frame of reference, I don't think this actually serves to make any point.
Each Nvidia rack is 120 kilowatts, so the power consumption is about the same as 10 racks of GPUs, FWIW.
Ok I'm sold, where to buy?
:D
This conversation seems chock full of comments from ML experts. Im really impressed.
Extreme speeds are actually interesting if you think about it. We are apparently getting some interesting results from LLMs nowadays. If they would put 100x effort in 10% the time it takes, or imagine 1% where how long tool calls take become the bottleneck, that would actually probably feel like a really special kind of "super intelligence".
It would absolutely be a phase change in practical utility, but it still wouldn't be "super intelligence".
I think that just like there's tons of different kind of "superheroes" and nobody disputes whether they're genuinely deserving of the name or not, there could be many valid types of super intelligence.<p>If something is roughly as intelligent as a normal human expert in a given area, but 100 times faster, that would be a valid example of super intelligence.<p>I'm sure that by some definitions like this we've already discovered narrow super intelligence.<p>For example - breadth of knowledge. Is knowledge a kind of intelligence? If so then current agents absolutely have a breadth of knowledge which is superhuman.
Is computational ability a kind of intelligence? If so then 20 year old calculators absolutely have a computational ability which is superhuman.<p>I'm not sure what my (or yours) point is, but like with all of AI it's a moving target - I still remember how image recognition was deep in researchy Artificial Intelligence territory. Nowadays it's such an obvious building block, nobody thinks of image classifiers as "intelligence".
Or you'd just be making the same mistakes experts do but at a 100x clip.<p>Tortious really does win the race, and there is no such thing as a free lunch.
<a href="https://www.lesswrong.com/posts/semvkn56ZFcXBNc2d/superintelligence-5-forms-of-superintelligence" rel="nofollow">https://www.lesswrong.com/posts/semvkn56ZFcXBNc2d/superintel...</a><p>5 Forms of Superintelligence (2018, review of Bostrom's <i>Superintelligence</i> from 2014)<p>Featuring .... the SPEED SUPERINTELLIGENCE! It really has been considered that way, even in scifi prior to the current AI epoch.
It would be a super intelligence in terms of thinking speed, not necessarily be super intelligent, mind you.
More like lighting a fire in your codebase at the speed of light.<p>Nobody needs this, and the quality of models they're deploying, or all llms for that matter, require heavy human review (if you care at all about maintainability).<p>One developer can barely keep up with the daily output of multiple llm sessions writing code at 50 tps.<p>Im convinced anyone confidently thinking 2k tps is a no brainer is probably not concerned about the quality of their work (many such cases).
You’re thinking about this the wrong way. Yes, any individual developer doesn’t need this much power all the time. But any individual developer would like to get an answer quickly, which this does. You share the speed and capacity of this over a whole organization. Any one developer only uses a fraction of its capabilities at any given time. Getting an “instantaneous” answer is quite helpful. Think of it like a shared compiler; a developer is not constantly compiling, but having an “instantaneous” compiler is quite useful.
most of the time developers arn't looking for answers. They want to implement some feature. And I'd hazard most developers arn't using these things as a crafter implementer but a frankenstein of architect, implementer, tester, builder and deployers.<p>And the dev, if they're sane, wants to actually evaluate what it's doing in those arenas. Speed might be a psychology killer then. Actually being able to see the development process as it does it's good think and it's wrong think is a benefit for local development.<p>I don't think speed of tokens is all that magical if there's a huge imbalance of ability to "human in the loop" it's processs.
speed gets you to have flow though, rather than breaking your concentration with various waits that aren't actually times where you can stop and think because its 7s surrounded by tiring reading.<p>a milisecond response, that you can then spend 30s focused on is better.<p>its the same as a build tool. just slow enough build tools really suck, compared to either an instant build or a 30min build<p>evaluating the areas isnt the same as watching the llm harnes say "thinking" between submitting a request and getting a response
Code is an answer to a coding question or direction. Getting an answer quickly is always better than getting an answer slowly. You can evaluate the answer at your leisure. The human can take as long as the human wants. Meanwhile, the infrastructure is generating answers quickly for other people. Faster is always better, even with a human in the loop.
No, faster is not always better, there's a reason why we have cliches and literal nursery rhymes about it.<p>You're larping too, can we bring NFTs back so the the dunning kruger vibe coder commenters on the internet find something else to fixate on?
Occasionally it takes me at most 10 minutes to figure out that a solution the LLM took an hour to churn out is not up to par and I need to throw it away entirely.<p>There are diminishing returns, but half of the rate you mentioned would already help me keep focus.
At those token rates you could treat the LLM as a compiler and rebuild your software from scratch every week. That would mean the end for AI slop because people who spend the effort to build a process that creates high quality software from scratch can take full advantage of the one shot performance of an LLM without having to release the process itself.
Youre completely missing my argument. Speed is not the bottleneck, its the review burden.<p>You need human review when working on anything not table stakes. You cannot write shitty code for an avionics package on an airplane under the assumption that if we encounter bugs, its no big deal because we can rewrite it super fast..<p>It seems like all these bullish takes on AI are coming from people who think all software runs in a browser, and simply exists to take stripe payments or serve ads. Most software is actually very high stakes and lives depend on it and have high compliance burdens.<p>Just because it can "one shot" (which actually means 90 tools calls and a whole lot of brute forcing against tests/linter/compiler, and possibly reward hacking) to something that runs, does not mean it will continue to run well.<p>People treat llm made software like disposable code. You're the only one using it, typically on a single machine with no real world multi-user testing, but for some reason we derive from the fact it worked well the 3 times we used it, on a single machine with a single user that it is evidence of llm solving programming. Its bogus.
You can lessen the human review burden by pitting multiple agents against each other. Ask 3 agents to implement the same solution, then set them up against each other in differential tests. This can find bugs in them or clarify the spec. When tokens get cheap it is worth having multiple replicas, taking advantage of different programming languages, libraries and agents. You don't keep one version, you have a population, and it tests itself. It's easier to implement again than to imagine testing scenarios anyway.
This is why I think there will be selective pressure for fast tools. It’s like an evolutionary process and fast tools will be preferred and created. It’s already happening and also why I think computer use is a dead end.
Lol these guys have another stupid quantization scheme they had mentioned earlier. Most models are stupider at least from them gateways.<p>Such a weird company with a terminally online on linkedin ceo, garbage devrel on X. Cant really be friends with CSPs coz everyone has their own chip, no consumer solution either, no neocloud would want to onboard them either coz its baggage no one wants to risk dev time/customer money with.
A wafer looks pretty big. I'm surprised it has only 44GB of memory available. Doesn't that seem a tad .... low? I'm not a hardware person, but I would have assumed much more memory would fit into a wafer?
It's SRAM. It's the equivalent of 44 GB of CPU cache.
traditional SRAM memory has 6 transistors per bit. Dense SRAM can have two transistors.<p>Wafers can have trillions of transistors, 10-16 times more than the 4 trillion of the WSE-3.<p>1 Terabyte of SRAM per wafer will be the upper limit for a few years [1].<p>[1] <a href="https://youtu.be/vbqKClBwFwI?si=gbBeZ1B66_eGI81Z" rel="nofollow">https://youtu.be/vbqKClBwFwI?si=gbBeZ1B66_eGI81Z</a>
SRAM doesn't shrink as well as other types of circuitry do when process nodes advance.
SRAM is not very dense.<p>IMO the next big shift in AI hardware is when some of the upcoming logic-compatible memory types mature. Inference inherently likes many small local pools of memory, with their own interface each. This is impractical with currently available memory other than sram.
You mean it is unimaginably massive.
What's the cache input discount relative to the regular input?
asdf
could push kv-cache to off chip cheaper storage like dram. Inference engine manages decode for a single user to run to completion before decode a different users so kv-cache fetch from dram to sram is limited.
[dead]
[flagged]
[flagged]
[flagged]
[dead]
[flagged]
[flagged]
[flagged]
I'd want to know is what a 59-wafer pipeline costs in failure domain. The GPU deployment can lose a node and degrade; Because, here the model is the rack, so one bad wafer takes the whole thing down, and a pipeline that has to be live end to end doesn't get a rolling upgrade. How do you drain one?
[flagged]
[dead]