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.
Ah, this prompted a few questions from me. Perhaps someone with more hardware background could chime in.
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.
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?
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.
I guess, what are the tradeoffs in place such that so many chip designs seem to reuse square tilings both locally and globally?
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.
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.
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.
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.
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.
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
> 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.
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.
I think the point is they need several wafers to fit the model. Data parallelism between requires too much bandwidth (they only have so much edge real estate for io) so you have to do pipeline parallelism with say one layer per wafer. No one does it on gpu unless you can’t fit the model in one fast network domain (which is 72 GPUs nowadays for Nvidia). There is a Dwarkesh episode from like two months ago that explains this reasonably accurately.
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.
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".
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.
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?
Lol these guys have another stupid quantization scheme they had mentioned earlier. Most models are stupider at least from them gateways.
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.
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.
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".
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.
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.
I'm sure that by some definitions like this we've already discovered narrow super intelligence.
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.
More like lighting a fire in your codebase at the speed of light.
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).
One developer can barely keep up with the daily output of multiple llm sessions writing code at 50 tps.
Im convinced anyone confidently thinking 2k tps is a no brainer is probably not concerned about the quality of their work (many such cases).
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.
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.
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.
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.
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?
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.
Openrouter on nitro gets me there. Although the TPS is 5X less still.
I have tried GPT OSS 120b on cerebras and it flies. But it's mostly wrong.
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.
GPT-5.6 Sol Ultrafast is running on Cerebras hardware.
It is not available to "developers".
Sure it is, you just have to pay for it. TANSTAAFL.
Ah, this prompted a few questions from me. Perhaps someone with more hardware background could chime in.
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.
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?
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.
I guess, what are the tradeoffs in place such that so many chip designs seem to reuse square tilings both locally and globally?
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.
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.
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.
Do you need diagonal wires for a topological hexagonal tiling?
Wouldn't a rectangular brick layout be topologically a hexagonal tiling
A few years old now, but their co-founder whiteboards how their chips work https://www.youtube.com/watch?v=7GV_OdqzmIU
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.
Citing full precision benchmarks then just going with 4 bit quants
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.
What's your use case?
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
> 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.
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.
I think the point is they need several wafers to fit the model. Data parallelism between requires too much bandwidth (they only have so much edge real estate for io) so you have to do pipeline parallelism with say one layer per wafer. No one does it on gpu unless you can’t fit the model in one fast network domain (which is 72 GPUs nowadays for Nvidia). There is a Dwarkesh episode from like two months ago that explains this reasonably accurately.
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.
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".
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.
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.
Exactly, which is why it’s so incredibly fast. And expensive.
SRAM doesn't shrink as well as other types of circuitry do when process nodes advance.
traditional SRAM memory has 6 transistors per bit. Dense SRAM can have two transistors.
Wafers can have trillions of transistors, 10-16 times more than the 4 trillion of the WSE-3.
1 Terabyte of SRAM per wafer will be the upper limit for a few years [1].
[1] https://youtu.be/vbqKClBwFwI?si=gbBeZ1B66_eGI81Z
You mean it is unimaginably massive.
Lol these guys have another stupid quantization scheme they had mentioned earlier. Most models are stupider at least from them gateways.
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.
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
If you made the mistake of buying their stock the past few weeks, you might wish you hadn't (or maybe it's an opportunity to buy low yada yada)
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.
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.
I'm sure that by some definitions like this we've already discovered narrow super intelligence.
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.
Or you'd just be making the same mistakes experts do but at a 100x clip.
Tortious really does win the race, and there is no such thing as a free lunch.
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.
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).
One developer can barely keep up with the daily output of multiple llm sessions writing code at 50 tps.
Im convinced anyone confidently thinking 2k tps is a no brainer is probably not concerned about the quality of their work (many such cases).
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.
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.
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.
Cerebras is the Transmeta of the AI wave.
Not sure why people would downvote this. Your not allowed to criticize any AI company on this site it seems.
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.
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?