> At over 2,500 t/s, Cerebras has set a world record for LLM inference speed on the 400B parameter Llama 4 Maverick model, the largest and most powerful in the Llama 4 family.
This is incorrect. The unreleased Llama 4 Behemoth is the largest and most powerful in the Llama 4 family.
As for the speed record, it seems important to keep it in context. That comparison is only for performance on 1 query, but it is well known that people run potentially hundreds of queries in parallel to get their money out of the hardware. If you aggregate the tokens per second across all simultaneous queries to get the total throughput for comparison, I wonder if it will still look so competitive in absolute performance.
Also, Cerebras is the company that not only was saying that their hardware was not useful for inference until some time last year, but even partnered with Qualcomm with the claim that Qualcomm’s accelerators had a 10x price performance improvement over their things:
Their hardware does inference with FP16, so they need ~20 of their CSE-3 chips to run this model. Each one costs ~$2 million, so that is $40 million. The DGX B200 that they used for their comparison costs ~$500,000:
You only need 1 DGX B200 to run Llama 4 Maverick. You could buy ~80 of them for the price it costs to buy enough Cerebras hardware to run Llama 4 Maverick.
Their latencies are impressive, but beyond a certain point, throughput is what counts and they don’t really talk about their throughput numbers. I suspect the cost to performance ratio is terrible for throughput numbers. It certainly is terrible for latency numbers. That is what they are not telling people.
Finally, I have trouble getting excited about Cerebras. SRAM scaling is dead, so short of figuring out how to 3D stack their wafer scale chips, during fabrication at TSMC, or designing round chips, they have a dead end product since it relies on using an entire wafer to be able to throw SRAM at problems. Nvidia, using DRAM, is far less reliant on SRAM and can use more silicon for compute, which is still shrinking.
>Each one costs ~$2 million, so that is $40 million.
Pricing for exotic hardware that is not manufactured at scale is quite meaningless. They are selling tokens over an API. The token pricing is competitive with other token APIs.
Last year, I took the time to read through public documents and estimated that their annual production was limited to ~300 wafers per year from TSMC. That is not Nvidia level scale, but it is scale.
There are many companies that sell tokens from an API and many more that need hardware to compute tokens. Cerebras posted a comparison of hardware options for these companies, so evaluating it as such is meaningful. It is perhaps less meaningful to the average person who cannot afford the barrier to entry to afford this hardware, but there are plenty of people curious what the options are for the companies that sell tokens through APIs, as those impact available capacity.
Some context is needed for this. The only way to get a 4 orders of magnitude difference would be to compare incomparable things, like OpenAI’s most expensive model versus llama 3.1 8B.
I agree on the first. On the second: I would bet a lot of money that they aren't actually breaking even on their API (or even close to). They don't have a "pay as you go" per-token tier, it's all geared up to demonstrate use of their API as a novelty. They're probably burning cash on every single token. But their valuation and hype has surely gone way up since they got onto LLMs.
> This is incorrect. The unreleased Llama 4 Behemoth is the largest and most powerful in the Llama 4 family.
Emphasis mine.
Behemoth may become the largest and most powerful llama model, but right now it's nothing but vaporware. Maverick is currently the largest and more powerful llama model today (and if I had to bet, my money would be on Meta discarding Llama4 Behemoth entirely it eventually without having released it, and moving on to the next version number).
> Also, Cerebras is the company that not only was saying that their hardware was not useful for inference until some time last year, but even partnered with Qualcomm with the claim that Qualcomm’s accelerators had a 10x price performance improvement over their things
I'm /way/ outside my expertise here, so possibly-silly question. My understanding (any of which can be wrong, please correct me!) is that (a) the memory used for LLMs is dominantly parameters, which are read-only during inference; (b) SRAM scaling may be dead, but NVM scaling doesn't seem to be; (c) NVM read bandwidth scales well locally, within an order of magnitude or two of SRAM bandwidth, for wide reads; (d) although NVM isn't currently on leading-edge processes, market forces are generally pushing NVM to smaller and smaller processes for the usual cost/density/performance reasons.
Assuming that cluster of assumptions is true, does that suggest that there's a time down the road where something like a chip-scale-integrated inference chip using NVM for parameter storage solves?
The processes used for logic chips, and the processes used for NVM are typically different. The only case I know of the industry combining them onto a single chip would be Texas Instruments’ MSP430 microcontrollers with FeRAM, but the quantities of FeRAM are incredibly small there and the process technology is ancient. It seems unlikely to me that the rest of the industry will combine the processes such that you can have both on a single wafer, but you would have better luck asking a chip designer.
That said, NVM often has a wear-out problem. This is a major disincentive for using it in place of SRAM, which is frequently written. Different types of NVM have different endurance limits, but if they did build such a chip, it is only a matter of time before it stops working.
> The only case I know of the industry combining them onto a single chip would be Texas Instruments’ MSP430 microcontrollers with FeRAM
Every microcontroller with on-chip NVM would count. Down to 45 nm, this is mostly Flash, with the exception of the MSP430's FeRAM. Below that... we have TI pushing Flash, ST pushing PCM, NXP pushing MRAM, and Infineon pushing (TSMC's) RRAM. All on processes in the 22 nm (planar) range, either today or in the near future.
> This is a major disincentive for using it in place of SRAM, which is frequently written.
But isn't parameter memory written once per model update, for silicon used for inferencing on a specific model? Even with daily writes the typical 10k - 1M allowable writes for most of the technologies above would last decades.
> I have trouble getting excited about Cerebras. SRAM scaling is dead, so short of figuring out how to 3D stack their wafer scale chips
AMD and TSMC are stacking SRAM on the chip scale. I imagine they could accomplish it at the wafer scale. It'll be neat if we can get hundreds of layers in time, like flash.
Yes, and it's TSMC enabling this. Lots of TSMC's customers going this route, not just AMD. Seemed odd to call out AMD as if they've got any special sauce here.
One caveat is that this paper only covers training, which can be done on a single CS-3 using external memory (swapping weights in and out of SRAM). There is no way that a single CS-3 will hit this record inference performance with external memory so this was likely done with 10-20 CS-3 chips and the full model in SRAM. Definitely can’t compare token/$ with that kind of setup vs a DGX.
Thanks for the correction. They are currently using FP16 for inference according to OpenRouter. I had thought that implied that they could not use FP8 given the pressure that they have to use as little memory as possible from being solely reliant on SRAM. I wonder why they opted to use FP16 instead of FP8.
By the time the CSE-5 is rolled out, it *needs* at least 500GB of SRAM to make it worthwhile. Multi-layer wafer stacking's the only path to advance this chip.
The memory bandwidth for that is 150GB/sec. Inference speed is memory bandwidth bound, so that memory is useless for inference. Discrete GPUs will run circles around the CSE-3 at inference if they tried using the external DRAM.
I think it is too risky to build a company around the premise that someone won't soon solve the quadratic scaling issue. Especially, when that company involves creating ASICs.
Attention is not the primary inference bottleneck. For each token you have to load all of the weights (or activated weights) from memory. This is why Cerebras is fast: they have huge memory bandwidth.
> The most important AI applications being deployed in enterprise today—agents, code generation, and complex reasoning—are bottlenecked by inference latency
Is this really true today? I don't work in enterprise, so don't know how things look like, but I'm sure lots of people here do, and it feels unlikely that inference latency is the top bottleneck, even above humans or waiting for human input? Maybe I'm just using LLMs very differently from how they're deployed in a enterprise, but I'm by far the biggest bottleneck in my setup currently.
It is if you want good results. I’ve been giving Gemini pro prompts for 200+ seconds multiple times per day this week and for such tasks I really like to make it double/triple check and sometimes give the results to Claude for review, too (and vice versa).
Ideally I can just run the prompt 100x and have it pick the best solution later. That’s prohibitively expensive and a waste of time today.
Assuming you experience is working within enterprise, you're then saying that cost is the biggest bottleneck currently?
Also surprising to me that enterprises would use out-of-the-box models like that, I was expecting at least fine-tuned models be used most of the time, for very specific tasks/contexts, but maybe that's way optimistic.
Cost is irrelevant when compared to the salaries of the people using them so they will do basic cost controls but nothing too onerous. And cost is never a reason to prevent solutions being built and deployed.
And most enterprises aren't even doing anything advanced with AI. Just doing POCs with chat bots (again) which will likely fail (again). Or trying to do enterprise search engines which are pointless because most content is isolated per team. Or a few OCR projects which is pretty boring and underwhelming.
You give it a task from hell which the devil himself outsources, like ‘figure out how these fifty repositories of yaml blobs, jinja templates and code generating code generating hcl generating yaml interact to define the infrastructure, then add something to it with correct iams, then make a matching blob of yaml pipelines to work with that infrastructure’
Only an insignificant minority of companies are running their own AI LLM models.
Everyone else is perfectly fine using whatever Azure, GCP etc provide. Enterprise companies don't need to be the fastest or have the best user experience. They need to be secure, trusted and reliable. And you get that by using cloud offerings by default and only going third party when there is a serious need.
I think the key here is twofold. First “the cloud” as commonly understood isn’t what anyone here is talking about. The subject is commercial inference providers.
The “cloud”, or Commercial offerings in storage, VMs, etc are reasonably “secure” in a very general context these days, that is generally true.
OTOH “cloud” AI (commercial inference) is going to use your data for training, incorporating your business processes and domain specific competencies into its innate capabilities, which could eventually impact your value proposition. Empirically, this will happen, eventually, regardless of the user agreement that you signed.
Leakage of proprietary competencies is what is meant by being insecure, in this context.
Second, “cloud isn't secure enough for the enterprise” should be replaced with “enterprise actually cares about security except as a cost/benefit analysis”.
Sending your data to someone else’s data center is a really good way for your data to potentially end up on someone else’s computer. In fact, it’s pretty much the point. If security was the priority, they wouldn’t do that.
Some quant-heads endorsing the latest fad doesn't prove anything. Also they don't care if chinese hackers are vacuuming data cause ballstreet doesn't care about sustainability. But I grant you that secure and trust are just words that don't mean anything anymore anyhow.
LOL, all fintech are using or entering the "cloud" very heavily. Cloud is here for long enough that claiming it's insecure shows only the immense ignorance.
True, the biggest bottleneck is formulating the right task list and ensuring the LLM is directed to find the relevant context it needs. I feel LLMs in their instruction following are often to eager to output rather than using tools (read files) in their reasoning step.
I estimated last year that they can only produce about 300 chips per year and that is unlikely to change because there are far bigger customers for TSMC that are ahead of them in priority for capacity. Their technology is interesting, but it is heavily reliant on SRAM and SRAM scaling is dead. Unless they get a foundry to stack layers for their wafer scale chips or design a round chip, they are unlikely to be able to improve their technology very much past the CSE-3. Compute might somewhat increase in the CSE-4 if there is one, but memory will not increase much if at all.
I doubt the investors will see a return on investment.
>Their CEO is a felon who plead guilty to accounting fraud [...]
Whoa, I didn't know that.
I know he's very close to another guy I know first hand to be a criminal. I won't write the name here for obvious reasons, also not my fight to fight.
I always thought it was a bit weird of them to hang around because I never got that vibe from Feldman, but ... now I came to know about this, 2nd strike I guess ...
While the CEO stuff is a problem, I don't think the other stuff matters.
Per chip area WSE-3 is only a little bit more expensive than H200. While you may need several WSE-3s to load the model, if you have enough demand that you are running the WSE-3 at full speed you will not be using more area in the WSE-3. In fact, the WSE-3 may be more efficient, since it won't be loading and unloading things from large memories.
The only effect is that the WSE-3s will have a minimum demand before they make sense, whereas an H200 will make sense even with little demand.
I did the math last year to estimate how many wafers per year Nvidia had, and from my recollection it was >50,000. Cerebras with their ~300 per year is not able to handle the inference needs of the market. It does not help that all of their memory must be inside the wafer, which limits the amount of die area they have for actual logic. They have no prospect for growth unless TSMC decides to bless them or they switch to another foundation.
> While you may need several WSE-3s to load the model, if you have enough demand that you are running the WSE-3 at full speed you will not be using more area in the WSE-3.
You need ~20 wafers to run the Llama 4 Behemoth model on Cerebras hardware. This is close to a million mm^2. The Nvidia hardware that they used in their comparison should have less than 10,000 mm^2 die area, yet can run it fine thanks to the external DRAM. How is the CSE-3 not using more die area?
> In fact, the WSE-3 may be more efficient, since it won't be loading and unloading things from large memories.
This makes no sense to me. Inference software loads the model once and then uses it multiple times. This should be the same for both Nvidia and Cerebras.
Yes, on an ordinary GPU it loads the weights to GPU memory, but then these weights must be moved from GPU memory onto the chip. But on these the weights can presumably be kept on chip entirely-- that's basically their whole point, so with the Cerebras there's no need to ever move weights to the chip.
Of course these guys depend on getting chips, but so does everybody. I don't know how difficult it is, but all sorts of entities get TSMC 5nm. Maybe they'll get TSMC 3nm and 2nm later than NVIDIA, but it's also possible that they don't.
If you need anything else, you need to load it from elsewhere. In the CSE-3, that would be from other PEs. In Blackwell, that would be from on package DRAM. Idle time in Blackwell be mitigated by parallelism, since each SM has SRAM for multiple kernels to run in parallel. I believe the CSE-3 is quick enough that they do not need that trick.
The other guy said “you will not be using more area in the WSE-3”. I do not see this die area efficiency. You need many full wafers (around 20 with Llama 4 Maverick) to do the same thing with the CSE-3 that can be done with a fraction of a wafer with Blackwell. Even if you include the DRAM’s die area, Nvidia’s hardware is still orders of magnitude more efficient in terms of die area.
The only advantage Cerebras has as far as I can see is that they are fast on single queries, but they do not dare advertise figures for their total throughput, while Nvidia will happily advertise those. If they were better than Nvidia at throughput numbers, Cerebras would advertise them, since that is what matters for having mass market appeal, yet they avoid publishing those figures. That is likely because in reality, they are not competitive in throughput.
To give an example of Nvidia advertising throughput numbers:
> In a 1-megawatt AI factory, NVIDIA Hopper generates 180,000 tokens per second (TPS) at max volume, or 225 TPS for one user at the fastest.
Cerebras strikes me as being like Bugatti, which designs cars that go from start to finish very fast at a price that could buy dozens of conventional vehicles, while Nvidia strikes me as being like Toyota, which designs far lower vehicles, but can manufacture them in a volume that is able to handle a large amount of the world’s demand for transport. Bugatti can make enough vehicles to bring a significant proportion of the world from A to B regularly, while Toyota can. Similarly, Cerebras cannot make enough chips to handle any significant proportion of the world’s demand for inference, while Nvidia can.
Openai wanted to buy them. G42 the largest player in middle east owne a big chunk. You are simply wrong about big investors not touching them but my guess is they will be bought soon by Meta or Apple.
I can only imagine Apple being interested. Their NPU hardware is slower than Qualcomm's, their GPUs have been lagging behind Nvidia in all fields since launch, and they refuse to work with any industry leaders to ship a COTS solution. They don't have many options left on the table, "figuring out how to optimize Apple Silicon" has been the plan for 6 years now and no CUDA-killers have come up out of the woodworks since then.
Either Apple entirely forfeits AI to the businesses capable of supplying it, or they change their tactic and do what Apple does best; grossly overpay for a moonshot startup that promises "X for the iPhone". I don't know if that implicates Cerebras, but clearly Apple didn't retain the requisite talent to compete for commercial AI inference capacity.
Cerebras’ technology works by using an entire wafer as a chip and power draw is 23kW if I recall correctly. Their technology cannot be scaled down and only works when scaling up. They could not be more useless for Apple’s purposes. Acquiring them would only give them a bunch of chip design engineers that might or might not be able to make a decent NPU that uses DRAM.
That said, Apple has some talented people already and they likely just need to iterate to make their designs better. Bringing new people on board would just slow progress (see the mythical man month).
I tried some Llama 4s on Cerebras and they were hallucinating like they were on drugs. I gave it a URL to analyse a post for style and it made it all up and didn't look at the url (or realize that it hadn't looked at it).
I love Cerebrus. 10-100x faster than the other options. I really wish the other companies realized that some of us prefer our computer be instant. I use their API (with Qwen3 reasoning model) for ~99% of my questions, and the whole answer finishes in under 0.1 seconds. Keeps me in a flow state. Latency is jarring. Especially the 5-10 seconds most AIs take these days, where it's just enough to make switching tasks not worth it. You just have to sit there in statis. If I'm willing to accept any latency, might as well make it a couple minutes in the background, and use a full agent mode or deep research AI at that point. Otherwise I want instant.
Yes, the issues were fixed ~1-2 weeks after release. It's a good "all-rounder" model, best compared to 4o. Good multilingual capabilities, even in languages not specifically highlighted. Fast to run inference on it. Code is not one of its strong suits at all.
This is incorrect. The unreleased Llama 4 Behemoth is the largest and most powerful in the Llama 4 family.
As for the speed record, it seems important to keep it in context. That comparison is only for performance on 1 query, but it is well known that people run potentially hundreds of queries in parallel to get their money out of the hardware. If you aggregate the tokens per second across all simultaneous queries to get the total throughput for comparison, I wonder if it will still look so competitive in absolute performance.
Also, Cerebras is the company that not only was saying that their hardware was not useful for inference until some time last year, but even partnered with Qualcomm with the claim that Qualcomm’s accelerators had a 10x price performance improvement over their things:
https://www.cerebras.ai/press-release/cerebras-qualcomm-anno...
Their hardware does inference with FP16, so they need ~20 of their CSE-3 chips to run this model. Each one costs ~$2 million, so that is $40 million. The DGX B200 that they used for their comparison costs ~$500,000:
https://wccftech.com/nvidia-blackwell-dgx-b200-price-half-a-...
You only need 1 DGX B200 to run Llama 4 Maverick. You could buy ~80 of them for the price it costs to buy enough Cerebras hardware to run Llama 4 Maverick.
Their latencies are impressive, but beyond a certain point, throughput is what counts and they don’t really talk about their throughput numbers. I suspect the cost to performance ratio is terrible for throughput numbers. It certainly is terrible for latency numbers. That is what they are not telling people.
Finally, I have trouble getting excited about Cerebras. SRAM scaling is dead, so short of figuring out how to 3D stack their wafer scale chips, during fabrication at TSMC, or designing round chips, they have a dead end product since it relies on using an entire wafer to be able to throw SRAM at problems. Nvidia, using DRAM, is far less reliant on SRAM and can use more silicon for compute, which is still shrinking.