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.
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.
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.
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).
Mistral says they run Le Chat on Cerebras
https://www.cerebras.ai/press-release/cerebras-qualcomm-anno...
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?
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.
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.
Your analysis seems spot on to me.
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.
Care to explain? I don't see it.
400B parameters would need 18 chips. Then you need a bit more ram for other stuff
Just takes one breakthrough and it's all different. See the recent diffusion style LLMs for example
https://cerebras-inference.help.usepylon.com/articles/192554...
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.
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.
Is it as simple as stating in the prompt:
Spend 200+ seconds and review multiple times <question/task>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.
https://milled.com/theinformation/cerebras-ceos-past-felony-...
Experienced investors will not touch them:
https://www.nbclosangeles.com/news/business/money-report/cer...
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.
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 ...
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.
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.
see https://www.cnbc.com/2024/10/11/cerebras-ipo-has-too-much-ha...
IPO was supposed to happen in autumn 2024.
I was just at Dell Tech World and they proudly displayed a slide during the CTO keynote that said:
"Cost per token decreased 4 orders of magnitude"
Personally speaking, not a business I'd want to get into.
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.
You must be living under a rock if you think the cloud isn't secure enough for the enterprise.
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.
I can't imagine Apple being interested.
Their priority is figuring out how to optimise Apple Silicon for LLM inference so it can be used in laptops, phones and data centres.
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.
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).
> 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.
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.
https://hc2024.hotchips.org/assets/program/conference/day2/7...
Similarly, the SMs in Blackwell have up to 228kB of RAM:
https://docs.nvidia.com/cuda/archive/12.8.0/pdf/Blackwell_Tu...
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.
https://blogs.nvidia.com/blog/ai-factory-inference-optimizat...
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.
CSE systems also come with off-chip memory, comparable to a GPU's memory, but usually in the TB range.
Here [1] they imply they can reach 1.2Tbps (allegedly, I know), and that's the previous generation ...
1: https://f.hubspotusercontent30.net/hubfs/8968533/Virtual%20B...
Of course they're using the on-chip SRAM, why wouldn't they?
This is a press release from Cerebras about a Cerebras chip, ... of course they are using a Cerebras chip!
Is that not obvious?
Empirically we know that the data is the most valuable input to cloud services, and eventually it will be used, regardless of the user agreement. When the stored data becomes worth more than the company, it will be eaten and stripped by vulture capital. Law of the jungle, baby.
Just one of the later examples of a very long list of cloud data breaches affecting millions of users. But hey who cares as long as it does not affect your own bottom line.
Any fintech (and these can afford smart people) is building with defense in depth, encrypting everything with their own keys, using ephemeral credentials (eg issued by hashicorp vault), etc, etc.
You're seemingly applying your own experience with cloud-based storage, like Dropbox, to the enterprise cloud-based infrastructure.
I don't feel like I should spend any time laying out my professional experience with these environments, I guess you could just skim through one of the books and watch a couple hours long video explaining layers of the leading "cloud" offerings.
And yes, eventually the breach will happen. Like it happens on premise all the time. 2014 Sony and 2020 Solar Winds are good examples.
Let's agree to disagree, I really don't want to spend any more time on this, I know how a good solution (passing multiple audits and pentests) looks like, you however have your opinion. I'm not going to fight you :)
Take care!
Such a bizarre interpretation considering we still use SMS
https://www.cerebras.ai/blog/cerebras-cs-3-vs-nvidia-b200-20...
https://www.cerebras.ai/blog/announcing-the-cerebras-archite...
It is useless for inference, but it is great for training. It used to be more prominent on their website, but it is harder to find references to it now that they are mimicking Groq’s business model.
Edit: yeah, double checked their site and everything. Dang, their IO is indeed "slow". They claim 1 microsecond latencies, but still, an H100 can move much more data than that.
I agree that Cerebras manufacture <300 wafers per year. Probably around 250-300, calculated from $1.6-2 million per unit and their 2024 revenue.
I don't really see how that matters though. I don't see how core counts matter, but I assume that Cerebras is some kind of giant VLIW-y thing where you can give different instructions to different subprocessors.
I imagine that the model weights would be stored in little bits on each processor and that it does some calculation and hands it on.
Then you never need to load the the weights, the only thing you're passing around is activations with them going from wafer 1, to wafer 2, etc. to wafer 20. When this is running at full speed, I believe that this can be very efficient, better than a small GPU like those made by NVIDIA.
Yes, a lot of the area will be on-chip memory/SRAM, but a lot of it will also be logic and that logic will be computing things instead of being used to move things from RAM to on-chip memory.
I don't have any deep knowledge of this system, really, nothing beyond what I've explained here, but I believe that Mistral are using these systems because they're completely superb and superior to GPUs for their purposes, and they will made a carefully weighed decision based on actual performance and actual cost.
Mistral is a small fish in the grander scheme of things. I would assume that using Cerebras is a way to try to differentiate themselves in a market where they are largely ignored, which is the reason Mistral is small enough to be able to have their needs handled by Cerebras. If they grow to OpenAI levels, there is no chance of Cerebras being able to handle the demand for them.
Finally, I had researched this out of curiosity last year. I am posting remarks based on that.
On WSE-3s however, there's enough memory that the model can actually be stored on-chip provided that you have a sufficient number of them. 20 are enough for some of the largest open models.
This, depending on how it's set up, allows more efficient use of what logic is available, for actually doing computations instead of just loading and unloading the weights. This can potentially make a system like this much more efficient than a GPU.
It doesn't matter whether Mistral are small fish or not. I don't agree that they are small fish, but whether or not they are they are experts. They are very capable people. They haven't chosen Cerebras to be different, they've chosen it because they believe it's the best way to do inference.
If you do the math you will find that Cerebras loses in all of them. They need 460 kW from 20x CSE-3 nodes to do inference for Llama 4 Maverick. A single DGX-200 node only needs 14.4kW. If you buy 32 nodes so that power consumption is the same and naively give each a full copy of the model, you would get 32,000 T/sec aggregate from a batch size of 1 while the 20 CSE-3 node cluster only gets 2,500 T/sec aggregate from a batch size of 1. This is having spent only $16 million for the 32 DGX B200 nodes versus the $40 million for the 20 CSE-3 nodes. Each DGX B200 node has 1.4TB of memory, while the CSE-3 cluster has only 880GB of memory. The CSE-3 cluster will run out of memory as you scale the batch size and context length. Now, if you buy another 15 CSE-3 nodes, you could match the memory of a single DGX B200, but then you could just store partial models on each DGX-200 like how Cerebras stored partial models on each CSE-3, and suddenly, you have more memory to scale to higher batch sizes on the Nvidia hardware. At some point, you will likely become compute bound and cannot keep scaling up the batch size, but that is hard to predict without actually testing for it. The prediction for what the CSE-3 could do based on advertised memory bandwidth was off by a factor of >1000 when given real data. It seems reasonable to think that what it can do as far as compute will similarly be limited to well below the theoretical capability.
Note that my numbers for power consumption were from Cerebras:
https://www.cerebras.ai/blog/cerebras-cs-3-vs-nvidia-b200-20...
Interestingly, the peak number for the DGX B200 is based on the power supplies for the DGX B200 and is actually 0.1 kW higher than Nvidia’s specification that puts it at 14.3kW:
https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-...
PSU peak output is always in excess of the maximum power usage capability of the hardware, but I did not know how Cerebras determined their 23kW figure, so I went with the Cerebras figure for Nvidia, even though I know it is unrealistically high. This likely gave Cerebras the benefit of a handicap on Nvidia’s hardware in the comparison, such that reality is even more in favor of Nvidia.
Calling Cerebras’ hardware the best way of doing inference is ridiculous. We are talking about doing mostly linear algebra. There is no best way of doing it. Pointing at Mistral to say that Cerebras has the best way is an absurd appeal to authority. None of the major players are using them, since they are incapable of handling their needs. The instant responses are nice and are a way for mistral to differentiate itself, but their models are not as good as those from others and few people use them, which is why Cerebras has the capacity to handle their needs.
From a historical standpoint, Cerebras is very similar to Thinking Machines Corporation, which went out of business after 11 years when there was a market downturn because they could not secure business. Cerebras is hemorrhaging money and is only in business because they found some investors willing to cover their losses. Once they run out of people willing to give them money (likely during the next AI winter), they will become insolvent, no matter how good their technology is. When the next AI winter hits, Mistral will likely become insolvent too, since they similarly are hemorrhaging money and are only in business because they found some investors willing to cover their losses.
By the way, you are lecturing someone who actually has worked on code for doing inference:
I will have to think through your comment, but won't be able to do so properly this month.