It makes fairly little sense for consumer hardware, unless you're embedding an absolutely tiny model. Since you're embedding the entire model in the chip itself, required chip size scales directly with model size. The reason they have llama3.1-8B as their model in https://chatjimmy.ai/ (their demo) is because it's very hard to do anything bigger than that, you literally start running out of physical space on the chip, and bigger chips cost money. You can't have a dishwasher using a chip the size of a B200 die to run a model you could run on a Raspberry Pi.
What's worse is that this is when it's already quantized to ~3 bits per parameter (which is fairly lobotomized). Yes, the chip will run it 1000X faster than the Raspberry Pi, but it will only be stupid faster.
Their press release explicitly lists that their HC1 puts the entire Llama 3.1 8B model on one 815 mm² TSMC N6 die, with about 53 billion transistors.
815 mm² is fucking huge. An RTX 5090 is 750 mm². A mid-to-high end consumer CPU die is something like 300 mm², with a lot of budget parts being significantly smaller than that, down to ~70 mm². Every square mm costs money.
If they upgraded from N6 to N3P they MIGHT be able to get as much as a ~35B class model on a a chip which fits in the reticle limit, maybe, probably not, because model weights aren't the only thing that needs to fit on the chip.
There are very serious issues with agentic performance in this setup, which is exactly where you would want something really fast. Their Llama 3.1 demo lists a context of 6,144, which is dramatically lower than the 131,072 Llama-3.1 supports.
Reasoning models are barely usable with contexts that short.
The reason for this is that to actually get those speeds, the KV cache needs to live in SRAM. You can't bake the KV cache into the circuitry since it... changes. They clearly don't have enough SRAM, and the problem gets worse the bigger you make the model since KV cache grows (sort of) with model dim. The longer you want to make your context, the more of your chip needs to be SRAM.
Frankly, I don't see the use-case for this tech. It's too expensive and too inflexible. Just doing what Cerebras did and making a wafer-sized chip which is mostly SRAM is a much better solution to serving LLMs at extreme speeds and you don't need to make a new chip every time a new model comes out.
Yeah, makes sense, if it's good for very small models only, then there's no point, as those van already run on cheap consumer hardware.
Yet, maybe it can work well enough, so that as a manufacturer, you don't pay $50 for a PI, but only $0.50 for a tiny "hard-coded" chip.
The advantage can be that, as a LLM, as opposed to other types of chips, the use-cases could be more varied, so same chip could be use in different devices (robo vacuums, security cameras, ball-shooting training robots, etc.)
Lots of chiplet interconnects hurt your max performance pretty severely and drives up your complexity a lot. There's a reason they want to keep it monolithic.
I already harped on Cerebras, but their approach of just copy/pasting a whole bunch of identical functional blocks, over-provisioning the chip by ~8%, and then just fusing off blocks with defects allows them to effectively have 100% yield on a wafer-sized monolithic chip. This is very desirable, and just another reason I like their approach better.
Kimi Delta Attention (KDA), despite having "Attention" in the name, isn't really attention at all in any conventional sense. It's more like an RNN which can be efficiently parallelized during training. It's a very small modification to Gated DeltaNet, which can be described as an RNN whose hidden state acts like a small, editable attention memory.
Because it's RNN-like, it has an inherent idea that X comes before Y which comes before Z in the sequence XYZ. Transformers, by default, don't have that. They operate on sets, unordered collections of unique items. They have no idea where those items are in relation to each-other so you have to clue them in.
Because There are 3 KDA layers per attention layer, and 3 KDA layers before the first attention layer, every single token position is going to be able to learn information about where it is in the sequence before the first actual attention layer.
RoPE is actually a bit destructive, so being able to omit it like this is very convenient. Models like Gemma-4 have a similar structure with 5:1 Sliding Window Attention (SWA) layers for every global attention layer. These are cheap, shitty attention layers which handle local information and which go in-between the big powerful ones, KDA serves the same role in this model. In Gemma only the SWA layers have RoPE while the global attention layers omit it. SWA is actual attention, even if it only operates on a small sliding window, so it needs the positional embedding. KDA isn't, so it doesn't.
>Curious to see if it holds up at frontier scale.
I don't know how much more frontier scale you can get than this, but yes, there's no reason why that wouldn't work at larger scales. Honestly, more parameters just makes it easier for the KDA layers to communicate that positional information better.
I see. That paper is looking a median income from a PPP perspective but without accounting for taxes and transfers, which is a tiny bit unorthodox in this exact context which tries to get into the weeds slightly more than just standard PPP adjusted GDP per-capita or something like that. It's valid statistically, but typically if you're trying to adjust for cost of living it makes a lot of sense to use the amount of actual income the individuals have left over after paying taxes and receiving transfers from social programs. Otherwise you'll have a number which accounts for differences in prices but not differences in the proportion of "income" which is actually retained by the earner. This is what that wikipedia article I linked is reporting. Comparing two different countries with very different tax/social program policies with PPP adjustments but without tax/transfer adjustments is less than ideal.
It's not wrong, it's just not very useful. It's typically a more intermediary statistic rather than a final one when doing this type of comparison.
Yes, comparisons are difficult between countries. With respect to unions and wages, not sure hourly wages are that bad a starting point - but happy to look at research there on differential impact. Adjusting for taxes and social programs also can create issues in terms of accounting for things accruing from those for the future (pay as you go pensions come to mind).
You'd similarly factor in comparable medical insurance/
avg. out of pocket medical expenses, taxes etc. for a median income earner the other side then, right?
At median income, does that really tilt the scales much when all is said and done?
Medical expenses would be a combination of cost of insurance / treatment (PPP relevant) and government transfers in a country with partially or fully public healthcare, which is why it's so important to do both.
The idea of periodically stopping to write blocks of recent context into a fast-weight state is interesting, but I think it liked it better when E2E-TTT[1] did it. It's a more flexible and elegant continuous learning approach.
Essentially it goes "You know how your model can remember its training data? Well, what if you treated its recent context like more training data and updated (some of) the weights using (mostly) the same process used to train it?"
The end result is very good at remembering things but also really good at adapting to new unseen distributions.
Yah I think E2E-TTT is a lot more like what people in this comments section are picturing. I can't tell that this method updates model weights at all during the "sleep" period, only the usual SSM state updated by any Mamba model after each token. They just optimized the model to use that SSM state _more_ when an eviction is about to happen.
Yes, since the weights being updated are a small subset of the overall total it's manageable. Just like how each separate conversation currently requires you to store a separate KV cache, you'd need to store the fast weights separately. Both KV cache and fast weight content stores have to be conversation specific, so just setting a bit of extra RAM aside for "memory" isn't really a new ask, just a different format for an old problem.
I wonder if we can get children to make something their life’s dream if we make the cool books about it when they are growing up? I wonder how flexible the human mind can be in convincing itself that it is fulfilling its dream?
>Its turn around time is ridiculous, it has to be maintained with specialized equipment/hangers, along with external contractor assistance.
>Compared to the Gripen, as an example, which can land on a freeway and be up in the air again in a few minutes.
I have no idea where people got the idea that the F-35 requires a major refit after each sortie or that it needs climate controlled hangars, but there's literally no truth to any of it.
The turnaround time for an F-35 after a mission in a wartime scenario isn't going to be much different from any other older fighter jet. Refuel, rearm, get back in the air.
One of the key requirements for the F-35 programs was to minimize extra care needed for the RAM (Radar Absorbent Material). Unlike older stealth aircraft the F-35's ram is "baked in" to the aircraft skin, rather than being a coating. The F-117 and B-2 require climate controlled hangars because their coatings are old and delicate, the F-22 doesn't, but needs regular touch-ups for its coating, the F-35 is just left sitting outside most of the time regardless of where it's operating, a desert, the arctic, a jungle, the deck of a ship, you just leave it out there. The only common maintenance done on the F-35's RAM is replacing a relatively small amount of special RAM tape which is usually used around the edges of the access panels which are opened for other types of maintenance.
Cerebras has effectively 100% yield on these chips. They have an internal structure made by just repeating the same small modular units over and over again. This means they can just fuse off the broken bits without affecting overall function. It's not like it is with a CPU.
That's what it's running on. It's optimized for very high throughput using Cerebras' hardware which is uniquely capable of running LLMs at very, very high speeds.
What's worse is that this is when it's already quantized to ~3 bits per parameter (which is fairly lobotomized). Yes, the chip will run it 1000X faster than the Raspberry Pi, but it will only be stupid faster.
Their press release explicitly lists that their HC1 puts the entire Llama 3.1 8B model on one 815 mm² TSMC N6 die, with about 53 billion transistors.
815 mm² is fucking huge. An RTX 5090 is 750 mm². A mid-to-high end consumer CPU die is something like 300 mm², with a lot of budget parts being significantly smaller than that, down to ~70 mm². Every square mm costs money.
If they upgraded from N6 to N3P they MIGHT be able to get as much as a ~35B class model on a a chip which fits in the reticle limit, maybe, probably not, because model weights aren't the only thing that needs to fit on the chip.
There are very serious issues with agentic performance in this setup, which is exactly where you would want something really fast. Their Llama 3.1 demo lists a context of 6,144, which is dramatically lower than the 131,072 Llama-3.1 supports.
Reasoning models are barely usable with contexts that short.
The reason for this is that to actually get those speeds, the KV cache needs to live in SRAM. You can't bake the KV cache into the circuitry since it... changes. They clearly don't have enough SRAM, and the problem gets worse the bigger you make the model since KV cache grows (sort of) with model dim. The longer you want to make your context, the more of your chip needs to be SRAM.
Frankly, I don't see the use-case for this tech. It's too expensive and too inflexible. Just doing what Cerebras did and making a wafer-sized chip which is mostly SRAM is a much better solution to serving LLMs at extreme speeds and you don't need to make a new chip every time a new model comes out.