This looks very interesting. Possible to get those rates without exotic hardware.
But I have to say that the comparison is not really fair. Comparison is done with a 2 B model vs frontier models that are likely 100s of times larger. Also taalas with their 15000 tok/s inference are suspiciously missing from the comparison.
We need to see the comparison with this framework and useful models, which at present seems to mean ~30 B.
We strived to be fair as possible in the benchmark, but it's indeed not perfect.
Taalas should have been added in the dedicated hardware section, even though they use 3-bit quantization when we are on FP16 (to be fair in both directions) and they burn the model directly on the card.
Our tech preview is about the speed (hence the small dense model, it was easier to implement).
The math checks out though to allow support for large frontier MoE models at similar speeds:
- At batch size 1, GPT-OSS-120B has 5.1B active parameters - in FP8, it's in the same size ballpark than our 2B model in FP16 (5.1 GB vs 4GB).
- DeepSeek V4 Flash has 13B in mixed FP4/FP8, so let's say ballpark around 3x bigger than 4GB - so in theory we could reach >1,000 tok/s on it with MI300X/H200 and up to 4k on next generation GPUs.
Your playground/write-up is very interesting and I would be really interested when you can have something like Deepseek V4 Flash model (49B) running as you are suggesting.
I haven't read the article at the moment and I will try to read them hopefully but I wish to ask a question regarding, can this approach be done for say trillion or large parameter models as well or is there some wall which gets hit that makes it valuable for only smaller parameter model.
That being said, its still really incredible because in future, because these small models are really getting good for many use cases and speed becomes their bottleneck, with greater speeds at consumer hardware, I think its gonna be amazing work!
Consumer inference scenarios tend to be highly bespoke so it's difficult to apply a monokernel approach based on deep manual optimization. I suppose this could become applicable to rare scenarios where both the model and the hardware are fixed and self-contained, e.g. I'm running Apple's AI model on the latest Apple Silicon hardware. Then this becomes a viable approach even for 'consumer' use.
The authors' approach also encompasses multi-node approaches that won't apply easily to consumer inference since consumer GPUs have very low-performance interconnects, hence why layer parallelism is usually favored. (But that doesn't work very well with the monokernel approach, since it involves running distinct logic on each separate GPU. It also doesn't speed up single inference, though you can get that throughput back by pipelining small minibatches.)
scenarios where both the model and the hardware are fixed and self-contained
That's basically antirez's DS4 and it works pretty well because there are few leading models and few hardware platforms (Apple, GB10, Strix Halo) that are worth using.
The last section of the article lays out the scaling laws that apply when porting this approach to another model. In a nutshell, DeepSeek V4 Pro with 49B active params is close to the upper bound.
Also worth noting that our results are currently for standard datacenter GPUs. On consumer hardware, though the same low-level optimization approach applies, the bandwidth limitations will cap the achievable speed.
It looks like DTP is a distinct architectural choice that would require training new models accordingly? This wouldn't be able to just run inference for existing models.
Totally, though DTP is not required for these kind of speeds.
Standard TP works also.
DTP is something we built for our roadmap in order to get to extremely high speeds (like 10k+ tokens/s). When the budget is under 10 µs per layer, any little overhead matters.
For 1k to 5k tokens/s, regular TP still works because we are able to optimize the inter-GPU all-reduce collectives at under 3 µs, which allows to continue streaming model weights in shared memory, registers and caches while GPUs exchange data.
I have been lamenting for a while that the memory-bandwidth <-> tps relationship was pretty much working for small models on consumer cards, but not at all on datacenter hardware.
It's great to see that with proper care on the inference engine implementation the relationship can be restored.
The blog makes it clear that "standard" GPU here is in opposition to purpose-built hardware like Cerebras. The selling point is reaching the same order of magnitude in generative speed as those approaches.
What a lot of use on here are salivating for is the ability to run these on prosumer hardware at home. So we tend to jump to the conclusion that "standard" means "consumer-grade" because that's what we want to see.
Still, very cool work!
I think many would assume "not enterprise" or "not datacenter grade" when someone says "Standard GPUs", but maybe that specific phrase have a specific meaning I'm not familiar with.
Edit: I just tried a 4B model on a RTX Pro 6000, getting ~500 tok/s with llama.cpp not even trying to optimize or change anything, just default settings. I'm sure with vLLM it'd be a lot faster already, still before manually tuning configs. I wouldn't call that card "Standard GPU" either FWIW, but it makes the claimed performance numbers feel not as exciting, especially given the hardware they were using.
- model size: 2B is just for this preview (it was faster to implement), our article explains how we expect to support large frontier MoE at 1,000 to 5,000 tokens/s
- reaching 500 tok/s, or even up to ~1,000 tok/s, on a consumer GPU card is possible with existing inference engines like vLLM. But there is a ceiling.
The hard part comes we you try to be faster than that: these frameworks won't scale higher just by adding GPUs or using faster GPUs. There is a "glass ceiling" due to microseconds lost everywhere in the stack (grid syncs, inter-GPU comms, kernel launches, CPU sampling, etc.).
All our work at Kog is about removing these bottlenecks.
Thank you for explaining. Do you think there are still opportunities for stack optimizations to meaningfully speed up inference on single consumer-grade GPUs?
I'm sure there are, and I really hope we can work on consumer-grade GPUs at some point.
It should be possible to apply the same methodology (digging deep into the hardware details to understand all its little characteristics, and rethinking the inference stack around that).
You can get something pretty fast right now with a Cerebras Coder subscription, sadly I think the best model they had last I checked was the somewhat dated GLM 4.7: https://inference-docs.cerebras.ai/models/overview
I feel like if they got DeepSeek V4 Flash and Pro running on their hardware, even if at less than 1000 tok/s, they’d still be crushing it with any subscription they’d provide, given how generous their token limits were.
As for the demo it's fast and extremely dumb like expected for 2B. I asked how to stop drinking habit and in just one follow-up message it recommended trying 8% ABV. Hilarious.
True, and for third-party models we'll just re-use their public open weights.
There is a time-consuming part, though, that is performed manually by our (human) team: implement the logic of the model in C++ and assembly code in a super-optimized way, co-designed for each specific hardware card.
This can take months.
We hope to accelerate the process with AI agents, but we're not there yet.
For me it's 3.4k tok/s of pure nonsense, the model is bad, you tell it it's wrong, it acknowledge it's wrong and repeats the same nonsense. It reminds me my nephew though. Ask it something like: "I want to play the guitar on the surface of the Moon. What speakers do you suggest." and then "But Moon has no atmosphere, how the sound will travel?".
For new open weights models, will you need to adapt model code and optimization for your inference engine by hand?
It's true that BS=1 is king when it comes to agentic workflows, however these kinds of system serve multiple requests concurrently with dynamic batching. Do you think it will scale as well ?
- yes, we rewrite the whole model code (while keeping the same logic) in CUDA/HIP and assembly, in order to optimize by hand for each GPU type. It's quite tedious for sure, but I guess this is the price to pay to get this kind of results.
- the batching question is a great one. In agentic systems, there is probably a trade-off between sequential thinking/iterations vs parallel exploration of multiple solutions. Also, there could just be multiple independent tasks running in parallel, depending on the use case.
We plan to support a small amount of batching, but it quickly becomes a trade-off vs speed. Pick one for your use case, I guess.
Also to consider: because we answer requests much faster, we are also able to process lots of them without needing high batches - and scaling on multiple nodes is possible.
- open sourcing: maybe, maybe not. I'm still undecided on this. We are a small startup and I'm told that giving our IP away might be shooting ourselves in the feet. On the other side, I think it could be of great benefit to the community and for us... we'll see
Do you think the work will still apply to speculative/alternative decoding methods like MTP and block diffusion, which are making batch=1 decoding less memory bound? Kernel launch overhead and memory transfer become less and less significant as a % of time when computing multiple tokens at once.
Why not, it's one way to look at it!
Although I have yet to see other work with speculative decoding higher than ~1,000 tokens/s., because the other bottlenecks start to matter at that point, and they need to be solved to go further.
Our view is that MTP / speculative decoding could help getting a X multiplier (X = 2 to 6) on the tokens per second speed we currently achieve.
We are a bit greedy, we want to stack optimizations on top of each other to get the maximum speed possible.
It involves additional compute to verify the predicted tokens during the forward pass (it's like a small batch), which should be totally doable for dense models, and will be more tricky for MoEs because it could mean activating more experts and thus more active parameters.
disclaimer: I've known the founder for a while, as legitimate as it gets in deep tech, real years of research and engineering behind this, not vaporware
Huh, interesting. Some parts of this do generalize even to an RTX 6000 Pro Blackwell, I imagine, though we're going to be solidly bottlenecked then on inter-card throughput through the PCIe interface.
Making these claims on a 2B parameter model seems a bit like seeing linear scalability from 1 to 4 cores and then assuming 256 cores will give you a 256x speedup. Or demonstrating massive improvement on datasets that fit in cache and then assuming the same improvements will be present on problem sizes that span the memory of multiple machines. Something tells me that scaling to larger models will be more difficult than assumed.
Yeah, I agree: I'm actually not expecting it to be easy, and there will certainly be several unknown unknowns we'll discover along the way.
Our process has been, and will continue to be, a sequence of (tedious) R&D experiments where the GPU never behaves as expected when pushed to its limits in ways no-one really tested before (I still have nightmares of the L3 cache cross-IOD bottlenecks on MI300X).
IMHO, we did solve the multi-GPU memory bandwidth scaling problem, and thus the linear scaling of the size of the model towards infinity.
But the main difficulties will come from keeping the speed, with steady and continuous memory streaming, while implementing the much more complex architecture of modern frontier MoEs (attention compression tricks, hash layers, routing logic, etc.)
An article with a title saying tokens per second throughput without any qualifier e.g. what size the model is should immediately be classified as spam.
Who cares about token speed? What is the quality of the results like? I don't know why people are so fixated on token speed, since no one cares how quickly it can spew garbage. Most reasonable people are happier waiting a bit more for accurate results.
It also matters for thinking models and for agentic workflows, especially in software engineering, where a lot of tokens need to be output in iterative loops before the user sees any result.
That sounds a little bit like the 64kb memory is enough, then someone invented electron ;P
But joke aside, I think we don't even know yet what is possible if you hit very fast very high token / second numbers if your whole ecosystem behind it can handle it.
You could literaly implement the same solution 100x and benchmark all of them and get only the best result.
You could build and architecture a whole stack in parallel.
You could do massive thinking token / chain of thought.
You could let the LLM analyse everything around you while you type. Like it could tell you that this might create a bug in a different file and why.
We could start doing some type of monte-carlo search with this.
Token generation speed matters for sequential agentic workflows, like software engineering / vibe coding, where a lot of reasoning tokens, code generation, refactoring, testing, etc. happen in a loop before an actual outcome is served to the user.
About model performance, we plan to support the latest frontier models (this tech preview is about the speed of the engine)
In theory yes, although not in a linearly proportional way, because in practice our memory streaming is not yet perfect. There are still some fixed costs that we did not fully optimize (for now).
NVIDIA H200 Is not a standard GPU. 8 of them in a box with a cpu and ram costs close to the same as a house.
I am 100% all about using local models instead of sending someone else all my data and paying for the privilege of doing so, this article is misleading.
I can get a 27b model to kick out 40 tok/s on 16 gb vram. This is the area ripe for development.
If you can’t connect a monitor, it isn’t a standard GPU, at least not in the way people have spoken about GPUs until a few years ago.
Do you think maybe changing your articles title from "Real-time LLM Inference on Standard GPUs" to "Real-time LLM Inference on Standard Datacenter GPUs" might make sense here? Given more people seem confused by the title than not, and you could clear this up relatively easily, at least on your website although might be late to fix the HN title.
But I have to say that the comparison is not really fair. Comparison is done with a 2 B model vs frontier models that are likely 100s of times larger. Also taalas with their 15000 tok/s inference are suspiciously missing from the comparison.
We need to see the comparison with this framework and useful models, which at present seems to mean ~30 B.
We strived to be fair as possible in the benchmark, but it's indeed not perfect. Taalas should have been added in the dedicated hardware section, even though they use 3-bit quantization when we are on FP16 (to be fair in both directions) and they burn the model directly on the card.
Our tech preview is about the speed (hence the small dense model, it was easier to implement).
The math checks out though to allow support for large frontier MoE models at similar speeds: - At batch size 1, GPT-OSS-120B has 5.1B active parameters - in FP8, it's in the same size ballpark than our 2B model in FP16 (5.1 GB vs 4GB). - DeepSeek V4 Flash has 13B in mixed FP4/FP8, so let's say ballpark around 3x bigger than 4GB - so in theory we could reach >1,000 tok/s on it with MI300X/H200 and up to 4k on next generation GPUs.
Check out the math at the end of our blog post:
https://blog.kog.ai/real-time-llm-inference-on-standard-gpus...
I haven't read the article at the moment and I will try to read them hopefully but I wish to ask a question regarding, can this approach be done for say trillion or large parameter models as well or is there some wall which gets hit that makes it valuable for only smaller parameter model.
That being said, its still really incredible because in future, because these small models are really getting good for many use cases and speed becomes their bottleneck, with greater speeds at consumer hardware, I think its gonna be amazing work!
The authors' approach also encompasses multi-node approaches that won't apply easily to consumer inference since consumer GPUs have very low-performance interconnects, hence why layer parallelism is usually favored. (But that doesn't work very well with the monokernel approach, since it involves running distinct logic on each separate GPU. It also doesn't speed up single inference, though you can get that throughput back by pipelining small minibatches.)
That's basically antirez's DS4 and it works pretty well because there are few leading models and few hardware platforms (Apple, GB10, Strix Halo) that are worth using.
The last section of the article lays out the scaling laws that apply when porting this approach to another model. In a nutshell, DeepSeek V4 Pro with 49B active params is close to the upper bound.
Also worth noting that our results are currently for standard datacenter GPUs. On consumer hardware, though the same low-level optimization approach applies, the bandwidth limitations will cap the achievable speed.
they seem to think it scales up because theyre shortening the stack.
Monokernel deep dive (GPU Engineering): http://blog.kog.ai/building-a-single-kernel-latency-optimize...
Delayed Tensor Parallelism (research): http://blog.kog.ai/delayed-tensor-parallelism-for-faster-tra...
To try the speed on the playground: http://playground.kog.ai
DTP is something we built for our roadmap in order to get to extremely high speeds (like 10k+ tokens/s). When the budget is under 10 µs per layer, any little overhead matters.
For 1k to 5k tokens/s, regular TP still works because we are able to optimize the inter-GPU all-reduce collectives at under 3 µs, which allows to continue streaming model weights in shared memory, registers and caches while GPUs exchange data.
I have been lamenting for a while that the memory-bandwidth <-> tps relationship was pretty much working for small models on consumer cards, but not at all on datacenter hardware.
It's great to see that with proper care on the inference engine implementation the relationship can be restored.
In contrast, not enterprise GPUs that cost as much as a car.
> 8× NVIDIA H200
Edit: I just tried a 4B model on a RTX Pro 6000, getting ~500 tok/s with llama.cpp not even trying to optimize or change anything, just default settings. I'm sure with vLLM it'd be a lot faster already, still before manually tuning configs. I wouldn't call that card "Standard GPU" either FWIW, but it makes the claimed performance numbers feel not as exciting, especially given the hardware they were using.
- model size: 2B is just for this preview (it was faster to implement), our article explains how we expect to support large frontier MoE at 1,000 to 5,000 tokens/s
- reaching 500 tok/s, or even up to ~1,000 tok/s, on a consumer GPU card is possible with existing inference engines like vLLM. But there is a ceiling.
The hard part comes we you try to be faster than that: these frameworks won't scale higher just by adding GPUs or using faster GPUs. There is a "glass ceiling" due to microseconds lost everywhere in the stack (grid syncs, inter-GPU comms, kernel launches, CPU sampling, etc.).
All our work at Kog is about removing these bottlenecks.
It should be possible to apply the same methodology (digging deep into the hardware details to understand all its little characteristics, and rethinking the inference stack around that).
Did the article headline not say Standard GPU?
Feels like a preview of the future
I feel like if they got DeepSeek V4 Flash and Pro running on their hardware, even if at less than 1000 tok/s, they’d still be crushing it with any subscription they’d provide, given how generous their token limits were.
There is a time-consuming part, though, that is performed manually by our (human) team: implement the logic of the model in C++ and assembly code in a super-optimized way, co-designed for each specific hardware card.
This can take months.
We hope to accelerate the process with AI agents, but we're not there yet.
You can ask it to implement an algorithm; we provide suggested prompts you can test.
Also, this tech preview is really about the speed of the inference engine (not the model itself) so I'm glad you got 3.4k tok/s!
For new open weights models, will you need to adapt model code and optimization for your inference engine by hand?
It's true that BS=1 is king when it comes to agentic workflows, however these kinds of system serve multiple requests concurrently with dynamic batching. Do you think it will scale as well ?
Any plans to release it open source?
Congratz again for the release
To answer your questions:
- yes, we rewrite the whole model code (while keeping the same logic) in CUDA/HIP and assembly, in order to optimize by hand for each GPU type. It's quite tedious for sure, but I guess this is the price to pay to get this kind of results.
- the batching question is a great one. In agentic systems, there is probably a trade-off between sequential thinking/iterations vs parallel exploration of multiple solutions. Also, there could just be multiple independent tasks running in parallel, depending on the use case.
We plan to support a small amount of batching, but it quickly becomes a trade-off vs speed. Pick one for your use case, I guess.
Also to consider: because we answer requests much faster, we are also able to process lots of them without needing high batches - and scaling on multiple nodes is possible.
- open sourcing: maybe, maybe not. I'm still undecided on this. We are a small startup and I'm told that giving our IP away might be shooting ourselves in the feet. On the other side, I think it could be of great benefit to the community and for us... we'll see
Our view is that MTP / speculative decoding could help getting a X multiplier (X = 2 to 6) on the tokens per second speed we currently achieve.
We are a bit greedy, we want to stack optimizations on top of each other to get the maximum speed possible.
It involves additional compute to verify the predicted tokens during the forward pass (it's like a small batch), which should be totally doable for dense models, and will be more tricky for MoEs because it could mean activating more experts and thus more active parameters.
The demo is very impressive!
disclaimer: I've known the founder for a while, as legitimate as it gets in deep tech, real years of research and engineering behind this, not vaporware
Our process has been, and will continue to be, a sequence of (tedious) R&D experiments where the GPU never behaves as expected when pushed to its limits in ways no-one really tested before (I still have nightmares of the L3 cache cross-IOD bottlenecks on MI300X).
IMHO, we did solve the multi-GPU memory bandwidth scaling problem, and thus the linear scaling of the size of the model towards infinity. But the main difficulties will come from keeping the speed, with steady and continuous memory streaming, while implementing the much more complex architecture of modern frontier MoEs (attention compression tricks, hash layers, routing logic, etc.)
This is our main use case.
For instant code generatio, 400-500 tok/s should be sufficient, though most frontier models give us closer to 70 tok/s.
But joke aside, I think we don't even know yet what is possible if you hit very fast very high token / second numbers if your whole ecosystem behind it can handle it.
You could literaly implement the same solution 100x and benchmark all of them and get only the best result.
You could build and architecture a whole stack in parallel.
You could do massive thinking token / chain of thought.
You could let the LLM analyse everything around you while you type. Like it could tell you that this might create a bug in a different file and why.
We could start doing some type of monte-carlo search with this.
About model performance, we plan to support the latest frontier models (this tech preview is about the speed of the engine)
each time getting 3300+ tps.
I guess with 1B or 500M model inference would be even faster?
I am 100% all about using local models instead of sending someone else all my data and paying for the privilege of doing so, this article is misleading.
I can get a 27b model to kick out 40 tok/s on 16 gb vram. This is the area ripe for development.
If you can’t connect a monitor, it isn’t a standard GPU, at least not in the way people have spoken about GPUs until a few years ago.
Sorry for the confusion
The math checks out though to allow support for large frontier MoE models at similar speeds.
At batch size 1, GPT-OSS-120B has 5.1B active parameters - in FP8, it's in the same size ballpark than our 2B model in FP16 (5.1 GB vs 4GB).
DeepSeek V4 Flash has 13B in mixed FP4/FP8.
Check out the math at the end of our blog post: https://blog.kog.ai/real-time-llm-inference-on-standard-gpus...
That means Jensen can add another 30 times faster when comparing Rubin to Blackwell without having to actually do anything.
Hopefully that means he won't have any problem to make another 150 billion in profit in the next year.
Sorry for the sarcasm. Looks like interesting work.