0.2 tok/s is fine for experimentation, but it is not interactive in any meaningful sense. For many use cases, a well-quantized 8B or 13B that stays resident will simply deliver a better latency-quality tradeoff
I didn't really understand the performance table until I saw the top ones were 8B models.
But 5 seconds / token is quite slow yeah. I guess this is for low ram machines? I'm pretty sure my 5950x with 128 gb ram can run this faster on the CPU with some layers / prefill on the 3060 gpu I have.
I also see that they claim the process is compute bound at 2 seconds/token, but that doesn't seem correct with a 3090?
DRAM speeds is one thing, but you should also account for the data rate of the PCIe bus (and/or VRAM speed). But yes, holding it "lukewarm" in DRAM rather than on NVMe storage is obviously faster.
Four channels of DDR4-3200 vs two channels of DDR5-6400 (four subchannels) should come out pretty close. I don't see any reason why the DDR4 configuration would be consistently faster; you might have more bank groups on DDR4, but I'm not sure that would outweigh other factors like the topology and bandwidth of the interconnects between the memory controller and the CPU cores.
I can imagine a couple scenarios in which a high-quality, large model would be much preferred over lower latency models, primarily when you need the quality.
yeah, actually I wanted to see if this was possible at all. I managed to get around 3000 tokens/s on a ps2 with classic transformers, since the emotion engine is capable of 32 bit addresses, but it has like 32gb of ram. So I ran into the question of why was that fast and I couldn't get that speed even with small models, and the deal is that the instructions went right of the memory to the gpu and that's the main difference that does when a regular computer does inference: it has to request the instructions to the cpu every time. As I mentioned too, on professional cards you can avoid these problems naturally, since they got instructions precisely for this, but sadly I don't have 30k bucks to spare on a gpu :(
You can get lots of tokens per second on the CPU if the entire network fits in L1 cache. Unfortunately the sub 64 kiB model segment isn't looking so hot.
But actually ... 3000? Did GP misplace one or two zeros there?
nvmlib/ssd-gpu-dma and BaM (based on the same code base) are pretty cool as they allow you to initiate disk reads/writes directly from a CUDA kernel (so not only reading/writing directly to GPU memory but also allowing the GPU to initiate IO on its own). Sometimes called GPU-initiated I/O or accelerator-initiated I/O.
Could be neat to see what giving the 8b like 6gb ram instead of 10gb. Something in-between, where you still need NVMe, but not like the 3x ratio of the 70b model on 23GB.
Nice work. PCI-P2P (GPU-Direct (tm)) is such great stuff. Cool to see!
Nice. I've been looking at doing something similar, more on the order of running a 1T model with less than half the available VRAM.
One workup indicated it was theoretically possible to modify a piece of SGLang's routing layer to support JIT predict-ahead expert swaps from Gen5 NVMe storage straight into GPU memory.
I'm hoping that proves true. The setup relies on NVIDIA Dynamo, so NIXL primitives are available to support that.
That would be nice to see. Actually I was thinking about getting another 3090 and a mobo upgrade since I'm bottlenecked by pcie3 to tryna run glm 4.7 or 5 at q4_k_m, it should be possible.
Yeah I’ve often wondered why folks aren’t training two tier MoEs for VRAM + RAM. We already have designs for shared experts so it cannot be hard to implement a router that allocated 10x or 100x as often to “core” experts vs the “nice to have” experts. I suppose balancing during training is tricky but some sort of custom loss on the router layers should work.
I’ve also wondered why the routers aren’t training to be serially consistent so you can predict layers to swap into VRAM a few layers ahead to maximize available bandwidth.
1) This is basically the intention of several recent MoE models: keep particular generally useful experts hot in VRAM.
2) Unless you can swap layers in faster than you consume them there is no point to predicting layers (what does this even really mean? did you mean predicting experts?).
It seems at the moment the best you can do is keep experts and layers more likely to be used for a given query in VRAM and offload the rest, but this is work-dependent.
I think part of the issue is that in production deployments, you're batching high enough that you'll be paging in those long tail experts constantly.
Unless you're handing that in some kind of fancy way, you'll be holding up the batch while waiting for host memory which will kill your throughout.
It makes much more sense for non batched local inference, especially if you can keep the MoE routing stable like you say, but most folks aren't optimising for that.
Ideally, you should rearrange batches so that inference steps that rely on the same experts get batched together, then inferences that would "hold up" a batch simply wait for that one "long tail" expert to be loaded, whereupon they can progress. This might require checkpointing partial inference steps more often, but that ought to be doable.
I think this is doable for very long tail experts that get swapped in for specialised topics - say, orbital mechanics.
But for experts that light up at, say, 1% frequency per batch, you're doing an awful lot of transfers from DRAM which you amortize over a single token, instead of reads from HBM which you amortize over 32 tokens.
I feel like we need an entirely new type of silicon for LLMs. Something completely focused on bandwidth and storage probably at the sacrifice of raw computation power.
This is an interesting area for experiments. I suspect that in the longer term model optimization (knowing which bits you can leave out without affecting the functioning of the model) will become the dominant area of research just like it did with compression algorithms because effectively a model is a lossy compression scheme.
And that's good because that increases democratization of AI away from the silos that are being created.
The compression analogy is interesting. Another way of looking at it could be fine-tuning as "knowing what to leave out" - a 3B model for example tuned for a narrow task doesn't need the capacity that makes 70B good at many things.
Yeah, GPUdirect should allow you to dma straight to a storage device.
I wonder... what if the m.2 storage was actually DRAM? You probably don't need persistence for spilling a model off the GPU. How would it fare vs just adding more host memory? The m.2 ram would be less flexible, but would keep the system ram free for the CPU.
Isn't m.2 storage but DRAM - hopefully, meaning NVMe/PCIe not SATA speed - already exists as Compute Express Link (CXL), just not in this specific m.2 form factor? If only RAM wasn't silly expensive right now, one could use 31GB/s of additional bandwidth per NVMe connector.
I gave a talk a few years ago at dask summit (conf?) on making the stars align with dask-cudf here. We were helping a customer accelerate log analytics by proving out our stack for nodes that look roughly like: parallel ssd storage arrays (30 x 3 GB/s?) -> GPUDirect Storage -> 4 x 30 GB/s PCIe (?) -> 8 x A100 GPUs, something like that. It'd be cool to see the same thing now in the LLM world, such as a multi-GPU MoE, or even a single-GPU one for that matter!
0.2 tok/s is slow for chat but perfectly fine for batch/async workloads. I run automated content generation pipelines where a single job kicks off dozens of LLM calls (script generation, metadata, descriptions) and none of them need to be interactive. The whole job takes 20 minutes anyway because of image generation bottlenecks. Being able to run a 70B model locally for those batch calls instead of paying per-token API costs would be a significant cost reduction, even at this speed.
Cost wise it does not seem very effective. .5 token / sec (the optimized one) is 3600 tokens an hour, which costs about 200-300 watts for an active 3090+system. Running 3600 tokens on open router @.4$ for llama 3.1 (3.3 costs less), is about $0,00144. That money buys you about 2-3 watts (in the Netherlands).
Great achievement for privacy inference nonetheless.
Something to consider is that input tokens have a cost too. They are typically processed much faster than output tokens. If you have long conversations then input tokens will end up being a significant part of the cost.
Why is this so damn important? Isn't it more important to end up with the best result?
I (in Norway) use a homelab with Ollama to generate a report every morning. It's slow, but it runs between 5-6 am, energy prices are at a low, and it doesn't matter if it takes 5 or 50 minutes.
Really cool. I'm wondering: what background did you need to be able to think of the question that resulted in this project?
I know you said you're involved in some retrogaming and were experimenting, but as someone who works in a world where hardware is pretty heavily abstracted away, even if I got into retrogaming I don't know that I'd consider that there may be a systems improvement lying around. Beyond the creative aspect, it feels like there is some systems and hardware background that helped put the idea together (and I'd be interested to go learn about of that systems/hardware knowledge myself).
The idea was basically to run a llm on a ps2, then I ran into some problems as the 32mb ram cap with 4mb vram cap; so I had to figure out a way to stream layers on the forward pass. Given that ps2 manages to give instructions directly to the vram that's capable of 32bit addresses, it gave an insane amount of tok/s, then I wondered if I could do the same on my puter
Cool project. Can you provide more details about your DKMS patching process for consumer GPUs? This would be fun to try out, but I’d need some more details on that patch process first.
I updated the documentation to provide more info for the patching process, I added the patches themselves too and provided some risk info about the patches
I've often wondered doing this with extreme compression. What if you did extreme compression + decompression on the GPU? Because you're leaving a lot of compute unused.
I did it, but with different quantization compressions, It ran into quality issues, I will try to rerun with the same quants if that fixes the issue, but the most that looks unused, its being used by rotating layers that are being swapped by the cpu from the ram itself, that manages to keep layers warm, ready to use while inferencing and discarding already used ones
I'm not sure, but I suspect that LLM weights don't compress all that well. The intuition here is that training an LLM is compression of the training data into the weights, so they are probably very information dense already. Can't squeeze them down much.
Bypassing CPU for NVMe-to-GPU transfer is clever. The bottleneck for running large models locally has always been the memory hierarchy — this essentially treats NVMe as extended VRAM with direct DMA.
I wonder how this compares to Apple's unified memory approach on M-series chips for similar workloads. The M4 Max can fit 70B models entirely in memory without any offloading tricks, though at lower throughput than a 3090.
Would be interesting to see comparative benchmarks: this NVMe approach on a 3090 vs M4 Max native, especially for batch inference where the NVMe latency might be amortized.
Cool hack but 0.5 tok/s on 70B when a 7B does 30+ on the same card. NVIDIA's own research says 40-70% of agentic tasks could run on sub-10B models and the quality gap has closed fast.
Really interesting experiment i should have done this before
Do you have numbers on effective throughput vs PCIe theoretical bandwidth?
I’m curious whether this is primarily latency-bound or bandwidth-bound in practice
Can some tell me??
Actually is purely bandwidth-bound, the major bottleneck of the whole process, for me in this case, is the B450 mobo I got that's only capable of pcie3 and 1x8 in the pcie lanes for gpu instead of 1x16; so I'm capped until I get an X570 maybe. I should get around twice or triple the tok speed with that upgrade alone
randomtoast | 23 hours ago
tyfon | 22 hours ago
But 5 seconds / token is quite slow yeah. I guess this is for low ram machines? I'm pretty sure my 5950x with 128 gb ram can run this faster on the CPU with some layers / prefill on the 3060 gpu I have.
I also see that they claim the process is compute bound at 2 seconds/token, but that doesn't seem correct with a 3090?
tgrowazay | 22 hours ago
DDR4 tops out about 27Gbs
DDR5 can do around 40Gbs
So for 70B model at 8 bit quant, you will get around 0.3-0.5 tokens per second using RAM alone.
vlovich123 | 22 hours ago
someguy2026 | 22 hours ago
zozbot234 | 21 hours ago
uf00lme | 21 hours ago
wtallis | 20 hours ago
[OP] xaskasdf | 20 hours ago
Wuzado | 22 hours ago
[OP] xaskasdf | 21 hours ago
derstander | 20 hours ago
anoncow | 18 hours ago
fc417fc802 | 17 hours ago
You can get lots of tokens per second on the CPU if the entire network fits in L1 cache. Unfortunately the sub 64 kiB model segment isn't looking so hot.
But actually ... 3000? Did GP misplace one or two zeros there?
eleventyseven | 16 hours ago
Do you have $2/hr to rent an RTX 6000 96GB or $5/hr for B200 180GB on the cloud?
superkuh | 16 hours ago
[OP] xaskasdf | 7 hours ago
fluoridation | 18 hours ago
throwaway2027 | 23 hours ago
someguy2026 | 22 hours ago
[OP] xaskasdf | 18 hours ago
jonassm | 12 hours ago
jauntywundrkind | 22 hours ago
Nice work. PCI-P2P (GPU-Direct (tm)) is such great stuff. Cool to see!
rl3 | 22 hours ago
One workup indicated it was theoretically possible to modify a piece of SGLang's routing layer to support JIT predict-ahead expert swaps from Gen5 NVMe storage straight into GPU memory.
I'm hoping that proves true. The setup relies on NVIDIA Dynamo, so NIXL primitives are available to support that.
Curious if anyone's tried this already.
[OP] xaskasdf | 21 hours ago
Wuzado | 22 hours ago
rao-v | 22 hours ago
I’ve also wondered why the routers aren’t training to be serially consistent so you can predict layers to swap into VRAM a few layers ahead to maximize available bandwidth.
hedgehog | 21 hours ago
svnt | 21 hours ago
1) This is basically the intention of several recent MoE models: keep particular generally useful experts hot in VRAM.
2) Unless you can swap layers in faster than you consume them there is no point to predicting layers (what does this even really mean? did you mean predicting experts?).
It seems at the moment the best you can do is keep experts and layers more likely to be used for a given query in VRAM and offload the rest, but this is work-dependent.
reitzensteinm | 21 hours ago
Unless you're handing that in some kind of fancy way, you'll be holding up the batch while waiting for host memory which will kill your throughout.
It makes much more sense for non batched local inference, especially if you can keep the MoE routing stable like you say, but most folks aren't optimising for that.
zozbot234 | 21 hours ago
reitzensteinm | 21 hours ago
But for experts that light up at, say, 1% frequency per batch, you're doing an awful lot of transfers from DRAM which you amortize over a single token, instead of reads from HBM which you amortize over 32 tokens.
exabrial | 21 hours ago
garethsprice | 9 hours ago
- https://taalas.com/the-path-to-ubiquitous-ai/ - https://chatjimmy.ai/
exabrial | 7 hours ago
jacquesm | 20 hours ago
And that's good because that increases democratization of AI away from the silos that are being created.
serendip-ml | 16 hours ago
01100011 | 19 hours ago
I wonder... what if the m.2 storage was actually DRAM? You probably don't need persistence for spilling a model off the GPU. How would it fare vs just adding more host memory? The m.2 ram would be less flexible, but would keep the system ram free for the CPU.
javchz | 19 hours ago
TechSquidTV | 18 hours ago
[OP] xaskasdf | 7 hours ago
ElectricalUnion | 17 hours ago
lmeyerov | 14 hours ago
I gave a talk a few years ago at dask summit (conf?) on making the stars align with dask-cudf here. We were helping a customer accelerate log analytics by proving out our stack for nodes that look roughly like: parallel ssd storage arrays (30 x 3 GB/s?) -> GPUDirect Storage -> 4 x 30 GB/s PCIe (?) -> 8 x A100 GPUs, something like that. It'd be cool to see the same thing now in the LLM world, such as a multi-GPU MoE, or even a single-GPU one for that matter!
umairnadeem123 | 17 hours ago
esquire_900 | 17 hours ago
Great achievement for privacy inference nonetheless.
Aerroon | 16 hours ago
It probably won't matter much here though.
teo_zero | 15 hours ago
IsTom | 12 hours ago
dotancohen | 7 hours ago
thatwasunusual | 12 hours ago
Why is this so damn important? Isn't it more important to end up with the best result?
I (in Norway) use a homelab with Ollama to generate a report every morning. It's slow, but it runs between 5-6 am, energy prices are at a low, and it doesn't matter if it takes 5 or 50 minutes.
qoez | 10 hours ago
eleventyseven | 16 hours ago
ekianjo | 14 hours ago
teaearlgraycold | 11 hours ago
[OP] xaskasdf | 7 hours ago
civicsquid | 16 hours ago
I know you said you're involved in some retrogaming and were experimenting, but as someone who works in a world where hardware is pretty heavily abstracted away, even if I got into retrogaming I don't know that I'd consider that there may be a systems improvement lying around. Beyond the creative aspect, it feels like there is some systems and hardware background that helped put the idea together (and I'd be interested to go learn about of that systems/hardware knowledge myself).
rustyhancock | 16 hours ago
Perhaps that's what made them think to try.
Perhaps the current batch of smart memory cards which on the PS2 I believe have quite complex DMA capabilities to stream from the SD card game data.
charcircuit | 15 hours ago
[OP] xaskasdf | 7 hours ago
[OP] xaskasdf | 7 hours ago
The idea was basically to run a llm on a ps2, then I ran into some problems as the 32mb ram cap with 4mb vram cap; so I had to figure out a way to stream layers on the forward pass. Given that ps2 manages to give instructions directly to the vram that's capable of 32bit addresses, it gave an insane amount of tok/s, then I wondered if I could do the same on my puter
Aurornis | 16 hours ago
Maxious | 13 hours ago
[OP] xaskasdf | 7 hours ago
timzaman | 15 hours ago
[OP] xaskasdf | 7 hours ago
sylware | 13 hours ago
spwa4 | 12 hours ago
[OP] xaskasdf | 7 hours ago
nathan_compton | 7 hours ago
MarcLore | 11 hours ago
I wonder how this compares to Apple's unified memory approach on M-series chips for similar workloads. The M4 Max can fit 70B models entirely in memory without any offloading tricks, though at lower throughput than a 3090.
Would be interesting to see comparative benchmarks: this NVMe approach on a 3090 vs M4 Max native, especially for batch inference where the NVMe latency might be amortized.
[OP] xaskasdf | 7 hours ago
stuaxo | 10 hours ago
7777777phil | 10 hours ago
davideom0414 | 9 hours ago
[OP] xaskasdf | 7 hours ago