Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPU
277 points
15 hours ago
| 17 comments
| github.com
| HN
Hi everyone, I'm kinda involved in some retrogaming and with some experiments I ran into the following question: "It would be possible to run transformer models bypassing the cpu/ram, connecting the gpu to the nvme?"

This is the result of that question itself and some weekend vibecoding (it has the linked library repository in the readme as well), it seems to work, even on consumer gpus, it should work better on professional ones tho

7777777phil
39 minutes ago
[-]
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. I was looking int this last year: https://philippdubach.com/posts/nvidia-likes-small-language-...
reply
valianteffort
22 minutes ago
[-]
Have you actually used anything 15B or less? Those models are borderline retarded.
reply
MarcLore
1 hour ago
[-]
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.

reply
01100011
10 hours ago
[-]
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.

reply
javchz
9 hours ago
[-]
Yeah a ramdisk would probably work wonders. It's a shame Intel optane didn't became a standard, those type of workflows would be amazing for it.
reply
TechSquidTV
8 hours ago
[-]
Ahhh damn it. Intel! Come back!
reply
lmeyerov
5 hours ago
[-]
This is exactly what I was wondering

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!

reply
ElectricalUnion
7 hours ago
[-]
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.
reply
umairnadeem123
8 hours ago
[-]
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.
reply
esquire_900
7 hours ago
[-]
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.

reply
teo_zero
6 hours ago
[-]
I think we use different units. In my system there are 3600 seconds per hour, and watts measure power.
reply
IsTom
3 hours ago
[-]
OP probably means watt-hours.
reply
qoez
38 minutes ago
[-]
Open router is highly subsidized. This might be cheaper in the long run once these companies shift to taking profits
reply
Aerroon
7 hours ago
[-]
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.

It probably won't matter much here though.

reply
thatwasunusual
2 hours ago
[-]
> Cost wise it does not seem very effective.

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.

reply
eleventyseven
7 hours ago
[-]
Are you taking into account energy costs of running a 3090 at 350 watts for a very long time?
reply
teaearlgraycold
1 hour ago
[-]
I doubt it’s at full TDP if it’s running at 0.2 tokens per second.
reply
ekianjo
4 hours ago
[-]
You can run a RTX3090 at 250w and still get a lot of its performance with nvidia-smi.
reply
randomtoast
13 hours ago
[-]
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
reply
xaskasdf
11 hours ago
[-]
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 :(
reply
derstander
11 hours ago
[-]
*32MB of RAM (plus 4MB of video RAM and a little sound and IOP memory).
reply
eleventyseven
7 hours ago
[-]
> I don't have 30k bucks to spare on a gpu :(

Do you have $2/hr to rent an RTX 6000 96GB or $5/hr for B200 180GB on the cloud?

reply
superkuh
6 hours ago
[-]
I'd rather not give money to scalper barons if I can avoid it. Fab capacity is going to that for rental rather than hardware for humans.
reply
anoncow
8 hours ago
[-]
3000 tokens per sec on 32 mb Ram?
reply
fc417fc802
7 hours ago
[-]
fast != practical

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?

reply
Wuzado
12 hours ago
[-]
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.
reply
fluoridation
8 hours ago
[-]
That's slower than just running it off CPU+GPU. I can easily hit 1.5 tokens/s on a 7950X+3090 and a 20480-token context.
reply
tyfon
13 hours ago
[-]
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?

reply
tgrowazay
12 hours ago
[-]
LLM speed is roughly <memory_bandwidth> / <model_size> tok/s.

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.

reply
uf00lme
12 hours ago
[-]
Channels matter a lot, quad channel ddr4 is going to beat ddr5 in dual channel most of the time.
reply
wtallis
10 hours ago
[-]
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.
reply
someguy2026
12 hours ago
[-]
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.
reply
vlovich123
12 hours ago
[-]
Faster than the 0.2tok/s this approach manages
reply
zozbot234
12 hours ago
[-]
Should be active param size, not model size.
reply
xaskasdf
11 hours ago
[-]
yeah, actually, I'm bottlenecked af since my mobo got pcie3 only :(
reply
stuaxo
1 hour ago
[-]
Interesting. Can AMD GPUs do direct io like this?
reply
jacquesm
11 hours ago
[-]
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.

reply
serendip-ml
6 hours ago
[-]
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.
reply
civicsquid
7 hours ago
[-]
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).

reply
rustyhancock
6 hours ago
[-]
I wonder too, DMA plays a huge role in most older gaming consoles when the CPUs were far more sluggish.

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.

reply
charcircuit
5 hours ago
[-]
Why not the PS5? That's when games started streaming assets straight from the NVME SSD to the GPU. In this case the assets are weights.
reply
rl3
13 hours ago
[-]
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.

Curious if anyone's tried this already.

reply
xaskasdf
11 hours ago
[-]
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.
reply
Aurornis
6 hours ago
[-]
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.
reply
Maxious
3 hours ago
[-]
the nvidia open source driver has been modded previously to unlock enterprise paywalled features like p2p gpu comms https://blog.chlc.cc/p/rtx4090-p2p-unlocked and vGPU splitting https://open-iov.org/index.php/VGPU_Unlock
reply
Wuzado
12 hours ago
[-]
I wonder - could this be used for multi-tier MoE? Eg. active + most used in VRAM, often used in RAM and less used in NVMe?
reply
rao-v
12 hours ago
[-]
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.

reply
reitzensteinm
12 hours ago
[-]
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.

reply
zozbot234
12 hours ago
[-]
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.
reply
reitzensteinm
12 hours ago
[-]
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.

reply
svnt
12 hours ago
[-]
Maybe I am misunderstanding something but:

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.

reply
hedgehog
12 hours ago
[-]
I don't have links handy but there is active research in this area.
reply
throwaway2027
13 hours ago
[-]
Didn't DirectX add an API for loading assets directly to GPU memory? Would that work?
reply
someguy2026
13 hours ago
[-]
My impression is that that is limited to assets and really needs to fit into the DirectX framework. From what I can tell, the gpu-nvme-direct is mostly similar to https://github.com/enfiskutensykkel/ssd-gpu-dma and https://github.com/ZaidQureshi/bam
reply
xaskasdf
9 hours ago
[-]
Actually this idea was fueled by those since I went to check if there was anything near to what I wanted to achieve, pretty useful tho
reply
jonassm
3 hours ago
[-]
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.
reply
exabrial
12 hours ago
[-]
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.
reply
spwa4
2 hours ago
[-]
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.
reply
sylware
3 hours ago
[-]
Isn't that linux DMA buf?
reply
timzaman
5 hours ago
[-]
Umm sorry but the cpu can easily keep up shuttling around to/from your nvme. Especially ancient gen3 pcie. Not sure why ud do this.
reply
jauntywundrkind
13 hours ago
[-]
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!

reply