News for hackers. Stuff that matters.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by randomtoast on Saturday February 21, 2026 @05:57PM (#47105824)
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
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by xaskasdf on Saturday February 21, 2026 @08:11PM (#47107002)
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 :(
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by derstander on Saturday February 21, 2026 @08:32PM (#47107152)
*32MB of RAM (plus 4MB of video RAM and a little sound and IOP memory).
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by anoncow on Saturday February 21, 2026 @11:14PM (#47108082)
3000 tokens per sec on 32 mb Ram?
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by fc417fc802 on Saturday February 21, 2026 @11:56PM (#47108321)
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?

Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by xaskasdf on Sunday February 22, 2026 @10:48AM (#47111944)
I wondered the same, but the rendering seems right, the output was almost instant. I'll recheck the token counter; anyway as you say, fast isn't practical. Actually I had to develop my own tiny model https://huggingface.co/xaskasdf/brandon-tiny-10m-instruct to fit something "usable", and it's basically a liar or disinformation machine haha
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by eleventyseven on Sunday February 22, 2026 @12:46AM (#47108562)
> 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?

Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by superkuh on Sunday February 22, 2026 @12:47AM (#47108575)
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.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by xaskasdf on Sunday February 22, 2026 @10:00AM (#47111512)
I thought about that, but idk if they allow me to modify the linux kernel and nvidia cuda kernel at all
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by green-salt on Sunday February 22, 2026 @12:23PM (#47112779)
I think you can do a bunch of that on Digitalocean's GPU droplets.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by jonassm on Sunday February 22, 2026 @12:37PM (#47112889)
In those systems you could probably leverage something like Nvidia SCADA or GDS directly.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by xaskasdf on Sunday February 22, 2026 @03:07PM (#47114159)
Actually since they have direct GDS it should perform really well on professional gpus
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by SmithRenaldo on Monday February 23, 2026 @07:07PM (#47130947)
The $5/hr B200 rate is fine for training, but cloud latency usually breaks real-time signal processing. I’ve been hitting similar walls with MemeRadar; when you're processing high-volume spikes, the bottleneck is memory bandwidth, not raw TFLOPS. Quantizing to fit L3 cache is an option, but you lose the precision needed for spotting subtle rug-pull patterns. For 24/7 production workloads, local hardware TCO usually beats cloud rentals.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by Wuzado on Saturday February 21, 2026 @06:57PM (#47106449)
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.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by tyfon on Saturday February 21, 2026 @06:30PM (#47106193)
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?

Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by tgrowazay on Saturday February 21, 2026 @06:59PM (#47106464)
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.

Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by uf00lme on Saturday February 21, 2026 @07:47PM (#47106836)
Channels matter a lot, quad channel ddr4 is going to beat ddr5 in dual channel most of the time.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by wtallis on Saturday February 21, 2026 @09:03PM (#47107352)
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.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by someguy2026 on Saturday February 21, 2026 @07:17PM (#47106600)
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.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by tgrowazay on Monday February 23, 2026 @03:00PM (#47127899)
Yes.

In general systems usually have PCIE version with bandwidth better than RAM of that system.

For example a system with DDR4 (27Gbs) usually has at least PCIE4 (32Gbs at 16x).

But you can bottleneck that by building a DDR5 (40Gbs) system with PCIE4 card.

Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by zozbot234 on Saturday February 21, 2026 @07:28PM (#47106680)
Should be active param size, not model size.
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by tgrowazay on Monday February 23, 2026 @02:55PM (#47127835)
Yes, you’re right.

LLama 3.1 however is not MoE, so all params are active.

For MoE it is tricky, because for each token you only use a subset of params (an “expert”) but you don’t know which one, so you have to keep them all in memory or wait until it loads from slower storage, potentially different for each token.

Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by vlovich123 on Saturday February 21, 2026 @07:13PM (#47106575)
Faster than the 0.2tok/s this approach manages
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by xaskasdf on Saturday February 21, 2026 @08:28PM (#47107112)
yeah, actually, I'm bottlenecked af since my mobo got pcie3 only :(
Re:Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-G... by fluoridation on Saturday February 21, 2026 @10:56PM (#47107992)
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.