Hard disagree. The difference in performance is not something you'll notice if you actually use these cards. In AI benchmarks, the RTX 3090 beats the RTX 4080 SUPER, despite the latter having native BF16 support. 736GiB/s (4080) memory bandwidth vs 936 GiB/s (3090) plays a major role. Additionally, the 3090 is not only the last NVIDIA consumer card to support SLI.

It's also unbeatable in price to performance as the next best 24GiB card would be the 4090 which, even used, is almost tripple the price these days while only offering about 25%-30% more performance in real-world AI workloads.

You can basically get an SLI-linked dual 3090 setup for less money than a single used 4090 and get about the same or even more performance and double the available VRAM.

If you run fp32 maybe but no sane person does that. The tensor performance of the 3090 is also abysmal. If you run bf16 or fp8 stay away from obsolete cards. Its barely usable for llms and borderline garbage tier on video and image gen.

Actual benchmarks show otherwise.

> The tensor performance of the 3090 is also abysmal.

I for one compared my 50-series card's performance to my 3090 and didn't see "abysmal performance" on the older card at all. In fact, in actual real-world use (quantised models only, no one runs big fp32 models locally), the difference in performance isn't very noticeable at all. But I'm sure you'll be able to provide actual numbers (TTFT, TPS) to prove me wrong. I don't use diffusion models, so there might be a substantial difference there (I doubt it, though), but for LLMs I can tell you for a fact that you're just wrong.

To be clear, we are not discussing small toy models but to be fair I also don't use consumer cards. Benchmarks are out there (phoronix, runpod, hugginface or from Nvidias own presentation) and they say it's at least 2x on high and nearly 4x on low precision, which is comparable to the uplift I see on my 6000 cards, if you don't see the performance uplift everyone else sees there is something wrong with your setup and I don't have the time to debug it.

> To be clear, we are not discussing small toy models but to be fair I also don't use consumer cards.

> if you don't see the performance uplift everyone else sees there is something wrong with your setup and I don't have the time to debug it.

Read these two statements and think about what might be the issue. I only run what you call "toy models" (good enough for my purposes), so of course your experience is fundamentally different from mine. Spending 5 figures on hardware just to run models locally is usually a bad investment. Repurposing old hardware OTOH is just fine to play with local models and optimise them for specific applications and workflows.

Even with something like a 5090, I’d still run Q4_K_S/Q4_K_M because they’re far more resource-efficient for inference.

Also, the 3090 supports NVLink, which is actually more useful for inference speed than native BF16 support.

Maybe if you're training bf16 matters?

That's a smart thing todo considering a 5090 has native tensor cores for 4bit precision...