NVIDIA GeForce RTX 3090

What hackers pick instead

• NVIDIA GeForce RTX 5090A top consumer GPU for very fast local AI inference and 4K gaming, but 32GB VRAM limits larger models and street prices run well above MSRP.102recommend
• NVIDIA RTX PRO 6000Recommended for local AI models and fine-tunes with 96GB VRAM and fast compute. Main caveat: very high price, and little gain over an RTX 5090 for inference.77recommend
• Nvidia GPUsPeople recommend Nvidia GPUs for top performance, mature software support and a strong CUDA ecosystem; critics point to high prices and closed-source drivers.67recommend
• AMD Radeon GPURecommended for its open-source Linux drivers, reliable kernel support, and good performance for the price. Critics say its software stack, especially for ML and AI work, lags behind Nvidia.48recommend
• AMD Radeon AI Pro R970032 GB of VRAM for much less than Nvidia cards, fast enough for local LLMs, with quiet, easy drivers. The most common caveat is weak vLLM support.45recommend

Every quote

The 40 most detailed of 208 comments. Each links to the original.

1. I did this with used parts and cheaper consumer cards (3090s) and did much of the same calculations. I found it was way cheaper for me as well.

   Recommends · May 2026 · in “Was my $48K GPU server worth it?”

2. 3090 and 2x3090 are quite popular. But if you uses gigantic (for local models) context of 200k it will go south pretty quickly

   Mixed · Jun 2026 · in “Local Qwen isn't a worse Opus, it's a different tool”

3. If you compare the performance of a 3090 and above vs any Apple hardware you would be insane to go with the Apple hardware.

   Recommends · Mar 2026 · in “Apple discontinues the Mac Pro with no plans for future hardware”

4. price on those is unfortunately high, in fact so high that getting a new Radeon AI PRO R9700 might be a better deal

   Mixed · Aug 2026 · in “llama.cpp”

5. if you want the best possible AI machine at below $5k... actually you should probably just buy an RTX 3090 or 5090

   Recommends · Jul 2026 · in “AMD Ryzen AI Halo – $4k AI Dev Kit”

6. even the most often recommended GPU which is a 3090 with 24 GB VRAM, you can find used going for $700 at minimum

   Mixed · Mar 2026 · in “Is the Future of AI Local?”

7. A used 3090 is probably the most recommended card in the local LLM community right now, and for good reason.

   Recommends · Mar 2026 · in “Quantization from the Ground Up”

8. two of them (or three 3090/4090s) have the same total memory as your Mac, are in the same ballpark in price

   Recommends · Dec 2025 · in “Micron Announces Exit from Crucial Consumer Business”

9. They're comparing old GPUs, so it's interesting that they can still best the DGX Spark in token generation.

   Mixed · Jun 2026 · in “DGX Spark vs. Mac Studio and Halo”

10. a dual 3090 setup which you can build for $3k with parts from facebook marketplace and run in your basement

    Recommends · Jul 2026 · in “Jamesob's guide to running SOTA LLMs locally”

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

    Recommends · Dec 2025 · in “A guide to local coding models”

12. 3090 gets 100+ tokens/second for QWEN, very close to what you would see with a cloud based model.

    Recommends · Mar 2026 · in “Can I Run AI locally?”

13. A couple of 3090s at about 800-900 euros each, as much RAM as you can afford is king right now.

    Recommends · Aug 2026 · in “Qwen3.8 27B scores 52 on Artificial Analysis”

14. A single 3090 will run Qwen3.6 27b fine, and its VRAM speed is twice of what the best Mac has.

    Recommends · Jun 2026 · in “Qwen 3.6 27B is the sweet spot for local development”

15. That 3090 is going to burn 750W and it will still cap you at a 4 bit quant and ~48K context.

    Mixed · Jun 2026 · in “Qwen 3.6 27B is the sweet spot for local development”

16. substantially more powerful than the highest end overhyped Apple product for 1/5th the price

    Recommends · Jun 2026 · in “Qwen 3.6 27B is the sweet spot for local development”

17. It's also unbeatable in price to performance as the next best 24GiB card would be the 4090

    Recommends · Dec 2025 · in “A guide to local coding models”

18. I also get a boost in tokens per second (not double, about 1.5x compared to a single GPU).

    Recommends · Jun 2026 · in “Qwen 3.6 27B is the sweet spot for local development”

19. people will be buying used (3|4|5)090 GPUs to run local models at much better performance

    Recommends · Jun 2026 · in “Apple to Skip High-End M6 Mac Chips in Favor of AI-Focused M7 Line”

20. a single 3090 already provides a great experience at a reasonable quant and context size

    Recommends · Jul 2026 · in “Qwen3.8 is launching and going open-weight soon”

21. It's already pretty insane how good of a model I can run on two old RTX-3090s locally.

    Recommends · Mar 2026 · in “Anthropic: Claude Code users hitting usage limits 'way faster than expected'”

22. I'm so happy for the two used 3090s I bought for $500 each after Ethereum mining ended

    Recommends · Sep 2026 · in “Show HN: Sunk Cost – How long until a local LLM rig pays for itself?”

23. If you pooled it into one system, that would be 72gb of some pretty damn fast memory.

    Recommends · Jul 2026 · in “GTX 1080s: Testing a Legend”

24. I’m having a great time with an NVIDIA 3090. 24 GB RAM will run a lot of neat models.

    Recommends · Jul 2026 · in “You Could Have Come Up with Kimi Delta Attention”

25. I had this problem on Fedora and my RTX 3090. Never got to the bottom of it.

    Mixed · Aug 2026 · in “There's no reason for software to be slow anymore”

26. It's a little bit power hungry but fairly impressive performance wise.

    Mixed · Dec 2025 · in “macOS 26.2 enables fast AI clusters with RDMA over Thunderbolt”

27. 48GB of VRAM is just not enough to use a local LLM like a cloud model

    Mixed · Oct 2026 · in “JetBrains reports revenue growth, net financial loss for 2025”

28. AIUI you can only connect two 3090s at a time with NV link? So you’d only have 48GB of fast combined memory?

    Criticizes · Aug 2026 · in “Apple introduces M6 and M5 Ultra”

29. do all nvidia cards have the same issue with Linux drivers, where the fan won't ever go below 30%?

    Criticizes · Dec 2025 · in “Pop_OS 24.04 LTS with COSMIC desktop environment”

30. AFAIK this comes down a lot to NVIDIA not doing enough efforts for the Linux drivers

    Criticizes · Jan 2026 · in “Linux is good now; to feel like you actually own your PC, put Linux on it”

31. The 3090's in my area seem pretty expensive for a several year old second hand card

    Criticizes · Apr 2026 · in “OpenClaw isn't fooling me. I remember MS-DOS”

32. Are you aware that your 3090s have nowhere close to 256GB of VRAM?

    Criticizes · Mar 2026 · in “Can I Run AI locally?”

33. With gfx cards, you can throw big codebases at the local models and it will happily crunch through them in very little time.

    Recommends · Oct 2026 · in “New iMac with M6”

34. the coil noise on my open-case 3090 has been a fascinating and joyful complement to my generative experiments

    Recommends · Jan 2026 · in “Mystery: Why do some LLMs produce more coil noise on Mac Studio M3 Ultra?”

35. It's much easier to just find a 3090/4090/5090 and keep pace with the software and hardware simultaneously.

    Recommends · Jul 2026 · in “RTX 2080 Ti Memory Upgrade to 22 GB”

36. I run my 3090 pretty heavily with inference workloads and it stays cooler than when I use it for gaming.

    Recommends · Aug 2026 · in “Qwen 3.8-Flash-Next releasing tomorrow (125B a6B)”

37. a 3090 would be a billion times better for games and can probably hold its own for a broader AI workload

    Recommends · Apr 2026 · in “Intel Arc Pro B70 Review”

38. RTX 3090 came out a long time ago and it may be 'outdated' at this point but still banging like a champ

    Recommends · Sep 2026 · in “GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price”

39. Two RTX 3090s, RTX 4000 Pros, or R9700s, provided the motherboard can bifurcate two physical x16 slots

    Recommends · Jul 2026 · in “Nvidia DGX Spark as a daily driver”

40. More VRAM (some wasted due to it being non continuous), more RAM bandwidth, more aggregate compute.

    Recommends · Sep 2026 · in “Speculative Decoding in vLLM on AMD GPUs”