I’m putting together a local AI workstation under a $1500 GPU budget, and I’ve been comparing used listings at my kitchen table while checking case clearance with a tape measure. I expected memory capacity to narrow the choices quickly, but the tradeoffs look less obvious than expected. Are others finding the same thing, and what should I prioritize?
Buy VRAM first.
For an LLM-focused machine, a used RTX 3090 is still the sensible baseline. Its 24GB gives you much more room for larger quantized models, longer context, and fine-tuning experiments than a newer 16GB card. If you find a trustworthy RTX 4090 within your limit, it will be considerably faster, but it still has 24GB, so it does not expand what fits. I would not choose an RTX 5080 for this specific job unless image generation or gaming matters more than model capacity. The 5080 has 16GB despite being a newer design.
Two 3090s can be attractive if the GPU allowance covers both, but treat that as a more complicated build rather than a single 48GB card. Your software has to split the model across the GPUs, and some workloads handle that better than others. You also inherit motherboard-spacing problems, a large PSU requirement, heat, noise, and reduced airflow. Current model frameworks support sharding across devices, but it requires the right configuration and may add transfer overhead.
Since you are measuring the case, check width and cable clearance as carefully as length. A card can technically fit between the front fans and rear slots while leaving no safe bend for the power cable. NVIDIA’s 4090 Founders Edition guidance calls for extra room beyond the card itself for cabling, and partner cards can be larger.
My priority order would be VRAM capacity, CUDA support, card condition, cooling and power, then raw speed. For a used 3090, insist on a return window and test sustained memory-heavy loads, temperatures, fans, and power connectors immediately. A clean single 3090 with a solid PSU is a much less annoying starting point than forcing two cards into a marginal case. Save the remaining money for system RAM, storage, and power delivery rather than paying extra for a newer GPU that leaves you stuck at 16GB.
The hidden downside of chasing the fastest GeForce is ending up with a very quick out-of-memory error. If a used RTX A6000 appears within your budget, I’d inspect that before the gaming cards. You get twice the memory of the common Ampere option on one device, which avoids model sharding and makes larger LLMs much less awkward. The tradeoff is lower generation speed than newer gaming hardware, plus blower noise.
If no clean A6000 listing turns up, @alphabear6747node’s basic direction holds: get a single used GeForce with 24GB and keep the build simple. I would skip dual GPUs unless you already know your chosen software handles them properly. A single large-memory card saves a lot of fiddling with layer placement and apps that insist on using only the first device.
Realistically, this budget points to a used RTX 3090. An A6000 priced that low would make me suspicious of a mislabeled card, damaged hardware, or a scam listing. Buy locally if possible, run a memory-heavy test before paying, and check hotspot temperatures. Saving some cash for a strong PSU and better airflow is smarter than spending every dollar on the GPU.
A card’s advertised memory is not the amount your model actually gets to use. Context cache, the inference backend, and whatever is driving your desktop all take a slice. A model that appears to fit on paper can still fall over when you raise the context length, load a second model, or enable an image encoder. Leave some headroom instead of shopping right at the model’s quoted requirement.
For that reason, I would still buy the previous-generation Ampere flagship rather than a newer mid-memory card, but I would set it up as a dedicated compute device. If your CPU has integrated graphics, connect the monitor to the motherboard and keep desktop applications off the main GPU. Check that your firmware supports this arrangement before assuming it will work. A cheap display card can do the same job, though it may steal a slot or block airflow.
The other overlooked factor is software compatibility. Decide what you actually plan to run, then verify that its backend supports your intended quantization and GPU splitting. Some applications handle partial CPU offload gracefully. Others become painfully slow or simply refuse models that do not fit entirely in VRAM. Extra system memory is useful as an escape hatch, but it is not a substitute for GPU memory if you care about generation speed.
I would avoid building around two cards unless you have a specific model and application in mind. The attractive combined-memory number hides uneven support, inter-GPU traffic, and idle power. A single used compute card with plenty of memory, run with a modest power limit, is usually the workstation you will actually use rather than the project you keep troubleshooting. Spend the remaining budget on enough system RAM to avoid swapping and an SSD large enough that downloading several quantizations does not turn into constant file cleanup.
If a second GPU is even a vague future plan, choose the motherboard before buying the first card.
Two full-length slots do not mean two useful GPU slots. The lower slot may run through the chipset at x4, lose lanes when certain M.2 sockets are populated, or sit directly under a three-slot cooler. Read the board’s lane-sharing table. Product photos are decorative fiction.
For a permanent single-card build, the big-memory used GeForce is the obvious answer here. Buy it, test it, and move on. For a planned dual-card build, budget for the board, chassis, riser layout, and power supply as one package. Otherwise that “easy upgrade later” becomes a replacement of everything except the GPU.
Nobody’s said how hot those used 3090s actually get on the memory side, and that’s the thing that bites people after purchase. The GDDR6X on those cards runs warm by design, and a lot of the cheap listings came out of mining rigs where the thermal pads are cooked. The core temperature can look fine in a quick test while the memory junction is sitting near throttle under a sustained load. If you’re doing LLM work, memory is exactly what you’ll be hammering. So when @novawizard7308 says run a memory-heavy test before paying, I’d push that further: watch the memory junction temp specifically, not just the GPU core. If it’s climbing past the mid 90s quickly, factor in a repad job or walk away.
On the dual-card talk, I think @greentiger nailed the real problem better than the sharding worries did. The spacing and lane-sharing stuff is where these builds quietly fall apart. Two cards jammed together with no gap means the top one breathes the bottom one’s exhaust, and your nice combined memory number turns into a thermal mess that throttles anyway. For a kitchen-table build in a normal room, I’d honestly drop the two-card idea entirely unless you’re ready to deal with noise and a space heater.
The part I’d weigh more heavily than the thread has: what the card does to your room and your power bill. A single 3090 under a modest power limit is genuinely fine, and limiting it costs you very little speed on inference while dropping heat and fan noise a lot. If you’re sitting next to this thing while it runs, that matters more than the last few percent of tokens per second. Set a power cap early, keep the desktop off the compute card like @stacklab410 suggested, and you end up with something you’ll actually leave running instead of a machine you avoid turning on because it roasts the room.
For your budget the single 24GB used card is the sane call. Just treat the memory temps as the real gatekeeper, not the sticker on the box.
You may end up happier spending less than the full allowance if your everyday workload is smaller than the models you’re shopping around. I’d tentatively favor the used high-memory GeForce route here, but I’m cautious about treating “fits the largest model” as the whole buying decision.
Before paying, I’d write down a small acceptance test: the model you actually want, a typical prompt or document, and how long you’re willing to wait for a useful answer. Then ask for a demonstration or look for a benchmark that resembles that task. For document work, I’d want to see the wait before the answer starts, not only the speed once text is appearing. For an interactive assistant, I’d care whether the pauses feel tolerable. A seller’s impressive result on a tiny prompt wouldn’t settle either question for me.
That’s where I’d qualify @alphabear6747node’s “buy VRAM first” advice. It’s a useful starting filter, but I wouldn’t pay a large premium for capacity reserved for a vague future experiment. I’d choose the least expensive candidate that passes my actual test with some room to spare. If none does, I’d rather revise the workload before buying than discover afterward that my definition of “usable” was different from the benchmark’s.