Bought 2 of these and they both are like new. Definitely recommend.

Probably the Best Investment You Can Make Right NowI’ve been deep into AI for years now, and at a certain point, my home system became less of a tool and more of a bottleneck-shaped paperweight. The time for an upgrade was long overdue.As luck would have it, this realization coincided with NVIDIA’s launch of the 5000 series. Naturally, I looked at the 4090 and the mythical 5090—until I saw the price tags.A 5090 (Nvidia’s current flagship card) has an MSRP of $1999 (lol, as if you’ll ever see it for that). After spending hours navigating a labyrinth of price-gouging, out-of-stock listings, and enough scalper nonsense to make concert ticket resellers jealous, I had a thought:”What ELSE could I buy with this money?”The list was long. A down payment on a used car. Three months of mortgage payments. A solid vacation. Or—craziest thought of all—a GPU that doesn’t require a second mortgage.So I passed on the latest highway robbery and picked up a 5070 Ti instead, assuming it would be a reasonable middle ground. After all, NVIDIA won’t stop hyping up the 5000 series as the ultimate AI powerhouses.They conveniently forgot one minor detail:The software isn’t ready.If you enjoy spending hours patching together a half-broken Frankenstein installation of PyTorch and CUDA—only for it to implode the moment you update anything—go ahead, grab a 5000 series card.If you’d rather just get work done, the 3090 is a no-brainer.For years, the 3090 has been a proven workhorse. No hoops to jump through, no compatibility nightmares—just raw power. Sure, it’s not the latest and greatest, but it’s also not a $5000 paperweight.I was hesitant about spending over $1000 on a 3+ year-old GPU with last-gen architecture… until I installed it.This GPU cooks.I’ve thrown everything at it—simultaneous instances of Stable Diffusion (SDXL) and Kobold, two massive 8GB+ AI models running side by side—and the card didn’t even break a sweat.The real kicker? I never had to troubleshoot a thing. I quite literally just slotted it into my PCI-e slot, booted up, and got to work.Since then, I’ve seen it crank out 90+ FPS in a benchmark test while also running an AI instance. With no other programs running, it easily pushes 160+ FPS.It might not be the most power-efficient card anymore, and it doesn’t have NVIDIA’s latest marketing gimmicks—but:✅ It works.✅ It won’t melt its power connectors.✅ And you can buy it for less than three months of mortgage payments.Unless you have a burning desire to be an unpaid beta tester for the 5000 series software stack, do yourself a favor—get a 3090, save a ton of money, and actually get stuff done.

About 1 month in and all cylinders are firing. I purchased this card for an AI project and I was sweating before I got it. I heard some horror stories on here about these refurb cards being recycled cards from crypto rigs and abused. Certainly wasnt the case thus far, the card arrived very new “looking”, couldnt find a grain of dust anywhere. The packaging was solid, although the original box serial does not match the card that came. The serial numbers are X out. But that seems to be common. I am running a 12 billion param LLM model on this guy and it barely breaks a sweat when you properly tune the model. Of the 24GB of vram, at model load it uses ~18gb with full stack riding the card, so plenty of headroom. Also the model I am running has a vision encoder, so you would think more opportunities for the card to run out of memory (OOM) when ingesting images, but nope, barely moves off that 18GB mark. When considering this or the newer RTX 5090 in the 24GB consumer class, I will take the extra $2k of savings and spend elsewhere, this card is a champ.