GoKawiilThe AI boom means that no matrix math FLOPS are disposable, and that means older Nvidia GPUs with Tensor Cores are getting a new lease on life. Services are popping up that will outfit your RTX 2080 Ti with 22GB of VRAM, doubling its original memory pool and making it more useful for modern LLM and diffusion workloads. If you’re hard up for compute and you don't have an RTX 2080 Ti to modify, however, eBay has just the thing. A Hong Kong-based seller will send you a pre-modded 22GB 2080 Ti in exchange for $499.
The listing only promises that you'll receive a "Turbo," i.e., blower-style, RTX 2080 Ti with the 22GB mod. Don't get too picky about the particular brand of card you might receive, as the seller says: "The GPU brand could be Gigabyte, MSI, ASUS, Leadtek or others. It depends on what we have in hand."
The eBay listing photos show a Gigabyte blower-style card along with a large box of presumably modified cards ready to ship and a GPU-Z screenshot confirming the availability of 22528 MB of memory from a running card.
Latest Videos From Tom's Hardware Watch full video here:
The seller only has 99.6% lifetime feedback on eBay, which is relatively low, but feedback for this specific listing shows that at least 38 happy buyers have received these cards and that they’re working as described.
Assuming you can count yourself as one of those happy buyers, a $500 22GB RTX 2080 Ti might be the best VRAM bang for your buck that you can find these days. That relatively large memory pool, combined with the RTX 2080 Ti's 616 GB/s of memory bandwidth and Tensor Cores, means that it's still useful for local LLM tasks, although its raw compute capacity and limited reduced-precision data type support compared to more modern products might make demanding diffusion workloads leisurely.
Among Turing cards, the 24GB Titan RTX still commands about $800 on eBay at today's prices, and a Quadro RTX 6000 with the same amount of VRAM is about $900. The local AI enthusiast's favorite RTX 3090, which has 24GB of faster GDDR6X offering 936 GB/s of memory bandwidth, looks to be selling for about $1200. Ampere RTX Pro cards with even more VRAM rapidly get more expensive from there.
The fact that all of these GPUs are still commanding such high prices many years after their introduction is a testament to the continuing utility and universal availability of Nvidia’s Tensor Core architecture across both consumer and data center products since 2018.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter Get Tom's Hardware's best news and in-depth reviews, straight to your inbox. Contact me with news and offers from other Future brands Receive email from us on behalf of our trusted partners or sponsors
AMD has had matrix math accelerators in its IP arsenal since CDNA 1 in 2020, but their availability has been limited to Instinct data center products until RDNA 4 arrived early last year. Intel included XMX matrix engines in the Alchemist architecture from the start in late 2022, but Arc graphics products have faced considerable challenges beyond the presence or absence of that capability. And Apple only just introduced Neural Accelerators to its GPUs with the M5 family.
... continue reading