GPU Prices per GB of VRAM

No — and it's the most important caveat on this board. A discrete GPU has dedicated VRAM (GDDR6/GDDR6X/HBM) wired straight to the GPU die at very high bandwidth. Unified memory (Apple Silicon, NVIDIA GB10/DGX Spark, AMD Strix Halo) is one pool of RAM shared by the CPU and GPU. For fitting a large AI model it behaves like VRAM — a 128GB unified machine can load a model that would need multiple 24GB cards — but the usable memory bandwidth is typically lower than a high-end discrete card's, so tokens-per-second on the same model can be slower even though it fits. We rank these systems in $/GB because AI-inference buyers genuinely cross-shop them against stacks of GPUs, but every unified row is badged "unified" in the Specs column, and you can use the Memory filter to include or exclude them. Buy on bandwidth AND capacity, not capacity alone.

The floor moves, so check the live board above — it re-ranks hourly. As a rule of thumb from what we see: older high-VRAM cards and generous used discrete cards can dip toward the low-single-digit dollars per GB, while current-gen cards with fast memory carry a big premium per GB because you're also paying for compute and bandwidth. $/GB VRAM alone doesn't tell you if a card is fast — a cheap 16GB card and an expensive 16GB card have the same $/GB but wildly different throughput. Use the $/TFLOP alternate unit alongside it: $/GB tells you what fits, $/TFLOP tells you how fast it runs. The best buy is the card that clears both bars for your workload.

Because $/GB VRAM only measures capacity, not speed, and old cards are cheap precisely because their compute and memory bandwidth are dated. An 8GB RX 580 or a 12GB used Pascal card can post an eye-catching $/GB number, but they lack the tensor cores, memory bandwidth, and driver/software support (CUDA versions, ROCm coverage, FP16/FP8) that modern AI and gaming workloads assume. They're honest value for light workloads, a media/transcode box, or learning on a budget. They're a false economy if you actually need to train or run large models fast. This is exactly why we surface $/TFLOP next to $/GB — a rock-bottom $/GB with terrible $/TFLOP is the board telling you "cheap capacity, slow silicon."

For a specific use case, yes: running large models locally for inference where you need a lot of memory in one box, quietly, at low power. A Mac Studio with 96 to 128GB of unified memory can hold models that would otherwise need several discrete GPUs and a small power plant, and it does it in a silent desktop. Where it loses is raw training throughput and any CUDA-locked workflow — the discrete-GPU ecosystem (CUDA, the widest framework support, higher memory bandwidth on top cards) still wins for heavy training and maximum tokens/sec. On this board Apple systems show up as $/GB rows (badged "unified") so you can compare them head-to-head, but remember their prices are vendor list prices verified on the date shown, not live auction quotes, and their bandwidth is lower than a flagship card's. Decide on workload: memory-bound local inference leans Apple/unified; compute-bound training leans discrete NVIDIA.

There are two honest paths and the board shows both. Path one: a single unified-memory system — a 128GB Apple Silicon Mac, an NVIDIA DGX Spark (128GB), or an AMD Strix Halo machine — gets you 128GB in one pool, one power cord, one box, and on a $/GB basis these are often the cheapest route to that much addressable memory. Path two: stack discrete cards (e.g. multiple 24GB or 32GB cards) for far higher aggregate bandwidth and CUDA support, at the cost of more money, power, heat, and system-build complexity, plus the software work of splitting a model across GPUs. If you just need the model to FIT and run at a reasonable speed, unified is usually the cheapest and simplest 128GB+ option; if you need it to run FAST or you're CUDA-locked, budget for a multi-GPU build. Filter by Memory and sort by $/GB to see the current cheapest rows in each camp.