NVIDIA L4

Single-slot, 72W, no external power connector — designed to drop into dense inference servers. Low memory bandwidth (300 GB/s) relative to price/VRAM makes it a poor decode-speed choice versus an A5000/A6000 despite being newer; its value proposition is power/density, not raw tok/s. No sourced llama.cpp or vLLM decode-speed benchmark found this session (see gaps file). New-old-stock 'new' figure of $2,750 (Newegg) is now available via the same gpudojo.com tracker -- notably BELOW the $3,500 used/eBay figure, which is unusual and likely reflects thin/noisy used listings for this still-in-production card rather than a stable secondary-market premium; treat the used figure with caution.