GPU Value Rankings: Which Card Is Actually Worth Your Money for Local LLMs?

Most "best GPU" lists just repeat spec sheets. These rankings are different: every number below is computed at build time from our own database of 12 consumer, workstation, and Apple unified-memory GPUs. The methodology:

• Price: used-market price first; cards without used data (RTX 5090, A100, H100, Apple machines) use MSRP or a market reference. Used prices follow eBay sold/listed listings and are updated monthly.
• Speed: a theoretical estimate — bandwidth × 0.75 ÷ bytes of weights read per token. Real-world results vary with framework, drivers, and CPU; expect roughly ±30%.
• VRAM: Apple unified memory is counted at 75% usable (it is shared with the OS and apps); discrete GPUs use their rated capacity.
• Data verified: 2026-09-01.

The RX 7900 XTX takes first place at roughly $650 used — 1.48 GB/s per dollar makes it the bandwidth sweet spot. The RTX 3060 12GB holds second at 1.39 even after the DRAM shortage pushed its used price up ~40% (from $170-220 in July to $250-275, with NVIDIA relaunching it at retail for $339). Even a price hike couldn't dethrone it.

The RTX 5090 ranks third at its $1,999 MSRP — but with secondary-market prices at about 2.1× MSRP ($4,288-4,381 new), it drops out of the top spots at real street prices. It's only a deal if you can actually buy it at MSRP. The A100 and H100 at the bottom prove one thing: datacenter bandwidth costs 10× more per unit. Individuals running inference shouldn't look up to them.

The RTX 3060 12GB leads by a wide margin at 4.6GB per $100 — which is exactly why it has topped entry-level recommendations for years. The RX 7900 XTX is second at 3.7GB; a 24GB card that launched at $999 and now sells for $650 used remains AMD's classic capacity-per-dollar play.

The RTX 3090 takes third at 2.2GB/$100 despite a persistent 24GB premium ($1,095-1,181 used), and it is the workhorse of budget multi-GPU 70B builds. The adjusted Mac mini M4 Pro 48GB (36GB usable) breaks into the top five — if you want a quiet, low-power box that swallows 30B-class models, it is the only answer at $1,799.

For 8B-class models, the RX 7900 XTX and RTX 3060 12GB are nearly tied on speed per dollar (22.5 vs 21.1 tok/s per $100) — the sweet spots of the $650 and $260 budgets respectively. The former is almost three times as fast; the latter costs a fraction. The RTX 5090 is third at 13.7: its absolute 273 tok/s is the fastest of any consumer card here, but the price dilutes the win.

The cautionary tale is the RTX 4090: discontinued, with used prices anchored to the 5090 and climbing to $2,350-2,500. Its speed per dollar is a third of the 3060's — buying one just for 8B models means every dollar goes to the discontinuation premium.

The cheapest ticket into 70B territory is 4× RTX 3060 12GB ($1,040) — but it's a tight fit, and four cards bring hidden costs in power delivery, cooling, and PCIe slots. The real sweet spot is 2× RX 7900 XTX ($1,300): 48GB across two cards for only $260 more, with a fraction of the engineering pain. A step up, 2× RTX 3090 ($2,200) buys more bandwidth and a smoother CUDA ecosystem.

Prefer a single card? The RTX A6000 48GB ($2,800) is the cheapest single-card answer (tight); 2× RTX 5090 ($3,998) is the lowest-priced "comfortable" setup; and the Mac Studio M3 Ultra 96GB ($5,299) is the only single desktop machine that runs 70B comfortably — at the cost of a ~125 tok/s theoretical ceiling. Silent and power-sipping, but not fast.