ailocalinference

← All hardware

RTX 4090

GPU only · new $3.5k on Amazon, used ~$2.4k

NVIDIAAdd-in GPU$3,499 · Aug 31, 2026View on Amazonpaid link

Memory
24 GB~22 GB usable
Bandwidth
1.0kGB/s — sets decode speed
Compute
83FP16 TFLOPS — sets prefill
Power
450 Wunder sustained load
Noise
45 dBAyou will hear it
Price per GB
$146of accelerator memory

Price history

tracking

Tracking since 2026-08-31 — not enough history for a trend yet.

1 observation recorded · Last checked 2026-08-31. Check the retailer before acting on this — we would rather say the data is old than pretend otherwise.

What it can run

2 of 6 reference workloads fit. The largest is 32B dense · Q4 (coding), at an estimated 34 tok/s of single-stream decode — #4 of the 17 machines here that can run it.

WorkloadNeedsFitsPrefill t/sDecode t/sBatched t/s$ per t/s
8B dense · Q4 (chat, drafts)
dense · 4.7 GB read per token
5.4 GBYes1.8k139723$25
32B dense · Q4 (coding)
dense · 19 GB read per token
21.8 GBYes45434177$103
70B dense · Q4 (strong general)
dense · 40 GB read per token
46 GBNo2071683—
120B MoE A5B · MXFP4 (gpt-oss class)
MoE · 2.9 GB read per token
72.4 GBNo2.8k146759—
320B MoE A18B · FP8 (frontier-ish)
MoE · 18 GB read per token
368 GBNo80724125—
671B MoE A37B · Q4 (DeepSeek class)
MoE · 18.5 GB read per token
437 GBNo39323120—

Figures are modelled from this machine's bandwidth and compute, not measured on one. How they are derived →

Alternatives that also run 32B dense · Q4 (coding)

MachinePriceMemoryDecode t/s$ per t/s
$65024 GB16$41
$1,46024 GB33$44
$1,40024 GB20$70
$4,40032 GB61$72
$1,70032 GB22$77
$1,39932 GB18$78
$1,30032 GB16$81
RTX 4090 — this machine
$3,49924 GB34$103
$3,89948 GB30$130
$4,500128 GB21$214
$2,19948 GB9$244
$2,76064 GB11$251
$15,49996 GB61$254
$13,500512 GB41$329
$9,499512 GB28$339
$3,999128 GB9$444
$5,199128 GB9$578