Desk-side LLM iron

Pick a model, see which desk-side machines can actually hold it and how fast they decode. Throughput is derived from memory bandwidth and compute, not typed in by hand — the formula is published and applied identically to all 17 machines. Click a column header to sort.

Prices captured Aug 31, 2026

Machines that fit 18B active / 320 GB2 of 17

Cheapest that fitsMac Studio M3 Ultra · $9,499

Fastest decode that fitsMac Studio M5 Ultra · 28 t/s

Best $ per t/s ♛ crowned belowMac Studio M5 Ultra · $482

Model

Machine	Price	Memory	BW GB/s	Prefill t/s	Decode t/s ↓	Batched t/s	t/s per 100W	Fits	$ per t/s	dBA
RTX PRO 6000 Blackwell GPU only · GDDR7 · $8.5k launch, $15,499 on Amazon today	$15,499 $161/GB	96 GB ~88 usable	1.8k	1.2k	42	218	7	No	—	45
RTX 5090 GPU only · median street $4.4–4.7k vs $1,999 MSRP	$4,400 $138/GB	32 GB ~29 usable	1.8k	1.0k	42	218	7.3	No	—	45
Mac Studio M5 Ultra♛ Complete box · $5,499/96GB · $9,499/256GB · 512GB late Oct, price TBD (~$13.5k est)	$13,500 $26/GB	512 GB ~384 usable	1.2k	250	28	146	10.4	Yes	$482	35
RTX 4090 GPU only · new $3.5k on Amazon, used ~$2.4k	$3,499 $146/GB	24 GB ~22 usable	1.0k	807	24	125	5.3	No	—	45
RX 7900 XTX GPU only · new $1.46k on Amazon, used ~$800	$1,460 $61/GB	24 GB ~22 usable	960	169	22	110	6.2	No	—	45
Radeon PRO W7900 GPU only · 48GB, blower, ROCm	$3,899 $81/GB	48 GB ~44 usable	864	169	20	104	6.8	No	—	40
Mac Studio M3 Ultra Complete box · 512GB config, now prior-gen	$9,499 $19/GB	512 GB ~384 usable	819	117	19	78	8.6	Yes	$500	40
AMD Radeon AI PRO R9700 GPU only · 32GB workstation card	$1,700 $53/GB	32 GB ~29 usable	640	133	15	78	5	No	—	40
MacBook Pro M5 Max Laptop · 16-inch, high memory config	$4,500 $35/GB	128 GB ~96 usable	614	125	14	73	10.8	No	—	45
Tenstorrent Wormhole n300 Add-in card · immature software stack	$1,400 $58/GB	24 GB ~22 usable	576	167	13	68	4.3	No	—	50
Tenstorrent Blackhole p150 Add-in card · immature software stack	$1,399 $44/GB	32 GB ~29 usable	512	278	12	62	4	No	—	50
Intel Arc Pro B70 GPU only · 32GB Xe2-HPG workstation card	$1,300 $41/GB	32 GB ~29 usable	456	67	11	45	5.8	No	—	45
Arc Pro B60 GPU only · 24GB — cheapest capacity per dollar here	$650 $27/GB	24 GB ~22 usable	456	56	11	38	9.2	No	—	45
MacBook Pro M5 Pro Laptop · 18-core CPU / 20-core GPU config	$2,760 $43/GB	64 GB ~48 usable	307	63	7	36	7.8	No	—	45
DGX Spark Complete box · $2,999 announced, $5,199 on Amazon (4TB)	$5,199 $41/GB	128 GB ~118 usable	273	86	6	31	4.3	No	—	35
ASUS Ascent GX10 Complete box · same GB10 superchip as DGX Spark, $1.2k less	$3,999 $31/GB	128 GB ~118 usable	273	86	6	31	4.3	No	—	35
Mac mini M4 Pro Complete box · 48GB/512GB, the config Amazon stocks	$2,199 $46/GB	48 GB ~36 usable	273	33	6	22	7.5	No	—	30

Reading the table

Decode t/sSingle-stream generation — the speed you feel. Memory-bandwidth-bound, so it tracks GB/s far more than TFLOPS.

Prefill t/sPrompt processing. Compute-bound, and heavily dependent on how mature the vendor's software stack is.

FitsWhether the weights plus KV-cache headroom clear usable memory. A "No" means the model will not load, whatever the other columns say.

$ per t/sPrice divided by decode speed — the honest value column. ♛ marks the lowest of any machine that fits the selected model.