The 24 GB build

Verdict

For a builder who wants 30B-class models on one card now and a machine that becomes the 48 GB build by adding a second used RTX 3090 later. 24 GB holds Qwen2.5 32B at Q4_K_M with an 8k context, or the Qwen3 30B-A3B mixture of experts at Q4_K_M with 32k. One used 3090 is the cheapest path to 24 GB on one card; the rest of the parts list is the 48 GB platform unchanged, so the board, PSU and case cost more than a single card needs. The card is rated at 350 W; power-limited to 300 W, this site's estimate puts the machine at about 611 W at the wall while generating.

Model fit

Parts list

Store links may earn us a commission at no cost to you. As an Amazon Associate we earn from qualifying purchases.

Basis

Compatibility warnings

• Length and width. The Founders Edition RTX 3090 is 313 mm long and three slots wide. The Define 7 XL takes cards up to 549 mm in its open layout and 359 mm in its storage layout, and the Phanteks alternate takes 503 mm, so length is not a problem in either. A partner 3090 or 4090 can be longer or thicker than the Founders Edition; check the exact card. One thicker than three slots also closes the air gap to a second card in PCIE2.
• Which slot. Put the card in PCIE1, the Taichi’s top x16 slot. ASRock’s manual gives the two CPU-attached slots x16 or x8/x8 modes, so with one card PCIE1 has all 16 lanes (PCIe 4.0 on the 3090, which is a PCIe 4.0 card). The PCIe lanes and risers guide covers slot wiring and spacing in more depth. On the Taichi the M2_1 slot sits between the socket and PCIE1 (manual layout, PDF page 11), so the card is not directly under the NH-D15 G2.
• Power connectors. NVIDIA’s Founders Edition takes two 8-pin PCIe cables through the 12-pin adapter in the box. Run a separate cable from the PSU to each 8-pin input; don’t feed both from one cable’s daisy-chained pigtail. The HX1200i lists five 8-pin PCIe connectors, so one card uses two of them. The 4090 alternate needs three 8-pin cables to its adapter (NVIDIA says it supports three or four) or one 450 W PCIe Gen 5 cable.
• Transient spikes and power-limiting. NVIDIA rates the RTX 3090 at 350 W and asks for a 750 W PSU for a single-card system; for the 4090 it is 450 W and 850 W. Igor’s Lab warns that the 3090’s short load peaks, at intervals of 1 to 10 ms, can trip a PSU’s protection, and hav4ik reports his 3090s’ “power draw sometimes spikes up to 600W for brief periods”. With one card the HX1200i has room to spare: 350 W + 250 W for the rest of the system is 600 W against its 960 W at 80 %. Power-limiting isn’t required for one card, but set it now anyway, because the second card will need it and the Power & noise estimate below assumes it: nvidia-smi -pl 300 (it needs root). The limit does not survive a reboot or a driver reload, so reapply it at boot: Puget Systems does it with a systemd service that turns on persistence mode (nvidia-smi -pm 1) and then sets the limit. A power limit caps average draw; it does not cap the millisecond transients. The PSU sizing guide works through both.
• Ready for the second card. The slot layout and the PSU were chosen for two 3090s. Measured to scale from the layout drawing in ASRock’s manual, PCIE1 and PCIE2 sit about 80 mm apart, four slot positions, so two three-slot Founders Editions seat directly with one empty position of air between them; with both filled the slots run x8/x8 from the CPU. This site’s PSU estimate gives the HX1200i two GPUs: 1200 W × 0.8, minus 250 W for the rest of the system, divided by 300 W per power-limited GPU, rounded down. The Define 7 XL has nine expansion slots and takes E-ATX boards, so there are openings below the board for the second card, which overhangs it. The 48 GB build covers that second card in detail.
• With the 4090 alternate, redo the arithmetic before adding a card. A 4090 and a 3090 unlimited come to 450 W + 350 W + 250 W = 1050 W, over the HX1200i’s 960 W at 80 %. They also need five 8-pin cables between them (three plus two), all five the HX1200i lists. Power-limit both, or take the HX1500i, the 48 GB build’s PSU alternate, which this site’s estimate gives three GPUs (1500 W × 0.8 − 250 W = 950 W, ÷ 300 W). Corsair rates it at 1200 W on 100 to 115 V input, so on that mains it gives no extra headroom.
• Display output. The Ryzen 9 9950X has integrated AMD Radeon graphics, and the Taichi has one HDMI output and two USB4 ports for it. Plug the monitor in there and keep the 3090’s 24 GB for models; the Model fit margins assume no desktop on the card. The headless setup guide covers running the box without a monitor.
• Household circuit. A US 15 A circuit at 120 V carries 1800 W, and a continuous load (one running more than three hours) should stay under 80 % of that, which is 1440 W. This Tier’s estimated 611 W at the wall while generating fits with room to spare; unlimited, (350 W + 250 W) ÷ 0.9 is about 667 W, which fits too. On 230 V mains a UK plug takes at most a 13 A fuse, about 3.0 kW, and a German Schuko socket is rated 16 A, about 3.7 kW. The second card in the 48 GB build is where the circuit starts to matter.
• Card memory temperature. Half of the 3090’s GDDR6X sits on the back of the board, cooled only by the backplate. Watch its memory junction temperature under sustained load as the used RTX 3090 checklist describes.
• CPU cooler. The NH-D15 G2 is 168 mm tall; Fractal’s product sheet gives the Define 7 XL 185 mm of cooler height, and Phanteks lists 195 mm for the Enthoo Pro 2 Server Edition. Noctua’s compatibility list for the X870E Taichi includes the NH-D15 G2 with no extra mounting kit and no top-PCIe-slot restriction noted. Check the RAM kit’s height against Noctua’s NH-D15 G2 specification before ordering.

Power & noise

Estimated 24/7 energy use, with every assumption stated:

• Load: the card power-limited to 300 W (the per-GPU figure in this site’s PSU estimate), plus 250 W for the rest of the system (the same estimate’s system reserve). That gives 300 W + 250 W = 550 W delivered by the PSU while generating.
• Idle: Igor’s Lab measured about 16 W for one RTX 3090 at idle. The rest of the system at idle is an assumption, not a measurement: 70 W for the board, CPU, RAM, SSD and fans. That gives about 86 W delivered by the PSU at idle.
• At the wall: the HX1200i is 80+ Platinum, which requires 89 to 92 % efficiency between 20 % and 100 % load on 115 V. This estimate assumes 90 %, so divide by 0.9: about 611 W at load and 96 W at idle. At idle the PSU runs at about 7 % load (86 W of 1200 W), below the 20 % where Platinum’s figures start, so the real idle draw is probably somewhat higher.
• Duty cycle: an assumption for a personal inference box: 4 hours of generation and 20 hours idle each day.

Arithmetic: 4 h × 0.611 kW ≈ 2.4 kWh, plus 20 h × 0.096 kW ≈ 1.9 kWh, gives about 4.4 kWh a day at the wall. That is about 132 kWh a month (× 30.4 days) and about 1,590 kWh a year (× 365). Nearly half of it is idle draw, which a platform sized for two cards carries even with one; the duty cycle moves the result far more than any part choice here. Multiply by your own tariff. The power and noise guide explains the method and its weakest inputs.

Noise: the 3090 is the loud part, and none of the Basis builds measured noise. With one card, nothing below it blows hot air into it, which hav4ik names as the reason closely stacked 30-series cards overheat; skeptric’s 3090 stayed under about 75 °C in a benchmark and did not pass 71 °C during training, in a Corsair 4000D Airflow with four extra 120 mm fans. Power-limiting to 300 W lowers both heat and fan speed. The Define 7 XL is sound-damped, but the damping also holds heat, so fill the front fan mounts. timroy keeps his machine in the basement rather than listen to it in his office, which is the other answer.

FAQ

Has anyone built exactly this list? Not that we know of. The three Basis builds share its core: one used RTX 3090 in a desktop, run for machine learning. Their platforms differ. skeptric’s is a Ryzen 5 5600G on a B550 board with an 850 W PSU in a Corsair 4000D Airflow. timroy’s is a Ryzen 9 5900X on an ASUS Crosshair VIII Hero with a 1200 W PSU in the original Phanteks Enthoo Pro. biagioantonelli’s is a Core i7-13700K on an ASUS Strix Z790-F with a 1300 W PSU in a Lian Li O11 Dynamic. None reports a tokens-per-second figure for a language model, so every Model fit row is our arithmetic.

Why build for two cards when I’m buying one? timroy, one of the Basis builders, planned for it: he chose his 1200 W PSU partly for a second GPU and expects to need a second 24 GB card for fine-tuning; his board drops its first slot to x8 with two cards installed, and the second slot is only half wired. This Tier’s Taichi runs both slots at x8 from the CPU instead. Buying the two-card PSU and board now means the upgrade is a second 3090 and nothing else.

How was the fit worked out? The VRAM sizing guide explains the method. The card holds 24,576 MiB, about 25.77 GB. Qwen2.5 32B has 64 layers, 8 key/value heads and a head size of 128 (5,120 ÷ 40 heads). A 16-bit KV cache, llama.cpp’s default, therefore takes 64 × 8 × 128 × 2 (keys and values) × 2 bytes = 262,144 bytes per token. At 8,192 tokens that is about 2.15 GB; with the 19.85 GB Q4_K_M file, about 22.0 GB, leaving roughly 3.8 GB for the CUDA context and compute buffers. At 16k the cache is 4.29 GB and the margin falls to about 1.6 GB. Qwen3 30B-A3B (48 layers, 4 key/value heads, head size 128) takes 98,304 bytes per token, about 3.22 GB at 32k. Qwen2.5 14B (48 layers, 8 key/value heads, head size 128) takes 196,608 bytes per token, about 6.44 GB at 32k.

Why the 3090 and not the 4090 alternate? Both have 24 GB, so they fit the same models. The 4090 has more memory bandwidth (1,008 GB/s against 936 GB/s) and much more compute, so it processes prompts faster. It costs much more than a used 3090, draws 450 W against 350 W, and the 48 GB build is two 3090s, so the 3090 is the card that carries over.

What’s the upgrade path? The 48 GB build: add a second used RTX 3090 in PCIE2, set the same 300 W limit on both, and you have that Tier’s main parts list. 48 GB fits Llama 3.3 70B at Q4_K_M with an 8k context, which 24 GB cannot hold: its 42.52 GB file alone is well over one card.