Sep 25, 2026

The DDR4 Escape Hatch: Building a Used RTX 3090 Rig on AM4 While the 2026 RAM Crisis Prices Out DDR5

Q: Should I wait out the whole crisis instead?

For RAM-heavy builds, waiting is a real option — [we've said so since September](/blog/ddr5-ssd-price-surge-ai-hbm-impact-local-builds-2026/). But this build is deliberately RAM-light ($190–$263 of exposure), and used 3090s went from roughly $1,010 in March to ~$1,250 in September — about 24% in six months. Here, waiting costs more than it saves.

By RunAIHome Team · 12 min read

gpulocal-llmbudget-buildddr4am4rtx-3090ram-crisis

TL;DR: DDR5 has quadrupled since summer 2025 — a 32GB kit now runs $399–$479 — while DDR4-3200 32GB still sells for $190–$263. Since local LLM speed lives on the GPU, an AM4 board plus a used RTX 3090 skips roughly $300 of the crisis with zero tokens/sec penalty. The catch: Samsung, SK Hynix, and Micron have all ended DDR4 production, so this build only exists while existing stock lasts.

	AM4 + used RTX 3090	AM5 + used RTX 3090	Strix Halo mini PC
Best for	Cheapest path to 24GB VRAM today	Same rig with an upgrade path	128GB capacity, no DIMMs to buy
Total (Sep 2026)	~$2,120–$2,480	~$2,440–$2,760	$2,199 (GMKtec EVO-X2 128GB)
tok/s on 20B-class models	161 tok/s (gpt-oss-20b)	161 tok/s — identical	~55 tok/s, but runs 120B models
The catch	Dead socket, discontinued RAM	+$300 for the same inference speed	3–4× slower on anything that fits in 24GB

Honest take: If you’re buying a used RTX 3090 anyway, put it on a $430–$580 AM4 platform and spend the ~$300 you save on nothing — keep it. The AM5 upgrade path is worth $300 only if you’ll actually drop a Zen 6 CPU in later, and most inference boxes never get a CPU upgrade.

The DRAM price crisis is no longer news — DDR5 pricing has been broken since summer 2025 and no forecaster expects relief before late 2027. What has changed is that the standard advice (“just pay up for DDR5, it’s the current platform”) now costs more than a used graphics card. A 32GB DDR5-6000 kit runs $399–$479 in September 2026, per Tom’s Hardware’s RAM price index. The 64GB kits the old guides recommended are $680–$1,070.

DDR4 got dragged up by the same shortage — TechPowerUp tracked a Corsair Vengeance LPX 32GB DDR4-3200 kit going from $71.99 to $262.99 in six months — but it’s still roughly half the price of DDR5 at the same capacity: $190–$263 for a 32GB kit as of this week. For a machine whose entire job is feeding a 24GB GPU, that gap is the last meaningful discount left in the DIY market.

This is a time-limited situation, not a stable equilibrium, and it’s worth being precise about why.

DDR4 is discontinued. That cuts both ways.

All three DRAM manufacturers have formally ended DDR4 production. Samsung stopped taking DDR4 orders in mid-2025 and wrapped final shipments by December 2025; Micron issued its end-of-life notice in June 2025 with final shipments in Q1 2026; SK Hynix — the last holdout — completed its exit in April 2026, per TrendForce’s exit-timeline reporting. Everything on shelves today is existing inventory plus a trickle from smaller Taiwanese fabs.

Two consequences for this build:

The discount can vanish. DDR4 chip spot prices already inverted above DDR5 once in late 2025 when supply first tightened. Retail kit pricing hasn’t inverted — the Tom’s Hardware index still shows DDR4 kits at roughly half DDR5 — but a discontinued part with steady demand from the enormous installed AM4/LGA1700 base only moves one direction long-term.
Buy your final capacity on day one. The classic play of “start with 32GB, add another kit next year” is dead on this platform. If your workload needs 64GB (mostly 70B-class CPU offload), price the 64GB kit now — and at DDR4’s $380–$520 for 64GB, the AM4 discount shrinks enough that you should reconsider the whole plan. This build is for people whose models live in VRAM.

Why the platform doesn’t touch your tokens/sec

The uncomfortable secret of every LLM build guide: once the model is resident in VRAM, the CPU, chipset, and system RAM are bystanders during decode. Token generation is bound by GPU memory bandwidth — the RTX 3090’s 936 GB/s — and system RAM at 51 GB/s (DDR4-3200 dual-channel) or 96 GB/s (DDR5-6000 dual-channel) never enters the loop.

The numbers this site has verified on a used RTX 3090 hold on any platform that gives the card a PCIe 4.0 x16 slot:

Model (fits in 24GB)	RTX 3090 decode speed
7–8B Q4	~95 tok/s
gpt-oss-20b (MXFP4)	161 tok/s
27B dense Q4_K_M	~40 tok/s
Qwen3.6-35B-A3B (MoE, ~22GB)	107 tok/s

Run those on a $141 Ryzen 5 5600 or a $485 Ryzen 9 9950X and the decode column doesn’t move. We made the same point in the $3K/$6K/$10K tier guide: budget buys VRAM capacity, not speed. The platform’s only jobs are a full-bandwidth slot, enough RAM to stage the model file, and not crashing.

Where the platform does matter is the failure mode: a model that doesn’t fit in 24GB and spills to CPU. Llama 3.3 70B Q4 on a single 3090 offloads to system RAM and crawls at 2–5 tok/s on DDR5; DDR4’s ~half bandwidth drags that toward 1–3 tok/s (bandwidth-scaled estimate — we haven’t bench-tested the DDR4 case). Neither is usable daily. If 70B residency is the actual goal, this is the wrong article — you want the dual-3090 $5,000 build.

The parts list

Prices verified September 25, 2026. Every number has a source at the bottom.

Part	Pick	Price
GPU	Used RTX 3090 24GB	$1,150–$1,350
CPU	Ryzen 5 5600 (includes cooler)	$141–$161
Motherboard	MSI MAG B550 Tomahawk (or any B550 ATX)	$96–$160
RAM	Corsair Vengeance LPX 32GB DDR4-3200 (2×16GB)	$190–$263
PSU	NZXT C850 850W ATX 3.1	~$100
SSD	Samsung 990 Pro 2TB	~$338
Case	Lian Li Lancool 216 (or any two-fan ATX case)	~$103
Total		~$2,120–$2,480

Notes that earn their place:

The 3090 is the build. It’s 55–60% of the total, which is the correct shape — our long-standing position on the used 3090 is that 24GB at 936 GB/s remains the cheapest real VRAM in the hobby. September street prices are $1,150–$1,350 per the BestValueGPU and ResalePrices trackers; buy from a seller with returns and test with a memory-stress load inside the window.
B550, not B450. B550 gives the primary slot PCIe 4.0 x16 straight from the CPU — the 3090’s native interface. B450/X470 boards run the card at PCIe 3.0; inference decode barely notices (the bus mostly idles during generation), but model loads slow down and you’re one BIOS quirk away from Resizable BAR headaches. The $30 between a used B450 and a B550 is not where to save.
The 5600 is enough CPU. Six Zen 3 cores at 65W keep llama.cpp’s sampling and the OS out of the GPU’s way. If you also run agentic coding workloads that hammer tool calls on the CPU, step up to an 8-core 5700X; past that, AM4 CPU spending buys nothing for inference.
850W is deliberate headroom. A 3090 spikes hard (transient draw well above its 350W rating). 850W ATX 3.1 units handle the excursions and leave room to power-limit experiments — see our PSU sizing guide. The NZXT C850’s $99.99 street price makes this the cheapest line on the sheet.
The SSD hurts and there’s no fix. NAND rode the same shortage up; ~$338 for a 990 Pro 2TB is September reality (verified for the $5K build last week). A 40GB model loads in under 15 seconds on Gen 4 NVMe versus 70+ seconds on SATA — our SSD guide has the tier math. Don’t downgrade the drive to save $60 and pay for it every model swap.

Check your target model actually fits in 24GB before buying anything — the VRAM calculator does model + quant + context math in ten seconds.

The same rig on AM5 costs ~$300 more. What does that buy?

Swap the platform trio and keep everything else:

Platform piece	AM4	AM5
CPU + board	5600 + B550: $237–$321	Ryzen 5 7600X + B650 bundle: $349–$390 (Micro Center bundles, sources below)
32GB RAM	DDR4-3200: $190–$263	DDR5-6000: $399–$479
Platform total	$427–$584	$748–$869

The ~$300 delta buys, concretely: a live socket (Zen 6 lands on AM5; AM4’s last CPU shipped in 2022), DDR5’s doubled offload bandwidth for models that spill out of VRAM, PCIe 5.0 you cannot use (the 3090 is PCIe 4.0, and no new consumer GPU generation is coming in 2026 to change that), and RAM you can still buy new in 2028.

That last item is the honest argument for AM5, and it’s stronger than the CPU-upgrade one. An inference box rarely gets a new CPU — but it does sometimes need more RAM, and DDR4 will only get scarcer. If you know your RAG pipeline or fine-tuning experiments will eventually want 64GB and you can’t commit today, AM5 keeps that door open at survivable prices. If the machine’s job description is “hold a 3090 and serve models that fit in it,” the door is worth $0.

The trap that halves your offload speed (and how to catch it)

Real problem from this exact class of build: Ryzen boards ship with DDR4 running at the JEDEC default — 2133 MT/s, not the 3200 printed on the box — until you enable the XMP/DOCP profile in BIOS. Model inference in VRAM won’t care, which is why people miss it for months; the moment anything touches system RAM (model staging, KV-cache spill, CPU offload), you’re at two-thirds of the bandwidth you paid for.

Check it on Linux:

$ sudo dmidecode --type 17 | grep -i "configured memory speed"
	Configured Memory Speed: 2133 MT/s   ← wrong, XMP is off
	Configured Memory Speed: 2133 MT/s

Fix: BIOS → OC/Tweaker → enable DOCP/XMP Profile 1 (MSI calls it A-XMP), save, reboot, re-run the check. Expected output after:

	Configured Memory Speed: 3200 MT/s

On Windows, Task Manager → Performance → Memory shows the live speed. Thirty seconds, and it’s the difference between 34 GB/s and 51 GB/s of real dual-channel bandwidth.

What to actually buy

Prices as of September 2026, all verified in the comparison above:

Your situation	The machine	Price	Where
Cheapest path to 24GB VRAM, models fit in it	AM4 build above (used RTX 3090)	~$2,120–$2,480	Check 3090 price
Same rig, but you’ll genuinely add RAM/CPU later	AM5 variant (+$300)	~$2,440–$2,760	Check 7600X bundle
Capacity over speed — 100B+ MoE models, no DIMM exposure	GMKtec EVO-X2 128GB (soldered LPDDR5X)	~$3,649	Our Strix Halo verdict
Undecided — test the workload before buying anything	Rented 3090, from $0.07/hr	pay per hour	Vast.ai

The mini PC row deserves one sentence of honesty: the EVO-X2’s soldered 128GB dodges the DIMM crisis entirely and holds gpt-oss-120b (~55 tok/s) — but on anything that fits in 24GB, the 3090 is three to four times faster. Capacity or speed; the crisis didn’t change that trade, it just repriced it.

If your endgame is a local coding backend for Cline or Cursor, the 24GB tier runs the current best local coding models well — our sister site has the BYOK setup side covered.

FAQ

Is buying into a dead socket actually fine? For this machine, yes. The upgrade AM4 can’t offer (a faster future CPU) is the one an inference box doesn’t need; the upgrade it also can’t offer cheaply forever (more DDR4) you should buy on day one regardless. Treat the platform as an appliance around the GPU.

Why not a used full system instead of parts? Often the right call — an off-lease Ryzen 5000 tower with 32GB of DDR4 already in it can undercut the platform parts list and skip the RAM purchase entirely. Check the PSU (most OEM units can’t feed a 3090) and the physical clearance (most 3090s are triple-slot cards over 300mm long). If you have to replace PSU + case anyway, the savings mostly evaporate.

Won’t DDR4 prices crash once AM4 demand dries up? The reporting points the other way: production has ended at all three majors while hundreds of millions of DDR4 systems still want upgrades, and the shortage is projected into late 2027. Betting on cheaper DDR4 later is betting on a discontinued part getting more plentiful.

Should I wait out the whole crisis instead? For RAM-heavy builds, waiting is a real option — we’ve said so since September. But this build is deliberately RAM-light ($190–$263 of exposure), and used 3090s went from roughly $1,010 in March to ~$1,250 in September — about 24% in six months. Here, waiting costs more than it saves.

Recommended Gear

Products linked in this build, first mentions above:

Used RTX 3090 24GB — the whole point
AMD Ryzen 5 5600 — enough CPU, cooler included
MSI MAG B550 Tomahawk — PCIe 4.0 x16 for the card
Corsair Vengeance LPX 32GB DDR4-3200 — the escape hatch itself
NZXT C850 850W — transient headroom
Samsung 990 Pro 2TB — model library

Sources

Last updated September 25, 2026. Prices and specs change; verify current rates before purchasing.

Was this article helpful?

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.