22, 20 or 48 GB through VRAM Mods: RTX 2080 Ti, RTX 3080 and RTX 4090D for Local LLMs

A VRAM modification can enable another model class. Its initial benefit is capacity rather than automatically greater compute performance. Offers for RTX 2080 Ti 22 GB, RTX 3080 20 GB and RTX 4090D 48 GB therefore need comparison with official GPUs and credible evidence.

No original test of these modifications is available here. The assessment separates manufacturer data, supplier claims and unverified variants. It does not endorse an AliExpress listing as a reliable purchase.

Three modifications with different evidence

GPU Solutions lists RTX 2080 Ti and RTX 4090 memory upgrades. This establishes an offered service, not independent sustained-load verification. A 4090 modification also does not automatically establish the 4090D variant. The correct smaller card name is RTX 2080 Ti with 22 GB.

Capacity that makes a difference

An assumed total model budget of 23 GB fits neither 20 nor 22 GB. An official 24 GB GPU might barely suffice with appropriate headroom, while a 48 GB variant provides more space. These figures illustrate capacity rather than demonstrating a specific runtime.

A modification does not automatically increase bandwidth or supported compute formats. An older architecture remains older. A modern quantisation kernel can impose additional requirements even when modified memory holds the weights.

Document the original GPU and modification

A credible offer identifies GPU, PCB revision, memory chips, VBIOS and cooling. Photos and device reports must match the offered revision. Reported capacity alone says little about faults in individual memory regions or stability after restarting.

Verification includes a full memory test, repeated model startup and sustained LLM load. Temperature, clocks and errors belong in the report. Another GPU or seller’s results do not validate the actual unit.

AliExpress and import costs

Offer comparisons require seller identity, actual total price, delivery window and a workable return process. Shipping and confirmed additional charges supplement the GPU price. Long-distance returns can consume the apparent saving.

A complete modified GPU also differs from a labour-only service. The named repair supplier notes that listed labour prices do not automatically include memory modules. Comparing listing headlines alone can therefore produce the wrong cost basis.

Include official alternatives

The RTX 3090 with 24 GB is an official alternative to the smaller memory modifications. Intel Arc Pro with 32 GB requires a different software setup. The RTX 6000 Ada offers an official 48 GB option on a single GPU, while Apple systems provide capacity through shared memory.

The decision combines price, demonstrated functionality and setup effort. A modified GPU becomes useful when its additional memory solves a specific task and its actual revision has documented operation. The toolbox estimator helps with the budget; an online calculator cannot replace hardware verification.

Memory planning for model selection: LLM VRAM estimator.

Sources

1. GPU Solutions: angebotene VRAM-Umbauten, Anbieterangaben
2. NVIDIA: offizielle RTX 4090 D 24 GB
3. NVIDIA: RTX 3090
4. NVIDIA: RTX 6000 Ada 48 GB
5. Intel: Arc Pro B70 – Datenblatt, 32 GB
6. vLLM: GPU installation
7. Ollama: Hardware support
8. NVIDIA: RTX 3080 – offizielle Speichervarianten

Sources checked: 5 October 2026.