VRAM TableGPU fit and local-LLM VRAM for OEM desktops, from published specifications Search by model
Browse categories

NVIDIA RTX 3090TI

Factory graphics option · 24 GB · catalogue identifier rtx-3090ti-24gb · last updated 2026-09-24

1 desktop in this database carries a verdict for the NVIDIA RTX 3090TI: 1 fits. Fits without conditions: the Dell Precision 3660 Tower. The largest listed model its 24 GB holds at 32k context is EXAONE 4.0 32B (Q4_K_M, needs 21.8 GiB).

This is a factory option, not a card you can buy. A computer manufacturer lists it in its own specification sheet for a machine, exactly as “NVIDIA RTX 3090TI”. The published specifications therefore record a card of this class as a factory option for that machine, but no card-level specification is published for the option itself, so this page carries no dimensions, no board power, no connector, and no purchase link.

We have not physically tested this combination. Data is compiled from official specification sheets and cited sources.

Card memory: 24 GB · A read 2026-09-21 [S1]

Each graded figure above cites: [S1] GPU - Discrete - Precision 3660 Tower Setup and Specifications.

Desktops whose verdicts reference this option

This card has been assessed against 1 desktop in this database. Each verdict is a statement about published specifications, not about a tested combination.

  • ✓ Fits: 1
  • △ Fits with conditions: 0
  • ✕ Does not fit: 0
  • ? Unverified: 0

✓ Fits 1 desktop:

  • confidence A (1): Dell Precision 3660 Tower (The manufacturer offers this card as a factory configuration for this model, so the chassis, the bracket and the supply are established by that shipping fact — for this model alone. The manufacturer's specification sheet gives no dimensions and no power draw for the card on its own, so the coverage requirement does not apply to it.)

Verify with the manufacturer before purchasing.

What each verdict means. Every one of them is a statement about published specifications.
VerdictMeaning
✓ Fits Published specifications show no physical or electrical conflict.
△ Fits with conditions Published specifications allow it only under the stated conditions.
✕ Does not fit Published specifications rule the combination out: the card needs a connector the machine does not have, draws more than the supply provides, or is physically too large. It is a statement about specifications, not a report of a failed attempt.
? Unverified There is not enough evidence to say.

A dashed Estimated badge at the start of a line marks every verdict on that line as published as an estimate, and the grade beside it applies to each of them. The document behind each verdict, and the desktop's own figures, are listed on the desktop's own page. This card has a verdict for 1 of the 82 desktops in this database; the remaining 81 have not been assessed against it, which is not the same as them being ruled out.

Local language models this card's memory holds

Arithmetic shared by the models below. Sentences that read the same for more than one model are stated once here; each row names the ones it carries by number.

  1. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B. (5 rows)
Calculated from the published file size and the model architecture.
ModelOutcome Needs at 32k contextThe arithmetic Confidence
EXAONE 4.0 32B Q4_K_M ✓ Fits in VRAM 21.8 GiB 24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens. B
Qwen3 30B-A3B Instruct 2507 Q4_K_M ✓ Fits in VRAM Estimated 21.3 GiB 24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. [1] BD
Gemma 3 27B Q4_0 (QAT) ✓ Fits in VRAM 20.0 GiB 24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens. B
Mistral Small 3.2 24B Instruct Q4_K_M ✓ Fits in VRAM Estimated 19.4 GiB 24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. [1] BD
Phi-4 14B Q4_K_M ✓ Fits in VRAM Estimated 15.6 GiB 24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. [1] BD
GPT-OSS 20B MXFP4 (the vendor's own native quantisation) ✓ Fits in VRAM 14.6 GiB 24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens. B
Qwen3 14B Q4_K_M ✓ Fits in VRAM 14.4 GiB 24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens. B
Gemma 3 12B Q4_0 (QAT) ✓ Fits in VRAM 10.9 GiB 24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. B
Qwen3 8B Q4_K_M ✓ Fits in VRAM 10.2 GiB 24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. B
Llama 3.1 8B Q4_K_M ✓ Fits in VRAM Estimated 9.6 GiB 24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B. BD
Qwen3 4B Q4_K_M ✓ Fits in VRAM 7.9 GiB 24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. B
Qwen3 1.7B Q4_K_M ✓ Fits in VRAM Estimated 5.6 GiB 24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. [1] BD
Qwen3 0.6B Q4_K_M ✓ Fits in VRAM Estimated 4.9 GiB 24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. [1] BD
Gemma 3 4B Q4_0 (QAT) ✓ Fits in VRAM 4.7 GiB 24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. B
Qwen3 32B Q4_K_M △ Fits at short context only 27.5 GiB 24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens. B

✕ Exceeds VRAM 3 models need more memory at 32k context than this card has, 2 of them published as an estimate. The arithmetic for each is on the model's own page.

Exceeds VRAM does not mean the model will not run: inference software can keep part of a model in system memory. How fast that is, and whether it is pleasant to use, is outside what this site measures. The sizing rule and its constants are on the Methodology page.

Cards whose published memory size is the same, so the local-model outcomes on this page apply to them as well. Whether a given desktop accepts them is a separate question, answered on each card's own page.

Sources

Every document this page draws on, with the date it was read.
DocumentPublisherKindRead on
GPU - Discrete - Precision 3660 Tower Setup and Specifications Dell Technologies primary 2026-09-21

We have not physically tested this combination. Data is compiled from official specification sheets and cited sources. Verify with the manufacturer before purchasing.