localLLMriglocal, on your table

Guide

RTX 3090 vs 4090 for local LLMs

Both cards have the same memory class. The formula says a single 24 GB card does not fully hold Llama 3.3 70B Q4 at 8,192 tokens. The 4090's catalog bandwidth is only a little higher, and that bandwidth is not a citation. Prices are not live shop listings.

Same 24 GB, same full-fit limit

GeForce RTX 3090 24GB lists 24 GB and 936 GB/s. bandwidthEstimated is false, but bandwidthSourceUrl is null, so that speed is not a citation. GeForce RTX 4090 24GB lists 24 GB and 1008 GB/s. bandwidthEstimated is false, but bandwidthSourceUrl is null, so that speed is not a citation.

GeForce RTX 3090 24GB has 23.4 GB usable after the reserve in this formula (catalog memory 24 GB, vram). GeForce RTX 4090 24GB has 23.4 GB usable after the reserve in this formula (catalog memory 24 GB, vram). The reserve is 0.6 GB per card on both, so usable memory matches.

Llama 3.3 70B Q4 at 8,192 tokens is 45.8 GB in this formula (42.4 GB weights, 2.7 GB KV cache, 0.8 GB overhead). Fit on GeForce RTX 3090 24GB: not a fit. Fit on GeForce RTX 4090 24GB: not a fit. Fit on Used PC with RTX 3090 24GB: offload only, not a full fit. Fit on PC with RTX 4090 24GB: offload only, not a full fit. A full fit is what this catalog means by holding the model. Offload is not that.

Qwen2.5 32B Q4 at 8,192 tokens is 22.6 GB in this formula (19.7 GB weights, 2.1 GB KV cache, 0.8 GB overhead). Fit on Used PC with RTX 3090 24GB: a full fit. Fit on PC with RTX 4090 24GB: a full fit. That is the 32B Q4 class in this catalog. Qwen2.5 32B: 64 layers, 8 KV heads, head dim 128, max context 131,072. Those fields come from https://huggingface.co/Qwen/Qwen2.5-32B/raw/main/config.json. The parameter count 32.8B is still an estimate (paramsSourceUrl null).

Bandwidth, and what the formula will not claim

The catalog lists 936 GB/s for the 3090 and 1008 GB/s for the 4090. bandwidthEstimated is false on both rows, but bandwidthSourceUrl is null, so those GB/s figures are not a citation. Decode in the methodology is treated as memory-bandwidth bound, with the same NVIDIA utilisation (0.78) on both cards. The catalog gap is 72 GB/s. That is a small share of either figure, and it is not a measured speed.

Tokens per second on the calculator are estimates from the methodology page. benchmarkSourceUrl stays null. This guide does not add a measured llama-bench result.

Prices in the catalog, not a shop

  • GeForce RTX 3090 24GB: $1,370 (USD is an FX estimate), €1,224 incl. VAT. Global prices (USD), EUR available. Marked used. used (median of sold listings). Used median €1,224, typical €1,136–€1,312, 232 eBay sold sales over 90 days, updated 16 Oct 2026. Not an EU retail listing. Source: pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-3090-price-used-and-specs), dated 2026-10-16. Also new (lowest in-stock EU) €2,218 from GPU PRIX (https://gpuprix.com/eu/gpus/geforce-rtx-3090), dated 2026-10-16. New low €2,218, Zotac Trinity OC, France. Page last updated 16 Oct 2026.
  • GeForce RTX 4090 24GB: $2,446 (USD is an FX estimate), €2,185 incl. VAT. Global prices (USD), EUR available. Marked used. used (median of sold listings). Used median €2,185, typical €2,074–€2,290, 111 sales, updated 16 Oct 2026. Not an EU retail listing. Source: pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-4090-price-used-and-specs), dated 2026-10-16. Also new (lowest in-stock EU) €3,790 from GPU PRIX (https://gpuprix.com/eu/gpus/geforce-rtx-4090), dated 2026-10-16. Current low €3,790, MSI Gaming X Slim, Belgium. Page last updated 15 Oct 2026, checked 16 Oct 2026. In-stock boards in that table start at €3,990.
  • Used PC with RTX 3090 24GB: $2,602 (USD is an FX estimate), €2,324 incl. VAT. Global prices (USD), EUR available. Marked used. GPU line is the used median €1,224, labelled used, from pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-3090-price-used-and-specs). 64 GB DDR4 (estimate) €600, from the 16 Oct 2026 fact-check. Rest of system (estimate) €500. Other parts from €500 in that fact-check. Not a shop listing. The total is the GPU price plus the RAM line plus that rest estimate. It is not a checked system price. price unverified. sourceUrl is null. Dated 2026-10-16.
  • PC with RTX 4090 24GB: $3,678 (USD is an FX estimate), €3,285 incl. VAT. Global prices (USD), EUR available. Marked used. GPU line is the used median €2,185, labelled used, from pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-4090-price-used-and-specs). 64 GB DDR4 (estimate) €600, from the 16 Oct 2026 fact-check. Rest of system (estimate) €500. Other parts from €500 in that fact-check. Not a shop listing. DDR5 was not separately priced, so this total uses the sourced 64 GB DDR4 line. The total is the GPU price plus the RAM line plus that rest estimate. It is not a checked system price. price unverified. sourceUrl is null. Dated 2026-10-16.
  • Dual RTX 3090, 48 GB: $6,054 (USD is an FX estimate), €5,408 incl. VAT. Global prices (USD), EUR available. Marked used. GPU line is 2× the used median €1,224 (€2,448), labelled used, from pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-3090-price-used-and-specs). 128 GB DDR4 (estimate) €2,460, from the 16 Oct 2026 fact-check. Rest of system (estimate) €500. Other parts from €500 in that fact-check. Not a shop listing. The total is the GPU price plus the RAM line plus that rest estimate. It is not a checked system price. price unverified. sourceUrl is null. Dated 2026-10-16.

Each amount above is the catalog EUR price, VAT included, plus a dollar figure from FX_EUR_USD. The stored 3090 and 4090 figures are used medians from pcprice.watch. Each card also cites a GPU PRIX new price. Build totals add a sourced RAM line and rest of system (estimate) and are not a checked system price, so sourceUrl is null on those build rows.

When the 4090 is the better card

Prompt processing is compute-heavy. This site does not measure prefill, so there is no table for it. The catalog TDP is 450 W on the 4090 card and 350 W on the 3090 card, with no separate source URL. More compute and a higher power limit are a reason to expect faster prompt processing on the 4090. That is reasoning about the generation of the card, not a benchmark we ran. The decode formula does not give the 4090 any advantage beyond the bandwidth numbers above.

When the 3090 is the better buy

Both bare cards are 24 GB. The used 3090 card row is the cheaper 24 GB entry in this catalog. Llama 3.3 70B Q4 at 8,192 tokens is 45.8 GB in this formula (42.4 GB weights, 2.7 GB KV cache, 0.8 GB overhead). Q4 means Q4_K_M. Fit on Dual RTX 3090, 48 GB: a full fit (usable 46.8 GB across 48 GB). Dual RTX 3090, 48 GB: $6,054 (USD is an FX estimate), €5,408 incl. VAT. Global prices (USD), EUR available. Marked used. GPU line is 2× the used median €1,224 (€2,448), labelled used, from pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-3090-price-used-and-specs). 128 GB DDR4 (estimate) €2,460, from the 16 Oct 2026 fact-check. Rest of system (estimate) €500. Other parts from €500 in that fact-check. Not a shop listing. The total is the GPU price plus the RAM line plus that rest estimate. It is not a checked system price. price unverified. sourceUrl is null. Dated 2026-10-16. Dual RTX 4090, 48 GB: $8,206 (USD is an FX estimate), €7,330 incl. VAT. Global prices (USD), EUR available. Marked used. GPU line is 2× the used median €2,185 (€4,370), labelled used, from pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-4090-price-used-and-specs). 128 GB DDR4 (estimate) €2,460, from the 16 Oct 2026 fact-check. Rest of system (estimate) €500. Other parts from €500 in that fact-check. Not a shop listing. DDR5 was not separately priced, so this total uses the sourced 128 GB DDR4 line. The total is the GPU price plus the RAM line plus that rest estimate. It is not a checked system price. price unverified. sourceUrl is null. Dated 2026-10-16. Two used 3090s are how this catalog opens 70B Q4_K_M. Each build adds rest of system (estimate).

Parts on the dual 3090 row: 2× GeForce RTX 3090 24GB, used; 128 GB DDR4 (estimate): €2,460. from the 16 Oct 2026 fact-check.; Rest of system (estimate): €500. Other parts from €500 in that fact-check. Not a shop listing.; Ryzen 9 5900X or similar; 1200W PSU, a case that can breathe. Combined TDP in the catalog is 700 W. That describes the row. It is not a power measurement.

Not the 5090 question

A 32 GB card is a different comparison. The RTX 5090 street situation is a separate guide, and that page labels the discontinuation as a leak. This page does not repeat it. RTX 5090 leak-labelled guide.

Open the calculator: Qwen2.5 32B Q4, Llama 3.3 70B Q4, Llama 3.3 70B Q4 under €2,000. Hardware table: cards and builds.

Related: Dual RTX 3090 build for 70B, Used RTX 3090 buying guide (EU), How much VRAM does a 70B model need?, Best GPU for local LLMs in 2026.

Open the calculator

RTX 3090 vs 4090 for local LLMs · localLLMrig