NVIDIA · Consumer GPU

GeForce RTX 3090

24 GB of GDDR6X on Ampere. Check which quantized models fit the memory budget, then inspect speed and cost evidence for this variant.

Older generation; purchase condition and current supply need a seller quote.

Chip maker: NVIDIA
Memory: 24 GB GDDR6X
Architecture: Ampere
Card class: Consumer

Single-card inference · estimated

What models fit the memory budget?

Q4_K_M weight files · FP16 KV cache · one request · 2 GiB runtime reserve. Context includes prompt and generated tokens.

3 of 4 sample models within the estimated budget at 4K context.

Qwen3 8B

8B · Q4_K_M

Within memory budget

Quantized weight files: 4.68 GiB
FP16 KV cache: 0.56 GiB
Runtime reserve: 2.00 GiB
Total estimated memory: 7.24 GiB

15.11 GiB remaining

Model files & architecture

Inspect the quantized files ↗

Qwen/Qwen3-8B-GGUF · revision 7c41481f57cb95916b40956ab2f0b139b296d974.

KV architecture source ↗: 36 layers, 8 KV heads, head dimension 128. Files checked 7 Nov 2026.

Qwen3 14B

14B · Q4_K_M

Within memory budget

Quantized weight files: 8.38 GiB
FP16 KV cache: 0.63 GiB
Runtime reserve: 2.00 GiB
Total estimated memory: 11.01 GiB

11.34 GiB remaining

Model files & architecture

Inspect the quantized files ↗

Qwen/Qwen3-14B-GGUF · revision 530227a7d994db8eca5ab5ced2fb692b614357fd.

KV architecture source ↗: 40 layers, 8 KV heads, head dimension 128. Files checked 7 Nov 2026.

Qwen3 32B

32B · Q4_K_M

Within memory budget

Quantized weight files: 18.40 GiB
FP16 KV cache: 1.00 GiB
Runtime reserve: 2.00 GiB
Total estimated memory: 21.40 GiB

0.95 GiB remaining

Model files & architecture

Inspect the quantized files ↗

Qwen/Qwen3-32B-GGUF · revision 938a7432affaec9157f883a87164e2646ae17555.

KV architecture source ↗: 64 layers, 8 KV heads, head dimension 128. Files checked 7 Nov 2026.

Llama 3.3 70B

70B · Q4_K_M

Over memory budget

Quantized weight files: 39.60 GiB
FP16 KV cache: 1.25 GiB
Runtime reserve: 2.00 GiB
Total estimated memory: 42.85 GiB

20.50 GiB over the budget

Model files & architecture

Inspect the quantized files ↗

bartowski/Llama-3.3-70B-Instruct-GGUF · revision b6c5c9f176f3279204034e1d16d393105e95cb88.

KV architecture source ↗: 80 layers, 8 KV heads, head dimension 128. Files checked 7 Nov 2026.

How we estimate memory

Exact quantized file bytes + FP16 KV cache + 2 GiB reserved for the runtime. KV bytes = 2 × layers × KV heads × head dimension × total context tokens × 2 bytes. 1 GiB = 2³⁰ bytes.

Nominal manufacturer GB labels use a conservative decimal-GB budget (1 GB = 10⁹ bytes). Actual free VRAM, engine allocation, driver and operating-system support must be checked. Passing this estimate does not establish a tested deployment or the largest model the card could run. CPU offload, multi-GPU and training are outside this calculation.

Published test evidence

How fast does it generate tokens?

No sourced speed test collected yet.

Memory fit can help shortlist GeForce RTX 3090, but it does not establish output tokens per second. Validate your chosen model, quantization and engine before budgeting around its speed.

The context selector above changes memory estimates only. Published speed records keep their original model and test conditions; these are external reports, not our measurements.

Ownership cost

Purchase price references.

No current purchase quote collected.

We have not recorded a seller, board SKU, condition and current price for this variant. Manufacturer specifications ↗

Dated compute quotes

Where can you rent this variant?

$0.29 /GPU-hourUSD · single-GPU instance

Lium

spot · GPU_INSTANCE · interruptible

Available at observation · observed 8 Nov 2026

Compute only. Storage and other fees are separate; confirm the current price and available capacity with the provider.

Full quote, configuration & evidence →

$0.3 /GPU-hourUSD · single-GPU instance

Lium

marketplace · GPU_INSTANCE

Available at observation · observed 8 Nov 2026

Compute only. Storage and other fees are separate; confirm the current price and available capacity with the provider.

Full quote, configuration & evidence →

$0.22 /GPU-hourUSD · single-GPU instance

Lium

spot · GPU_INSTANCE · interruptible

Available at observation · observed 8 Nov 2026

Compute only. Storage and other fees are separate; confirm the current price and available capacity with the provider.

Full quote, configuration & evidence →

$0.22 /GPU-hourUSD · single-GPU instance

Lium

marketplace · GPU_INSTANCE

Available at observation · observed 8 Nov 2026

Compute only. Storage and other fees are separate; confirm the current price and available capacity with the provider.

Full quote, configuration & evidence →

$0.22 /GPU-hourUSD · single-GPU instance

Lium

spot · GPU_INSTANCE · interruptible

Available at observation · observed 8 Nov 2026

Compute only. Storage and other fees are separate; confirm the current price and available capacity with the provider.

Full quote, configuration & evidence →

$0.64 /GPU-hourUSD · single-GPU instance

Lium

spot · GPU_INSTANCE · interruptible

Available at observation · observed 8 Nov 2026

Compute only. Storage and other fees are separate; confirm the current price and available capacity with the provider.

Full quote, configuration & evidence →

$0.22 /GPU-hourUSD · single-GPU instance

Runpod

community · GPU_INSTANCE · non-interruptible

Available at observation · observed 8 Nov 2026

Compute only. Storage and other fees are separate; confirm the current price and available capacity with the provider.

Full quote, configuration & evidence →

This is the matching evidence currently in our catalog. It is not a complete provider comparison or a market-wide lowest-price claim.

Hardware identity & sources.

Chip designer: NVIDIA. Older generation; purchase condition and current supply need a seller quote. Retail board brands, seller SKUs and regional availability need their own evidence.

Official hardware specifications ↗ · checked 7 Nov 2026.

Model files and pinned revisions are attached to each memory estimate. Speed and price records link to their sources above. Rebuilding this page does not refresh those observation dates.