GPU OBJECT :: nvidia-rtx-3090-24gb-used

GeForce RTX 3090 24GB (used)

Evidence-first hardware detail for local AI.

VERIFIED_SPECRANKING ELIGIBLE
24GB
FAST MEMORY
936GB/s
BANDWIDTH
€850PRICE
OBSERVATION
← All GPUsSpecsMarketVariantsModel fitPerformanceSoftwarePros / ConsCompareSources

DIRECT ANSWER

GeForce RTX 3090 24GB (used) provides 24 GB of fast memory, 936 GB/s of memory bandwidth, and reaches at least minimum memory fit for 7 of 9 currently listed BrunoSan models. Its known software path includes CUDA, PyTorch, vLLM, llama.cpp. BrunoSan publishes tokens/s only with model- and runtime-specific benchmark evidence.

01 · HARDWARE

Technical data

Canonical GPU facts. Board-specific differences stay in variant evidence.

Fast memory / VRAM24 GBVERIFIED_SPEC
Memory typeGDDR6XVERIFIED_SPEC
Memory bus384 bitSOURCE_OBSERVED
Memory bandwidth936 GB/sVERIFIED_SPEC
Board power350 WVERIFIED_SPEC
Recommended PSU750 WVERIFIED_SPEC
Graphics processorGeForce RTX 3090SOURCE_OBSERVED
Processor familyNVIDIASOURCE_OBSERVED

02 · MARKET

Current market

Observed price

€850MARKET_ESTIMATE

Observed 2026-09-13

Price history begins when repeated merchant observations are connected.

Buying guidance

Strong memory fit for Qwen3-Coder-30B-A3B, Qwen3-30B-A3B, Qwen3.5-35B-A3B and more. The listed configuration carries a 750 W minimum PSU recommendation. For used purchases, verify condition, cooling, warranty, power connectors and return rights.

03 · VARIANTS

Observed board variants

Dimensions, connectors and outputs stay variant-scoped.

VariantGTINPCIePSUPowerDimensionsDisplay
ASUSTUF-RTX3090-O24G-GAMING0192876922828, 192876922828PCI Express 4.0300 × 52 × 127 mmDP 3 · HDMI 2
GIGABYTEGV-N3090GAMING OC-24GD4719331307547, 0889523024485PCI Express x16 4.0750 W2x 8-pin320 × 55 × 129 mmDP 3 · HDMI 2

04 · LOCAL AI FIT

Which BrunoSan models fit?

Exact BrunoSan model profiles, not generic 7B/14B buckets. Memory fit does not claim maximum context or runtime compatibility.

ModelMemory targetFitContext
GLM-4.7-Flash18–24 GBRecommended fitContext/KV cache unbenchmarked
Nemotron 3.5 Lightning 30B-A3B20–24 GBRecommended fitContext/KV cache unbenchmarked
Qwen3-30B-A3B20–24 GBRecommended fitContext/KV cache unbenchmarked
Qwen3-Coder-30B-A3B20–24 GBRecommended fitContext/KV cache unbenchmarked
Qwen3.5-35B-A3B20–24 GBRecommended fitContext/KV cache unbenchmarked
Qwen3.8-27B18–24 GBRecommended fitContext/KV cache unbenchmarked
gpt-oss-20b16–24 GBRecommended fitContext/KV cache unbenchmarked
Qwen3-Coder-Next 80B-A3B48–64 GBDoes not fit fast memoryContext/KV cache unbenchmarked
gpt-oss-120b64–96 GBDoes not fit fast memoryContext/KV cache unbenchmarked

05 · PERFORMANCE

Evidence before speed claims

Memory bandwidth: 936 GB/s

BENCHMARK TRUTH

No model/runtime-specific benchmark is registered. No tokens/s number is published.

06 · SOFTWARE

Runtime path

CUDA-native path with broad support across common local-AI runtimes; exact model support still depends on the runtime version.

CUDAPyTorchvLLMllama.cppExLlamaTensorRT ecosystem

OS: windows · linux

07 · PROS

Strengths

  • 24 GB of fast memory
  • 936 GB/s evidenced memory bandwidth
  • Recommended-memory fit for 7 BrunoSan model profiles
  • Documented software path: CUDA, PyTorch, vLLM

08 · CONS

Limits

  • Fast memory is insufficient for 2 larger model profiles
  • No model/runtime-specific benchmark in the BrunoSan registry yet
  • Used-market condition and warranty vary

09 · COMPARE

Nearby hardware choices

Related by memory capacity first; never a hidden performance ranking.

10 · SOURCES

Provenance

Every displayed fact keeps its evidence scope. Curated SSOT wins over source observations; board-specific Icecat fields are not silently generalized to a GPU family.

Specs Icecat

Icecat source observations are provided AS IS and are not used for AI training by BrunoSan.

Icecat.biz → · AS IS disclaimer →

LOCAL AI FINDER

Put this GPU into your actual decision.

Model, budget, PSU, OS and software stack belong in the same decision.

Open Finder →