LIVE TELEMETRYGPT-5$1.25 / $10.00Sonnet 5.5$2.00 / $10.00Gemini 3.8$0.75 / $3.75DS V4.1$0.15 / $0.60RTX 50901,792 GB/sM4 Max546 GB/sUS Avg$0.182/kWh
TELEMETRY CONSOLE NAVIGATION

Guncelkal Telemetry Index

Search VRAM specs, API costs, or Dev tools

AI & LLM EconomicsQ6_K • 41.49 GB Total

High-Fidelity Mixtral 8x7B Q6_K on Dual 3090s

Run 46.7B MoE parameters without aggressive quantization loss in home workstations.

Hardware & Model Configuration

32,768 tokens
1 stream
64 GB
KV Cache Precision:
CLI LAUNCH SNIPPET
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 \
  ollama run mixtral:8x7b-instruct-q6_k --ctx-size 32768
Total Required
41.49 GB
Safe cap: 45.12 GB
Est. Speed
98.3 tok/s
Decode bandwidth bound
KV Cache
2 GB
Q8_0 @ 32K ctx
Max Context Fit
32,768
Tokens before spill
VRAM ALLOCATION & KV-CACHE MAP41.5 GB / 45.1 GB Usable
Safe Pool (45.12 GB)Physical 48 GB
Weights: 38.29 GB KV Cache: 2 GB Runtime: 1.2 GB
Status: Tight Fit (Near Memory Ceiling — Cap Context)

All 7 GGUF / Quantization Tiers on 2x NVIDIA RTX 3090 NVLink (48 GB)

QuantizationMemory NeededSpeed Est.Placement
FP16 / BF16 (16-bit)Reference Lossless96.6 GB4.2 tok/sRAM Offload
Q8_0 (8-bit)Near-Lossless52.82 GB12.4 tok/sRAM Offload
Q6_K (6-bit K-Quant)Studio Fidelity41.49 GB98.3 tok/sPure GPU VRAM
Q5_K_M (5-bit Medium)Minimal Loss36.47 GB125.2 tok/sPure GPU VRAM
Q4_K_M (4-bit Sweet Spot)Recommended Default31.51 GB146.3 tok/sPure GPU VRAM
IQ3_XS (3.3-bit I-Quant)Noticeable Degradation23.05 GB205.4 tok/sPure GPU VRAM
IQ2_XXS (2.1-bit Extreme)Emergency Fit Only16.04 GB308.8 tok/sPure GPU VRAM

All 11 Open-Weight LLMs Ranked for 2x NVIDIA RTX 3090 NVLink (48 GB) (Click Row to Switch)

ModelBest Pure-VRAM QuantMemoryEst. Speed
Llama 3.1 8B InstructUltra-fast RAG extraction, IDE autocomplete & classificationFP1618.96 GB72.6 tok/s
Qwen 2.5 Coder 14BMid-tier local VS Code copilot balancing speed & repo accuracyFP1633.3 GB39.8 tok/s
Mistral Small 24BClean instruction following on 24 GB single-GPU workstationsQ8_028.9 GB45.9 tok/s
Gemma 2 / 3 27BMultilingual & editorial writing in single 24–32 GB GPUsQ8_035.55 GB39.9 tok/s
Qwen 2.5 32B / DeepSeek-R1 32BGold-standard local coding & reasoning on RTX 3090/4090/5090 & 48GB MacsQ8_039.43 GB33.8 tok/s
Mixtral 8x7B MoE (46.7B Total)Sparse Mixture-of-Experts: needs 47B RAM footprint, computes at 13B speedQ6_K41.49 GB98.3 tok/s
Llama 3.3 70B / DeepSeek-R1 70BGPT-4 class private local reasoning for 48 GB–128 GB rigsIQ3_XS36.21 GB38.6 tok/s
DeepSeek R1 671B MoE (Full Frontier)Cluster or Mac Studio Ultra / dual EPYC server deployment (37B active per token)Needs Offload / Cluster411.81 GBOOM
Qwen3-30B-A3B (MoE — 3.3B Active)Ultra-fast 30B Sparse MoE decoding at 3.3B speed; effortless fit on 24 GB cardsQ8_034.81 GB318.4 tok/s
Qwen3-32B (Dense Dual-Thinking)2026 dense local reasoning & coding benchmark with dual thinking/non-thinking toggleQ8_039.75 GB33.5 tok/s
Llama 4 Scout 109B (MoE — 17B Active)Multimodal 109B MoE (17B active) built for 64–128 GB Macs, Dual-3090, or H100 podsIQ2_XXS34.18 GB231.8 tok/s

Telemetry Methodology & Benchmark Sources

Read full equations →

Pre-configured telemetry for Mixtral 8x7B MoE (46.7B Total) (Q6_K) on 2x NVIDIA RTX 3090 NVLink (48 GB). Requires 41.49 GB (38.29 GB weights + 2 GB Q8 KV cache) against 45.12 GB usable pool for ~98.3 tok/s decode.

Primary Technical & US Grid Benchmarks

  • llama.cpp GGUF Model Sizing
  • Apple Metal / CUDA Wired VRAM Limits
  • GQA Attention Context Scaling Spec

Engineering & Pricing FAQ

Can 2x NVIDIA RTX 3090 NVLink (48 GB) run Mixtral 8x7B MoE (46.7B Total) without CPU RAM offload?
Leaves over 7 GB spare VRAM for long-context retrieval agents. At Q6_K with 32K context, total required memory is 41.49 GB compared to 45.12 GB safe usable GPU VRAM.
What max context length fits inside 2x NVIDIA RTX 3090 NVLink (48 GB)?
At Q6_K quantization on single-stream inference, safe KV-cache capacity supports up to 32,768 tokens before spilling into host system RAM.

Continue in Console Workflow

AI & LLM Economics

Cloud vs. Rig TCO

Plot cumulative 36-month API spending against hourly RunPod/H100 rentals and owned RTX/Mac hardware with electricity and resale value.

Open Workbench
Hardware & Power

kWh Power Cost Lab

Calculate daily, monthly, and annual kWh bills using 80+ PSU efficiency, duty cycle, optional AC cooling load, and 50 US state EIA rates.

Open Workbench
Hardware & Power

PSU Wattage & Spike

Combine dual-GPU draw, CPU peak boost, NVMe arrays, and ATX 3.1 transient spikes to hit the 50–70% PSU efficiency sweet spot.

Open Workbench