← AI Hashrate

RTX 3090 24GB

NVIDIA · 24 GB GDDR6X · 936 GB/s · 350 W · MSRP $1,499 · MSI GeForce RTX 3090 Ventus 3X 24G OC (affiliate) · More listings

The RTX 3090 24GB is a NVIDIA accelerator with 24 GB GDDR6X and about 936 GB/s peak memory bandwidth. AI Hashrate estimates local LLM decode (batch 1) at 8K context: fit counts weights + KV cache + 1 GB runtime overhead — not weights alone. At Q4, about 33 curated models fit fully on this card; at FP16, about 17 fit. Top Q4 speeds: Qwen3-0.6B ≈ 744.5 tok/s (estimated, fits); gemma-3-1b-it ≈ 595.6 tok/s (estimated, fits); Llama-3.2-1B-Instruct ≈ 496.4 tok/s (estimated, fits); Qwen3-1.7B ≈ 297.8 tok/s (estimated, fits); Qwen3-30B-A3B ≈ 198.5 tok/s (estimated, fits). Relative ranking is more reliable than absolute tok/s. See methodology for the bandwidth formula and measured-anchor policy. Methodology.

Table: decode speed estimates (Q4 / FP16). Measured rows override estimates. MSRP is list price, not live retail. Context default 8K.

ModelParams (B)Quanttok/sFits?
Qwen3-0.6B0.8Q4744.5 est.Yes
gemma-3-1b-it1.0Q4595.6 est.Yes
Llama-3.2-1B-Instruct1.2Q4496.4 est.Yes
Qwen3-1.7B2.0Q4297.8 est.Yes
Qwen3-0.6B0.8FP16204.8 est.Yes
Qwen3-30B-A3B30.5Q4198.5 est.Yes
Qwen3-Coder-30B-A3B-Instruct30.5Q4198.5 est.Yes
Qwen3-30B-A3B-Instruct-250730.5Q4198.5 est.Yes
Llama-3.2-3B-Instruct3.2Q4186.1 est.Yes
gemma-3-1b-it1.0FP16163.8 est.Yes
gpt-oss-20b21.5Q4161.8 measured (llama.cpp-disc-15396)Yes
Qwen3-Coder-Next79.7Q4158.2 est.No
Phi-4-mini-instruct3.8Q4156.7 est.Yes
Phi-4-mini-reasoning3.8Q4156.7 est.Yes
gemma-3-4b-it4.3Q4152.7 est.Yes
Qwen3-Next-80B-A3B-Instruct81.3Q4152.3 est.No
Qwen3-4B-Instruct-25074.0Q4148.9 est.Yes
Qwen3-4B-Thinking-25074.0Q4148.9 est.Yes
Llama-3.2-1B-Instruct1.2FP16136.5 est.Yes
Mixtral-8x7B-Instruct-v0.146.7Q488.7 est.No
Qwen3-1.7B2.0FP1681.9 est.Yes
Qwen2.5-7B-Instruct7.6Q478.4 est.Yes
Qwen2.5-Coder-7B-Instruct7.6Q478.4 est.Yes
DeepSeek-R1-Distill-Qwen-7B7.6Q478.4 est.Yes
Qwen3-8B8.2Q472.6 est.Yes
DeepSeek-R1-0528-Qwen3-8B8.2Q472.6 est.Yes
Mistral-7B-Instruct-v0.37.2Q470.0 measured (llama-cpp-bench)Yes
Llama-3.1-8B-Instruct8.0Q462.0 measured (llama-cpp-bench)Yes
Llama-3.2-3B-Instruct3.2FP1651.2 est.Yes
gemma-3-12b-it12.2Q449.6 est.Yes
GLM-4.7-Flash31.2Q449.6 est.Yes
Qwen3-32B32.8Q445.6 est.No
Qwen2.5-32B-Instruct32.8Q445.6 est.No
Qwen2.5-Coder-32B-Instruct32.8Q445.6 est.No
DeepSeek-R1-Distill-Qwen-32B32.8Q445.6 est.No
Phi-4-mini-instruct3.8FP1643.1 est.Yes
Phi-4-mini-reasoning3.8FP1643.1 est.Yes
gemma-3-4b-it4.3FP1642.0 est.Yes
Qwen3-4B-Instruct-25074.0FP1640.9 est.Yes
Qwen3-4B-Thinking-25074.0FP1640.9 est.Yes
Phi-414.7Q440.5 est.Yes
Phi-4-reasoning14.7Q440.5 est.Yes
Qwen3-14B14.8Q440.2 est.Yes
Qwen2.5-14B-Instruct14.8Q440.2 est.Yes
Qwen2.5-Coder-14B-Instruct14.8Q440.2 est.Yes
DeepSeek-R1-Distill-Qwen-14B14.8Q440.2 est.Yes
gpt-oss-20b21.5FP1637.4 est.No
gemma-3-12b-it12.2FP1629.1 est.No
gpt-oss-120b120.4Q425.9 est.No
Devstral-Small-250723.6Q425.2 est.Yes
Mistral-Small-3.2-24B-Instruct-250624.0Q424.8 est.Yes
Magistral-Small-250924.0Q424.8 est.Yes
Qwen3-30B-A3B30.5FP1622.9 est.No
Qwen3-Coder-30B-A3B-Instruct30.5FP1622.9 est.No
Qwen3-30B-A3B-Instruct-250730.5FP1622.9 est.No
Mistral-7B-Instruct-v0.37.2FP1622.8 est.Yes
gemma-3-27b-it27.4Q421.9 est.Yes
Qwen2.5-7B-Instruct7.6FP1621.6 est.Yes
Qwen2.5-Coder-7B-Instruct7.6FP1621.6 est.Yes
DeepSeek-R1-Distill-Qwen-7B7.6FP1621.6 est.Yes
Llama-3.1-8B-Instruct8.0FP1620.5 est.Yes
Qwen3-8B8.2FP1620.0 est.Yes
DeepSeek-R1-0528-Qwen3-8B8.2FP1620.0 est.Yes
GLM-4.5-Air110.5Q419.8 est.No
Phi-414.7FP1616.5 est.No
Phi-4-reasoning14.7FP1616.5 est.No
Qwen3-14B14.8FP1616.2 est.No
Qwen2.5-14B-Instruct14.8FP1616.2 est.No
Qwen2.5-Coder-14B-Instruct14.8FP1616.2 est.No
DeepSeek-R1-Distill-Qwen-14B14.8FP1616.2 est.No
DeepSeek-V4-Flash158.1Q49.2 est.No
GLM-4.7-Flash31.2FP165.2 est.No
Llama-3.1-70B-Instruct70.6Q44.8 est.No
Llama-3.3-70B-Instruct70.6Q44.8 est.No
DeepSeek-R1-Distill-Llama-70B70.6Q44.8 est.No
Qwen2.5-72B-Instruct72.7Q44.4 est.No
Devstral-Small-250723.6FP164.1 est.No
Mistral-Small-3.2-24B-Instruct-250624.0FP163.9 est.No
Magistral-Small-250924.0FP163.9 est.No
gemma-3-27b-it27.4FP162.6 est.No
Qwen3-32B32.8FP161.5 est.No
Qwen2.5-32B-Instruct32.8FP161.5 est.No
Qwen2.5-Coder-32B-Instruct32.8FP161.5 est.No
DeepSeek-R1-Distill-Qwen-32B32.8FP161.5 est.No