RTX 3080
A high-end graphics card released by NVIDIA, based on the Ampere architecture and equipped with 10GB or 12GB of VRAM, suitable for gaming and AI computing.
Timeline (last 90 days)
Two RTX 3080s combined can run Strata at up to 109 tokens per second; officially, a single 3090 is expected to reach 100 to 140 tokens per second.
Qwen2.5-Coder-7B can run smoothly on consumer-grade GPUs (such as the RTX 3080 10G) and is capable of handling tasks such as code completion, function generation, and debugging suggestions; models of 32B or above are recommended for complex architecture design or multi-file refactoring
There is precedent for using QLoRA to complete fine-tuning on a single GPU with 8 GB of VRAM, such as early community practices of fine-tuning LLaMA 7B on an RTX 3080.
In Path C, an RTX 3080 paired with a DGX Spark improves Prefill speed from 1111.3 to 1600 (approximately 44%), but Decode speed drops from 33.36 token/s to 17.62 token/s.
In Path B, an RTX 3080 handles Prefill and an AI Max 395 handles Decode, compressing time-to-first-token from 4.825 seconds to 1.073 seconds under Q4 quantization.
Large diffusion models often face insufficient VRAM or slow inference on consumer-grade GPUs such as the RTX 3080 10GB.
The project's policy was trained from scratch using mjlab (a MuJoCo wrapper tool) and converged in approximately 4 hours on a single RTX 3080.
The policy was trained from scratch using mjlab (a wrapper for MuJoCo) and converged in approximately 4 hours on a single RTX 3080.
All Facts (9)
DeepSeek-R1 14B模型约需10-12GB显存，需要中高端独立显卡如RTX 3080/4070
75%UnverifiedTwo RTX 3080s combined can run Strata at up to 109 tokens per second; officially, a single 3090 is expected to reach 100 to 140 tokens per second.
50%UnverifiedQwen2.5-Coder-7B can run smoothly on consumer-grade GPUs (such as the RTX 3080 10G) and is capable of handling tasks such as code completion, function generation, and debugging suggestions; models of 32B or above are recommended for complex architecture design or multi-file refactoring
50%UnverifiedThere is precedent for using QLoRA to complete fine-tuning on a single GPU with 8 GB of VRAM, such as early community practices of fine-tuning LLaMA 7B on an RTX 3080.
50%UnverifiedIn Path C, an RTX 3080 paired with a DGX Spark improves Prefill speed from 1111.3 to 1600 (approximately 44%), but Decode speed drops from 33.36 token/s to 17.62 token/s.
50%UnverifiedIn Path B, an RTX 3080 handles Prefill and an AI Max 395 handles Decode, compressing time-to-first-token from 4.825 seconds to 1.073 seconds under Q4 quantization.
50%UnverifiedLarge diffusion models often face insufficient VRAM or slow inference on consumer-grade GPUs such as the RTX 3080 10GB.
50%UnverifiedThe project's policy was trained from scratch using mjlab (a MuJoCo wrapper tool) and converged in approximately 4 hours on a single RTX 3080.
50%UnverifiedThe policy was trained from scratch using mjlab (a wrapper for MuJoCo) and converged in approximately 4 hours on a single RTX 3080.
50%