About VRAMForge
VRAMForge estimates whether a given large language model or diffusion model fits on given hardware, how fast it will run, and what it costs to operate — before you buy the GPU.
How it works
Everything runs in your browser with a JavaScript engine (≈40 KB). It computes weights, KV cache (with sliding-window layers, MLA latent caches and 16-token block alignment), activations with chunked prefill, framework and driver overheads, multi-GPU/TP/PP/EP sharding, offloading, decode/prefill timing, power draw and energy cost.
Calibration
The speed model is calibrated against real llama.cpp benchmarks we ran on an RTX 4070 Laptop (Qwen2.5-3B/8B and Qwen2.5-14B in Q4/Q3), and the fine-tuning model against measured PyTorch AMP peaks. Expected accuracy: weights ±2%, KV ±5%, total VRAM ±8–12%, throughput ±20–30%, first-token latency ±30%. Your runtime, drivers and quantization quality will move these numbers.
Data
Model architecture fields are collected from public Hugging Face configurations, vendor model cards and similar sources, and are normalised to a common schema. Where a field is unpublished we mark the model as partially inferred and say so in the UI. GPU memory, bandwidth and dense FP16 throughput are taken from vendor specifications.
What is not modelled
Interconnect topology beyond simple constants, scheduler P99 effects, prefix-cache hit rates, speculative decoding, TP scaling curves, pipeline bubbles, quantization quality loss, multimodal encoders, allocator fragmentation, training data pipelines, power-supply efficiency and cloud storage/network costs.
Independence
VRAMForge is not affiliated with any GPU vendor or model provider. Results are estimates for planning, not guarantees.
Contact
关于 VRAMForge
VRAMForge 帮你在买卡之前判断:某个大模型或扩散模型能否装进指定硬件、跑多快、每小时多少电费。
工作原理
全部计算在浏览器内完成(JavaScript 引擎约 40 KB):权重、KV 缓存(支持滑窗层、MLA 潜向量与 16-token 块对齐)、分块预填充的激活、框架与驱动开销、多卡 TP/PP/EP 切分、卸载、解码/预填充耗时、功耗与电费。
校准
速度模型基于我们在 RTX 4070 Laptop 上跑的真实 llama.cpp 基准(Qwen2.5-3B/8B 与 14B 的 Q4/Q3),微调模型基于 PyTorch AMP 实测峰值。预期精度:权重 ±2%、KV ±5%、总显存 ±8–12%、吞吐 ±20–30%、首字延迟 ±30%。实际运行还受驱动、运行时与量化质量影响。
数据来源
模型结构字段来自公开的 Hugging Face 配置、厂商模型卡等来源,并归一化为统一 schema;未公开的字段会标记为“部分推断”并在界面说明。显卡显存、带宽与稠密 FP16 算力取自厂商规格。
哪些没有建模
互联拓扑(仅用常数)、调度 P99、前缀缓存命中率、投机解码、TP 效率曲线、流水线气泡、量化质量损失、多模态编码器、分配器碎片、训练数据管线、电源效率与云存储/网络成本。
独立性
VRAMForge 与任何显卡厂商、模型提供方无关联。结果用于规划参考,不构成保证。