Local AI VRAM & Hardware Capability Workbench

● WebGPU: Probing...
✓ Optimal Fit in 8GB VRAM (67.8% utilization)
Model weights, KV cache, and activations fit comfortably inside GPU VRAM without host RAM offloading.
VRAM Allocation Stack Breakdown Total: 5.43 GB / 8.00 GB
Weights 4.58GB
KV 0.50GB
Act 0.35GB
Free 2.57GB
Model Weights 4.58 GB
KV Cache 0.50 GB
Activation Overhead 0.35 GB
Total VRAM Required 5.43 GB
Predicted Speed 38.2 tok/s Memory bandwidth bound
VRAM Utilization 67.9% Target: 8 GB Physical VRAM
KV Memory / Token 128 KB GQA (Grouped Query Attn)
Host Offload Risk 0.00 GB No PCIe bottleneck slowdown
Side-by-Side Micro-Model Efficiency Comparison High utility local models on consumer hardware
Model Name Category Parameters VRAM Footprint Estimated Speed / Latency Consumer Viability
WebGPU Browser Acceleration Diagnostic
Scanning local browser WebGPU device capabilities...

Trace the saved memory allocations and calibration branches

Read the explanation

For the default eight-billion model at Q4 K M, four point five bits per parameter first gives four point five gigabytes. The source then calibrates this particular fixture to exactly four point five eight. Both bars use ninety pixels per assigned decimal gigabyte. The multiplier and explicit override are source calibration choices. They are not independently measured checkpoint sizes or numerical quality results. At exactly four thousand ninety-six context tokens, this model overrides cache allocation to point five gigabytes. At eight thousand one hundred ninety-two, it uses the general equation instead: two times thirty-two layers, eight heads, one hundred twenty-eight dimensions, two bytes per value, times tokens, divided by one billion. That gives one point zero seven three seven four one eight two four. Bars use four hundred pixels per decimal gigabyte. The special default branch does not scale with batch size; preserve that limitation. The default calibrated weights four point five eight plus cache point five plus fixed activation point three five give five point four three. Eight gigabytes of selected VRAM leaves two point five seven. Both bars use sixty pixels per assigned gigabyte. The source also overrides default throughput to thirty-eight point two tokens per second. Its fits and speed labels remain illustrative arithmetic, not verified execution, device support or an actual benchmark.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.