Target Hardware
Sort
Filter
No hardware matches this filter.
GB/s
GB
TFLOPS
Hardware Compare
Calculated Requirements
Total VRAM Needed
—
GB
Model Weights
—
GB
VRAM Headroom
—
GB
Fits in VRAM?
—
Decode Speed
—
tokens/sec
Prefill Speed (PP)
—
tokens/sec
Max Time (Full Context)
—
hh:mm:ss
KV Cache Memory
—
GB
Load from HuggingFace / ModelScope
e.g. https://huggingface.co/unsloth/Qwen3-30B-A3B-GGUF · https://modelscope.ai/models/Qwen/Qwen3-8B (find models at modelscope.ai/search)
Tinker
Runtime options & configuration
Runtime & attention
Support sources
Model architecture & weights
Attention architecture
32K tokens
Attention Buffer
0.0 GB
FA: scores stay on-chip
Weight estimation
Devices & placement
Hardware & shared RAM
TFLOPS
GB/s
GB
Component placement & CPU
Rig placement
GPU residency and inter-device transport are separate. Mixed rigs use layer splits on one machine; PCIe peer access is not assumed.
Paged Attention & throughput
Memory Savings
- KV per token
- —
- Contiguous
- —
- Paged + prefix
- —
- Saved
- —
Speedup / Throughput
%
- Per user
- —
- Aggregate
- —
- Speedup vs N=1
- —
- Bottleneck
- —
- Prefill → cached
- —
Offload to RAM / Split Layers
Presets use dual-channel peak bandwidth, rounded to 0.1 GB/s. Real bandwidth is lower; enter a measured value or a different channel count manually below.
GB/s
GB
- In VRAM
- —
- In RAM
- —
- Read / token
- —
- Decode
- —
- Time in RAM
- PCIe link
- —
- Prefill
- —
Speculative decoding
GB
- Standard
- —
- Speculative
- —
- Gain
- —
- Step bound
- —
Specialised model components
Hybrid / limited-context attention
GB
Model Quantization
Standard 1-bit / 3-bit choices estimate IQ1 / Q3-family storage at 1.75 / 3.9 bits per weight. Format and runtime support are unverified. Imported GGUF uses complete file bytes when available.