INTERACTIVE BENCHMARK SIMULATION
nexusvllm Engine vs Baseline PyTorch
Simulated throughput comparison: Continuous batching vLLM + TensorRT-LLM with PagedAttention vs vanilla PyTorch unbatched baseline.
Sample Prompts:
Legacy Hugging Face / PyTorch
TTFT: 185ms•32.4 tok/s
Awaiting race trigger... Slow cold start and high KV cache overhead will be visible here.
Time Elapsed:0s
VRAM Consumption:74.2 GB (FP16)
Cost / 1M Tokens:$2.20
NEXUS AI (TensorRT-LLM FP8)
TTFT: 14.8ms•165.2 tok/s
Awaiting race trigger... Continuous batching and FP8 Tensor Cores will stream tokens instantly.
Time Elapsed:0s
VRAM Consumption:41.2 GB (FP8 TensorRT)
Cost / 1M Tokens:$0.42 (-72%)