VRAM GB →GiB / MiB
A100 80 GB. nvidia-smi reports 81,920 MiB, not 74.5 GiB — GPU labels are binary, so the storage conversion runs the wrong way. ECC rows, CUDA context and library workspace eat the rest.
GB-to-GiB Converter →TFLOPS →PFLOPS
8× H100 nodes = 15.8 PFLOPS FP8. Your architecture review deck still says TFLOPS. Convert before the VP asks why the number looks small.
TFLOPS →PFLOPS Scaler →Mbps →MB/s Throughput
ISP quotes 1,000 Mbps. Your NVMe array benchmarks in MB/s. Divide by 8. Subtract 5—2% for protocol overhead. Now size the pipe.
Mbps →MB/s Converter →Gbps →TB / Day Egress
10 Gbps sustained = 108 TB/day. At AWS egress rates that's $9,180/day. Model the pipe before the CFO forwards you the bill with a question mark.
Gbps →TB/day Egress Meter →