Online ROI Calculator

How does the inference-storage-acceleration math work out? Drag the sliders to compute in real time for your cluster size and workload profile. The model shares its source with mingxin/accel_value.py (GPU-equivalents = uplift × cold-recovery share × GPU count); every parameter labels its measured/estimated basis, so any third party can cross-check.

Your cluster parameters

GPU nodes: 8GPUs per node: 8Cold-recovery / long-context workload share: 30%(estimated parameter, backfilled by pilot measurements)Throughput uplift: +35%(measured band 29–40%, R2/R3)

Ratio auto-computed on the measured anchor of one FX100 per 8 nodes: 8 nodes (64 GPUs) → 1 array(s), storage investment ¥371,200 (≈$51,556).

Your scenario result

6.7

GPU-equivalent capacity released
(uplift × cold share × GPU count)

$80,640

Equivalent compute value
(at MI308X $12,000/GPU)

1.56×

Value multiple on storage spend
(equivalent compute value ÷ array investment)

Conservative / mid / upper scenarios (same scale)

Scenario	Uplift	Cold share	Equiv. GPUs	Equiv. value	Multiple
保守	+29%	10%	1.9	$22,272	0.43×
中值	+35%	30%	6.7	$80,640	1.56×
上界	+40%	50%	12.8	$153,600	2.98×

Throughput uplift is a measured basis (R2/R3, 480B, MI308X, production form); the cold-recovery share is an estimate (no public statistics; listed for pilot backfill). Results are mid-scenario references, not a promise of returns.

Capacity ledger: $/GB compared

MI308X whole-GPU basis $62.5/GB ($12,000 ÷ 192 GB) vs FX100 fully populated $0.273/GB (188,744 GB) — a ratio of ≈228.8×.

Honest statement: HBM and all-flash differ by orders of magnitude in bandwidth/latency and are not interchangeable; this ledger only claims warm-context absorption (R5 HBM-efficiency methodology).

Cluster ledger: 128-GPU pilot (joint-test scale)

16 nodes of GPU investment ¥25,600,000; 8 arrays at ¥2,969,600 — storage takes only 11.6% of GPU spend for a cluster-wide +30% effective throughput (conservative mainline anchor).

The full TCO model (network / power / depreciation / Monte Carlo sensitivity) is reproducible Python after NDA.

Verify this math in a joint test View the test reports

This site presents business-cooperation information and constitutes neither an investment offer nor any promise of returns. Measured data come from signed / official test reports (see the Evidence Library); vendor specs, public sources and estimates are labeled as such.