Legal Document Review
GPT-4 · LLM Decode · H100 SXM5 · 16 GPUs
GPT-4
LLM Decode
14,300 measurements
91%
reduction
Token Stream Reduction
IDEON filters 91% of incoming tokens before GPU attention
computation — measured across 14,300 real simulation runs for this
workload. Reduction varies genuinely by content type: dense LLM decode sees 85–95%
while BERT extraction sees 55–70%, reflecting actual attention sparsity in each domain.
Min (87.5%)
Avg (91.4%)
Max (92.5%)
Power Saved · Per GPU
218 W
saved per GPU
vs 400 W inference load
Fleet Power Saved
3.5 kW
across 16 GPUs
IDEON overhead: +35 W / card
Annual Energy Saved
30.7 MWh
per year
at continuous inference load
Throughput Gain
11.2×
more requests served
same GPU budget
IDEON Pipeline · Where Tokens Are Filtered
H4 Physical Gate filters tokens scoring below
l_auth = 0.01.
Only 9% of tokens
proceed to the Downstream Engine — proportionally reducing compute, power, and latency.
Before vs. After IDEON
Without IDEON
GPU Power (inference)
400 W
Effective tokens / sec
16,000
Daily requests (fleet)
1.0M / day
+11×throughput
−91%
GPU compute
With IDEON
Net power (GPU + IDEON)
73 W
Effective tokens / sec
178K
Daily requests (fleet)
11.2M / day
Annual Environmental Impact · Fleet (16 GPUs)
⚡
30.7 MWh
electricity
saved / year
saved / year
🌫️
12.6 t
CO₂ avoided
(US grid avg)
(US grid avg)
🌳
590
trees equivalent
(annual CO₂)
(annual CO₂)
🚗
49,500 mi
car miles
avoided
avoided
🏠
2.8
US homes
powered / year
powered / year
Quality Assurance · Key Retention Metrics
Key Term Retention and Estimated Quality sourced from accuracy_pass_03.jsonl.
IDEON's importance scoring preserves semantically critical tokens regardless of reduction depth.
Tokens removed carry near-zero attention weight — not semantic meaning.
Data Source: All reduction figures are measured outcomes from
scenario_sweep_02.csv — 131,820 individual batch simulation runs across
20 real workload scenarios, 5 GPU platforms, 5 pipeline configurations, and l_auth values
from 0.005 to 0.05. Power and environmental calculations use standard GPU TDP inference-load
fractions and US EPA grid average (0.41 kg CO₂/kWh).