▸ CEH™ Standard · Benchmark Index v1.0

CEH™ Benchmark Index

Full ten-configuration LLM inference benchmark. Workload: Llama-class, batch 8, vLLM. PUE 1.2. Grid $0.085/kWh. Sources: MLPerf v5.1, CUDO, Koyeb, Spheron.

📊
Full Index · Coming Shortly
Detailed Benchmark Index
The full interactive benchmark index — including all ten hardware configurations, bar chart comparisons, grade scale visualization, and downloadable data — is being prepared for this page. The benchmark data is available now in the CEH™ Whitepaper v1.0.
Download Whitepaper →
▸ CEH™ Index v1.0 · LLM Inference · Full Table
Hardware
CEH™ Rate
Grid $/M tok
BTM $/M tok
Grade
NVIDIA B200
Blackwell · 2025
2.44×10⁻⁷
$0.02
$0.01
S
Google TPU v4
Google · 2023
4.27×10⁻⁷
$0.04
$0.01
A
NVIDIA H200
Hopper · 2024
4.83×10⁻⁷
$0.04
$0.02
A
NVIDIA L40S
Ada Lovelace · 2023
5.17×10⁻⁷
$0.04
$0.02
A
NVIDIA H100 SXM
Hopper · 2023
6.20×10⁻⁷
$0.05
$0.02
A
AMD MI300X
CDNA3 · 2024
7.32×10⁻⁷
$0.06
$0.03
B
NVIDIA A100 80GB
Ampere · 2021
7.62×10⁻⁷
$0.06
$0.03
B
NVIDIA RTX 4090
Ada Consumer · 2022
1.20×10⁻⁶
$0.10
$0.04
C
NVIDIA V100
Volta · 2018 · Legacy
2.14×10⁻⁶
$0.18
$0.07
D
CPU Cluster (EPYC)
x86 Baseline · No Accelerator
6.11×10⁻⁶
$0.52
$0.21
F

PUE 1.2 applied uniformly. Grid = $0.085/kWh (EIA 2025). Representative BTM rate = $0.035/kWh (source-agnostic benchmark range; actual rates vary). Carbon: 0.386 kg CO₂/kWh (EPA eGRID 2024).