Compute Resource Explorer

Where does your energy actually go?

Select an application, model size, and serving mode. This tool shows the physics-level energy breakdown of AI inference — specifically, how much energy is consumed by moving data vs. performing computation. Every number is cited.

Application
Model Size
Serving Mode
Data Center
Unlimited thermal headroom. GPU is viable here — CIM reduces TCO.
Thermal budget: 300 W

Energy Anatomy: Per Token

Where does the energy go for a single inference token on a 135M-parameter model?

GPU Architecture 648.0 µJ / token
99.6% data movement
Weight fetch from external memory Useful compute (0.4%)
BSA-CIM Architecture 32.4 µJ / token
100% compute — zero data movement
Weights resident in RRAM — never fetched, never refreshed

99.6% of GPU energy is spent moving weights from external memory to compute units. CIM eliminates this entirely — resulting in ~20× system-level energy efficiency for edge single-stream workloads.

Weight Memory
GPU
135 MB
External DRAM/HBM (requires refresh power)
CIM
135 MB
Embedded RRAM (zero standby power)
Energy Ratio
~20×
CIM is ~20× more energy-efficient
Modeled: batch-1 edge serving, foundry position paper
Estimated Power
GPU (Mobile NPU class)
2 W–5 W
CIM (derived from modeled ratio)
100 mW–250 mW
Fits Data Center?
Thermal budget: 300 W
GPU
✓ Fits
CIM
✓ Fits
Manufacturing Cost Context
GPU (3nm–5nm EUV)
Wafer: ~$16,000–$20,000
Requires TSMC or Samsung
Subject to export controls
CIM (22nm DUV)
Wafer: ~$3,000–$5,000
20+ foundries worldwide
Zero export restrictions
CIM Packaged Die
~$22 at 120mm² (modeled)
RRAM adds ~15% mask cost
Source: 10yr datacenter cost model
Sources & Methodology

DRAM access energy: 20 pJ/bit — Horowitz, "Computing's Energy Problem," ISSCC 2014

7nm ALU switching energy: ~0.01 pJ/op — same source, extrapolated to 7nm

GPU system-level pJ/op: 0.5–1.0 pJ/op — derived from published A100 (300W / 312 TOPS) and H100 TDPs at peak utilization

CIM system-level pJ/op: 0.120 pJ/op (weight-dominated) — modeled, BSA-CIM energy ledger, SmolLM2-135M workload shapes (weight distribution R̄=2.273 is measured; the pJ figure is modeled from it)

Energy ratios: ~20× (batch-1 edge), 4–8× (batched server), 4–8× at long context for modern serving models (sparse-attn / MLA / hybrid-SSM keep the weight matmul dominant, so the advantage holds; naive dense full-attention converges to parity) — modeled system-level analysis, foundry position paper

Data movement share: 99.6% (batch-1), 98.6% (batched) — derived from GPU system energy decomposition

GPU reference power: Approximate, based on published NVIDIA Jetson platform TDPs for edge inference and A100 for server

CIM power estimates: Derived by applying modeled energy ratios to GPU reference power. No silicon measurements exist — these are modeled estimates, not hardware specifications.

This calculator shows physics-level energy analysis, not product specifications. CIM power estimates are modeled from the energy ratios applied to published GPU references. Actual hardware performance will vary. The core insight — that data movement dominates AI compute energy and CIM eliminates it — is a measurable physical fact independent of any specific product.