Why benchmark?

A great metric in isolation is meaningless. Benchmarking puts your design next to known references so 'good' has a number. CrystalSim's Benchmarks module compares your saved designs against published silicon (Apple M4, NVIDIA H100, Intel Lunar Lake) and against research devices (Tokyo InGaOx GAA, Georgia Tech piezotronic).

Key Concept

TOPS (Tera-Operations Per Second)

Trillions of basic operations (typically INT8 multiply-accumulates) per second. The headline number for AI accelerators. Apple M4 NPU ≈ 38 TOPS; NVIDIA H100 ≈ 1979 TOPS.

Key Concept

TOPS/W (operations per watt)

Compute throughput divided by power. The single most important number for mobile and edge AI. Targets: > 10 TOPS/W for next-gen mobile, > 30 for edge inference.

Key Concept

Performance per area

Throughput per mm² of silicon. The number datacenter operators care about — it sets the per-rack ceiling.

Match metric to application

- AI workloads (datacenter): TOPS, perf/area, memory bandwidth. - Mobile / edge AI: TOPS/W first, then TOPS/mm². - High-performance computing: raw FLOPS, sustained perf/W under load, thermal headroom. - IoT / sensors: TOPS/W and quiescent power dominate; absolute throughput barely matters.

Rule: define the application before opening the benchmark page. Otherwise you'll cherry-pick the metric that flatters you.

Exercise — Crystal-EM vs the world

1. Open the Chip Builder, load the Crystal-EM Hybrid template (Templates menu). 2. Run the full test suite in the Test Lab. 3. Open Benchmarks. 4. Stack your design up against Apple M4 and Tokyo InGaOx GAA. 5. Read the radar chart and table side-by-side.

Try It Yourself

Step 1 — load Crystal-EM template

Open the Chip Builder, then Templates → Crystal-EM Hybrid.

Open Chip Builder

Try It Yourself

Step 2 — run full test suite

Test Lab → run every Material, Component, Integration, and Full-Chip test.

Open Test Lab

Try It Yourself

Step 3 — compare in Benchmarks

Open the Benchmarks page, add your design + Apple M4 + Tokyo InGaOx GAA, switch between bar / radar / table views.

Open Benchmarks

Checkpoint · +5 XP

What's the honest way to read a benchmark result?

Reading the result

Crystal-EM should typically win on TOPS/W (resonant gating + lower V_dd) and lose on absolute TOPS until volume manufacturing matures (low TRL + small dies). Use that pattern in your write-up: 'Crystal-EM wins where energy is the bottleneck, loses where raw throughput dominates.' That sentence is honest, defensible, and points to the next R&D step.

Lesson Summary

Use the Benchmarks module to put your design next to real silicon.
Match the right metric to the right application.
Read where Crystal-EM wins and where it still loses.
Treat benchmarks as questions, not answers.

Test Your Knowledge · +80 XP

TOPS/W is most important for which application?

Performance-per-area matters most to:

What should you decide BEFORE opening the Benchmarks page?

An honest benchmark write-up:

Where does Crystal-EM most likely BEAT mature silicon?