0.3186Best bandwidth/energy proxyUNKNOWNInspect metric contract
4-high / 1.6 TB/s / enhancedBest caseUNKNOWNInspect metric contract
18 casesSweep rowsUNKNOWNInspect metric contract
SIMULATION / SA-MEM-003-HBM-THERMAL-ROOFLINE-001
HBM thermal bandwidth roofline
HBM Roofline
Replay stack height, target bandwidth, and cooling class to see usable bandwidth, energy per bit, and throttle risk together.
EXPLORE
Start with the observable.
Use this page when the simulation is the primary artifact. The linked lesson explains the concept; the linked lab keeps the visual replay surface available.
- Do not read HBM bandwidth without energy per bit and thermal headroom.
- Compare target bandwidth and usable bandwidth under throttle risk.
- Use stack height as a package/system knob, not just a capacity slogan.
RUN CARD
Conditions before conclusions.
This page replays saved HBM roofline data and does not run thermal CFD, package extraction, or vendor HBM models.
Primary knobs
- Stack height
- Target bandwidth
- Cooling class
Saved outputs
- Power proxy
- Thermal headroom
- Usable bandwidth
- Energy per bit
EXPORT
Download the public artifacts.
These exports are the supported public files for review. PDK files and restricted third-party material are not redistributed.
Dataset JSON
/data/hbm-thermal-bandwidth-roofline-web-v1.json
Open file ->Dataset CSV
/data/hbm-thermal-bandwidth-roofline-v64.csv
Open file ->v64 deep simulation summary CSV
/data/v64-deep-simulations-summary.csv
Open file ->v64 deep model script
/data/v64_deep_semiconductor_sims.mjs
Open file ->v64 deep model card
/data/v64-deep-simulations-model-card.md
Open file ->METHODS
How the number was made.
Methods are written for citation discipline: the extraction rule matters as much as the plotted value.
Power proxy scales with target bandwidth and stack height.
Keep this method attached when reusing the figure or metric.
Thermal headroom penalizes higher power and improves under the enhanced cooling label.
Keep this method attached when reusing the figure or metric.
Usable bandwidth is reduced by a throttle-risk factor when headroom is low.
Keep this method attached when reusing the figure or metric.
LIMITS
What this page does not claim.
The MVP is useful because the limits are visible. These statements prevent a teaching simulation from being cited as silicon evidence.
No vendor HBM stack, interposer, TSV, microbump, package, CFD, or workload measurement is included.
Carry this boundary into any derivative note, paper draft, or slide.
No product comparison, procurement, or AI-system performance claim is made.
Carry this boundary into any derivative note, paper draft, or slide.
Real HBM work needs package thermal data, workload traces, memory-controller data, and measurement.
Carry this boundary into any derivative note, paper draft, or slide.
EVIDENCE
Trace the claim back to artifacts.
The evidence route remains the catalog-level browser; this page is the simulation-level reading card.
v64 deep results store every stack height, bandwidth, cooling, power, and headroom row.
SA-MEM-003-HBM-THERMAL-ROOFLINE-001
CSV export preserves the thermal roofline table.
SA-MEM-003-HBM-THERMAL-ROOFLINE-001
The memory hierarchy lesson separates HBM system role from DRAM cell physics.
SA-MEM-003-HBM-THERMAL-ROOFLINE-001
Source spine
These links anchor the public teaching model. They do not convert the replay into a qualified process recipe or signoff result.
Ramulator2External MIT-licensed DRAM/HBM simulator reference; v64 pages use original teaching data.DRAMsim3External MIT-licensed DRAM timing and thermal simulator reference.IEEE IRDSRoadmap source spine for memory hierarchy and emerging memory vocabulary.Google Security RowHammer research notePublic context for RowHammer as a DRAM charge-leakage, refresh, and security-evidence topic.
Public execution boundary
No server-side live ngspice execution, uploads, accounts, payments, comments, newsletter signup, or job/event submissions are active on this page.