Data Center Liquid Cooling Simulator · User Guide v1
Live: https://sim.nextgenergy.ai Catalog: https://sim.nextgenergy.ai/catalog/ Submit a product: https://sim.nextgenergy.ai/catalog/submit/
Version: 2026-10-29 (v1.27.0). The page opens in English; "中 / EN" in the header switches language.
1. What it is, what it is not
What it is: a browser-based simulator of the whole liquid-cooling thermal path. It models "ambient air → dry cooler / tower / chiller → CDU heat exchanger → secondary loop → cold plate → chip" segment by segment under energy conservation, returns junction temperature, supply/return temperatures, flow, pump power and pPUE, and judges "does this case hold" against the four-gate acceptance rules. It can also fold verified real CDU / dry-cooler / cooling-tower models to your operating point and answer "is this unit enough".
What it is not: every number is a model value, for comparison and pre-selection, not an equipment guarantee. It is not CFD (no in-rack flow distribution or hot spots), not vendor selection software (vendor curves take precedence), not an acceptance tool, and it does not represent any OCP or ASHRAE certification. Final selection, purchase and acceptance rest on vendor data and site measurements. This sentence sits at the top of the page, on every result card and in the report.
2. Three minutes to a first result (four-step wizard)
The wizard opens on first visit (returning users reopen it with "Wizard" in the header):
- Platform and racks: platform (GB300 NVL72 / GB200 / VR200 / H100 / MI300 / CPU DLC / immersion / air-cooled control), rack count, racks per row, load.
- Site and heat rejection: city (sets design dry/wet bulb, altitude and the weather-year bins), rejection method (dry cooler / tower / adiabatic / full mechanical), architecture (In-Row = one CDU per row, or central), redundancy.
- Supply and ΔT: supply temperature (the slider shows the platform inlet limit and the ASHRAE S class; from v1.18.0 the secondary supply carries an S class and the primary supply a W class, matching the ASHRAE Liquid Cooling Guidelines rule that W classes govern the FWS and S classes the TCS), ΔT, coolant, CDU approach.
- Result: after a 900 s run to steady state — junction temperature, supply, pPUE, CDU count, fan, trim cooling, and any fatal condition.
"Open the simulator" writes the four steps into the left column; everything else lives in the collapsible groups there. "Skip" goes straight to the left column with the same effect.
3. Interface map
- Header: app icon + site name; icon buttons (name on hover): Feedback (in-page, LinkedIn link inside) · Copy link · Copy summary · Wizard · Get report · Help "?" (the three-minute start in a page panel, with a link to this guide); language key (shows the current language, one click toggles) · units key (SI / IP likewise) · Free / Enterprise badge · "Sim time" (time elapsed inside the model at the chosen speed, not wall-clock). At ≥ 1400 px the icons carry labels.
- Left column (parameters), nine groups: Equipment & load | Secondary loop (TCS) | CDU & primary loop (with "pick a CDU from the catalog") | Secondary pipe screening | Heat rejection & climate (with "pick a rejection unit from the catalog", altitude) | Air-side management | Boundary & acceptance | Scenarios (eight buttons + speed + pause/reset) | N-1 failure what-if (three buttons).
- Centre (stage): three icon segment controls — schematic / 3D, zoom (fit / standard / large; "fit" on first visit, the level you pick is remembered in this browser), size (default / large / full screen), current segment highlighted, name on hover; label tri-state; tabs below: Temperature trend | Six boundaries | Temperature budget | Four gates | Sizing | Annual | TCO | Climate map. The four-gate verdict sits at the right end of the tab bar.
- Right column (readings): ten KPI tiles (junction, case, secondary supply / return, flow, specific flow, cooling power, pPUE, IT power, instantaneous PUE estimate), each with a hover note on its definition; tap the "?" beside a reading for an in-page note (you stay on the current case) → temperature chain → hydraulics → power → air side → diagnostics (alerts and red lines) → View assumptions.
- Phone: four switch buttons just below the header — Diagram / Charts / Inputs / Readouts (the "inputs panel" in the wizard and Help is the Inputs view); the six-boundaries table scrolls sideways and the B6 note sits below it; the schematic defaults to "fit", pinch to zoom.
4. Left column: what to change, what to watch
Every slider recomputes the design point at once and the transient model then settles toward steady state (clock in the header, speed 1–120×). Wait for the right-column KPIs to settle before reading, or press "Reset" to restart from the design point.
Common traps: ① after picking a platform preset, changing "Chip TDP" or "Rack power" triggers the chip-vs-rack power check; over 100% means contradictory inputs and the four gates fail; ② "IT load" also drives the annual profile; ③ "Loop fluid volume" only changes how fast transients move, not the steady state.
Secondary coolant and brand label (from v1.25.0): under "Coolant" there is now a "Brand label" drop-down listing the 12 PG25 coolants on the NVIDIA DSX Ready list (the list is maintained by NVIDIA; this tool only cites it and has not tested these products; the date the list was read is shown in the note). A brand is only a label and carries no property set of its own: picking a brand picks the property curve it maps to (Dynalene maps to the "generic PG25 curve", the other 11 to the "data-centre PG25 curve (based on DOWFROST LC 25 properties)"), the model only uses the curve, and the readings are identical to picking that curve directly. The note gives the product's rough deviation from the curve; 7 products have no public temperature-dependent datasheet and are marked "unverified" in the drop-down, 1 has a single data point and is marked "partly verified". Being on the list is not an endorsement and not a certification. Share links keep the property curve only, not the brand, so the brand is empty when someone opens your link; loading a template, opening a share link or running the wizard clears the brand. Immersion machines use non-PG25 fluids, so the brand drop-down is disabled there.
5. Picking a CDU from the catalog
Where: left column "CDU & primary loop" → the catalog dropdown (grouped by vendor; only foldable models are listed, the last greyed line says how many cannot be folded).
Input: the model itself. Its rated capacity is folded to your current approach (Q ≈ UA·ATD linear approximation) and then capped by the vendor's maximum flow (max flow × ρ·cp × ΔT); its internal pressure drop and pump head are written into the model. Changing approach, ΔT, coolant, rack count or architecture re-folds automatically.
Reading the result: the check card at the top of the "Pipe & capacity screening" tab — rating wording as published, folded line, flow-cap line, three-state verdict (Not enough / Enough with margin under 15% / Enough), "ΔT must be at least X K for this unit to carry the load", supply range, datasheet link (the catalog records no prices). Central architecture shows how many units are needed. On "Not enough" a red line appears in the right-column diagnostics; click it to jump to Sizing.
Why the rating wording matters: vendors rate at different conditions (Vertiv at 4 K approach, Boyd at 8 K, Motivair at the equivalent of 12.8 K). Two units both labelled 1,350 kW can differ by a factor of two at your 4 K. A ratio outside 1/3–3× is flagged "extrapolated"; a dielectric fluid flags "cap unreliable" — do not conclude while either flag is showing.
6. Picking a heat-rejection unit from the catalog
Where: left column "Heat rejection & climate" → the catalog dropdown (the list follows the rejection method: dry coolers or towers).
Result: a check card only; it is not fed into the solver. Models whose rated fluid-inlet and ambient temperatures can be parsed are scaled linearly by ITD to the site design condition (±20%) and compared with the design rejection duty; the rest show the original wording with "not comparable". The chiller catalog is not connected yet. The selected id is written to the URL and survives a refresh.
7. Scenarios (eight buttons)
Where: left column "Scenarios". A line under the buttons says what the current scenario does; pause/reset sit next to it.
| Button | Simulates | Watch |
|---|---|---|
| Steady | design point running on | baseline readings |
| Load step | 30% ↔ 100% every 5 min | swing of junction temperature and flow |
| Heat wave | ambient +14 K over 40 min, then back | when trim cooling cuts in after fans max out |
| Pump failure | one pump lost: secondary flow drops to (n−1)/n; with a standby it recovers after the "changeover time" | ΔT widens, junction rises |
| Half fans lost | half the rejection fans fail | how much the chiller must step in |
| Plate fouling | cold-plate resistance creeps up | fluid temperatures look normal, junction climbs — the easiest miss |
| QD contamination | branch resistance spread grows | worst-branch junction |
| Flow loss | pumps stop | seconds to the junction limit (ride-through) |
Trap: fouling and QD contamination clear the four-gate records of the affected boundary and do not restore them when the scenario ends; tick them again.
8. N-1 failure what-if (three buttons)
Where: left column "N-1 failure what-if": lose one CDU / lose one rejection unit / lose one secondary pump, plus "Clear N-1". Button wording follows the redundancy setting: with standby it reads "lose one, standby takes over in N s", without it "one fewer, permanently", and In-Row reads "this row loses its only CDU" (same path as flow loss).
Result: applied live to the simulation, plus a rehearsal (settle 900 s, then the failure for 1800 s) giving several groups of figures: the fault domain (racks affected / liquid-cooled load), supply to the failed row and to the other rows, site service load (nominal / solved, shown separately), remaining rated capacity of the online CDUs, peak/steady junction, seconds to the junction limit, and recovery time (since v1.22.2 there is no longer a single "remaining capacity %"). Zero CDUs online or an overheated row becomes a fatal item in the four gates.
Traps: the changeover time comes from the "Standby switchover time" slider, not from the catalog; when the catalog does not state pump redundancy it is treated as none and labelled "assumed"; In-Row assumes no interconnecting header.
9. Share link
"Copy link" in the header first lists what is and is not included, then packs all 61 inputs and the selected catalog models into the address bar (#s=). It copies the current design inputs: whoever opens it starts the simulation from those inputs (no wizard). It does not include triggered N-1 events or rehearsal results, a running scenario or the simulation time, transient history such as the temperature trend, gate items ticked by hand, or a pinned model / catalog version. The link is also the input for "three scenarios side by side" on the Annual tab and "Scenario B" on the TCO tab. A code that cannot be decoded falls back to the default case with a notice at the bottom of the page; a single out-of-range, invalid or fractional-count value is clamped / rounded / reset to its default, and the page says how many items were corrected.
10. SI / IP
"SI / IP" in the header converts display only: L/min ↔ GPM, kW ↔ ton, °C ↔ °F, kPa ↔ psi, m/s ↔ fps, mm ↔ in. Sliders stay SI, the model is all SI, and switching changes no result. Catalog entries rated in IP units show the converted value alongside on the check card.
11. Altitude and supply red lines
Altitude: left column "Heat rejection & climate" → altitude slider (follows the city library when the location changes). It sets atmospheric pressure, air density, wet bulb and fan power — at 3,000 m the fan power for the same duty roughly doubles.
Red lines (red text + fatal item, inputs are not refused): supply above the platform inlet limit (45 °C for GB/VR); return above inlet limit + 20 K; primary fluid freezing at ambient; wet-bulb formula out of range for evaporative units. The slider shows the ASHRAE W17/W27/W32/W40/W45 boundaries for reference against the platform class.
12. Annual profile
Where: "Annual" tab. Inputs: electricity price, water price, cycles of concentration, freeze dry-mode threshold (load follows the left column).
Method: each 2 K dry-bulb × wet-bulb weather bin is solved at steady state, power × hours. Twenty-four cities carry a typical weather year (Toronto, Montreal, Vancouver, Halifax, Ashburn, Dallas, Phoenix, Santa Clara, Beijing, Shanghai, Singapore, Dubai, Albuquerque, Bengaluru, Berkeley, Bristol, Chicago, Delhi, Golden, Lugano, Oak Ridge, Ottawa, San Francisco, St. Louis); other cities and any map point are synthesised from seasonal values and labelled "coarse".
Output: free-cooling hours, trim hours, fan/pump/chiller energy, annual pPUE, water (evaporation / blowdown / drift), WUE, WUI score (cities without Aqueduct data are not scored), electricity and water cost.
Side by side: paste two more share links into B / C; Δ = scenario − A. Trap: the design point uses the design dry/wet bulb, the annual profile uses a typical year — not the same data; the annual profile has no transients and no N-1.
13. TCO
Where: "TCO" tab. City defaults for electricity, water and sewer carry their sources (missing ones show "not recorded" for you to fill); maintenance 3%/yr, discount 8%, horizon 5/10 years are editable.
Scenario A CapEx is a range you enter yourself (the catalog records no prices; shown as "not recorded" until entered); Scenario B = same site with the other rejection type / other architecture / a pasted share link. Output: CapEx, annual electricity and water, maintenance, N-year total, present value, and B vs A ΔCapEx, annual saving, simple/discounted payback, NPV — all as ranges, never a single value, never "saves X%". Excludes civil works, piping, installation, labour, tax and residual value.
14. Performance curves
At the bottom of the "Pipe & capacity screening" tab. Rejection unit: capacity and fan power at inlet ±5 K × load 25/50/75/100% (linear ITD scaling from the catalog rating + fan law, ±20%); CDU: effective capacity vs ΔT 6–15 K with the flow-cap knee ΔT* visible (below it the maximum flow limits). For comparison only; vendor software curves take precedence.
15. Four gates and six boundaries
"Six boundaries" tab: facility water interface → heat rejection → CDU → manifold and quick disconnects → hoses → cold plate; each boundary carries the one question it must answer, reachable by clicking the numbered badge on the schematic or the equipment in 3D. "Four gates" tab: 18 criteria, 5 computed and judged by the model (material temperature margin, pressure test, flow balance, approach, baseline), 13 are records you tick. Three-state verdict: Model met · records confirmed / Records open / Design condition not met. Any input change voids all records — deliberately. "Temperature budget" is the waterfall from ambient dry bulb to junction; the segments sum exactly to the junction temperature, so you can see where the budget goes.
16. Report
Where: "Get report" in the header → the first time you enter name, email and phone (company optional) and pass the human check → the report unlocks on the page as "Report preview"; "Download PDF" in the toolbar uses the browser print dialog. The same browser is not asked again for 30 days. Step 4 of the wizard uses the same entry. Switch language first.
Content (the new customer report from v1.26.0): paged A4, in this order — cover (report number, input fingerprint, model version, review status and a one-line status bar: objective verdict, record progress, number of fatal conditions, domain state), document control and contents, 01 executive summary (domain notice at the top, then six key numbers, four gates and the junction limit, fatal conditions, "how to fix", efficiency rating eligibility, model notes and suggested checks), 02 design basis and system boundaries (topology sketch, inputs, objective summary, six boundaries and usual owners), 03 thermal and hydraulic analysis (design-point table, head budget bar), 04 equipment capability and performance curves (catalog folding, end-unit and CDU curves), 05 annual energy, water and economics, 06 failure rehearsal (N-1 overview, key case and event timeline), 07 verification and test recommendations (judgement table for the 18 gate items, FAT / SAT / IST recommendations, blank test record and issue-closure tables), Appendix A operating snapshot, pressure-test boundaries and model notes (with the non-thermal checklist), Appendix B source register, Appendix C delivery status. Every page carries a header, footer, "for reference only · model values" and a watermark; the page count depends on the case (default case about 24 pages in Chinese, about 26–28 in English depending on the font). The summary edition (tiering on, no access code) has no chapters 05 and 06 and no performance curves.
Prices and economics: the report prints no catalog price ranges; chapter 05 gives annual electricity, water and operating cost only. CapEx, present value and payback (as ranges) are printed only when you have filled in all four CapEx fields on the TCO tab yourself.
Tests are run by the project team: chapter 07 holds test recommendations generated from the criteria and blank record tables; acceptance limits, load conditions and test methods are set by the project team from codes, drawings and vendor documents. NextGenergy does not run, attend or sign these tests, and the report has no signature block.
Fingerprint and report number: the fingerprint is computed from the share code (taken in SI), the current scenario, the current run time (rounded to the second), the failure-rehearsal state, the domain state, the four-gate ticks, the coolant brand and the entries on the TCO tab (electricity price, water price incl. sewer, maintenance share, discount rate, years, heat price, IT-side saving, scenario B comparison, user-entered CapEx when all four cells are entered by the user); it does not change with language or units, and any change to these items changes the number.
If the new layout fails: if the new report fails while gathering data, laying out or paginating, the whole report falls back to the legacy layout and the preview toolbar shows "The new report layout failed to generate; the legacy report is shown instead".
Previous report layout: add ?rpt=legacy to the address to get the v1.25.0 layout (kept during the transition).
Note: the report is for reference only and carries no certification wording. If the mail service is configured, a link to the report is also emailed to you; the page says "sent" only when it really was.
17. View assumptions (S / D / E)
"View assumptions" at the bottom of the right column lists all constant tables (the panel title gives the current count) with a class and a source: S = standard or vendor-published value (source given); D = derived from physics (script and assumptions given); E = estimate with no public source (basis and range given). The platform library only states "derived by NextGenergy and calibrated against public measured points". Check this panel before quoting a number in a proposal — E-class values must be marked as estimates.
18. Submitting a product to the catalog
Entry: https://sim.nextgenergy.ai/catalog/submit/ . Two routes: a single-entry form (fields and a specs sub-form by product type) or a CSV batch (download the template; rows are validated in the browser, faulty rows are not submitted and can be downloaded, fixed and re-uploaded; up to 50 rows per upload). A datasheet link and the rating condition are required (CDU: approach, supply temperature, primary water temperature).
Flow: submit → status pending → NextGenergy checks the datasheet → verify → only then does it appear in the public catalog and the simulator dropdowns. Rule: public specifications only; the catalog does not take prices (price columns in a CSV are ignored), and quotes, discounts and contact names in text fields are refused.
19. Free and Enterprise tiers (live since 2026-09-25)
Enterprise features show a lock badge; opening one offers "Enter access code" or "Request Enterprise" (the same contact form). Access codes are issued by NextGenergy and expire; this is a front-end soft lock, not a security boundary, and the access code is the actual authorization.
| Feature | Free | Enterprise |
|---|---|---|
| Design-point simulation, schematic / 3D, eight scenarios | ✓ | ✓ |
| Four gates, six boundaries, temperature budget, diagnostics and advice | ✓ | ✓ |
| Catalog selection and check (CDU / rejection), performance curves | ✓ | ✓ |
| Wizard, share link, SI/IP, altitude and red lines | ✓ | ✓ |
| Paged report | unlocked after contact details; summary edition (no 05 annual and economics, no 06 failure rehearsal, no performance curves) | unlocked after contact details; full edition (with chapters 05, 06 and performance curves) |
| Annual profile (24 typical years + coarse) | — | ✓ |
| Three scenarios side by side | — | ✓ |
| TCO (range NPV, payback) | — | ✓ |
| N-1 failure what-if | — | ✓ |
| Catalog non-thermal checklist (20 items) | — | ✓ |
| MCP / API access (planned; pilot on request) | — | ✓ |
| AI brief reader (PDF → parameters → sanity assessment, 5 reads per email per day) | — | ✓ |
| Priority support (reply within one business day) | — | ✓ |
The free tier is enough to answer "is this unit enough, does this case hold"; the enterprise tier is for those who decide, write proposals and compare options. Pricing is in the separate pricing memo.
20. AI brief reader (from 2026-09-25)
Where: the "Brief" tab in the middle column (right of "Pipe & capacity screening"). Enterprise feature; it also needs two tokens: a contact token (the contact form pops up the first time; reads are counted per email, 5 per day) and an access-code token (enter your code once in the "Access code" box on this tab; users who unlocked earlier must enter it once more — the page does not store the code).
Steps: ① read the note at the top and tick consent (once per browser session); ② upload a PDF (click or drop onto the dashed box; ≤ 10 MB, ≤ 50 pages, must have selectable text) or switch to "Paste text" (≤ 40,000 characters); ③ pass the human check and press "Read the brief" — usually 30–90 s; ④ check the result table: each row has the extracted value (editable), confidence (green / amber / red), source (page · quote), and any conversion as "⇄ 104 °F → 40 °C"; rows marked "not given" are highlighted with the value that will be used in grey; the checkbox on the right decides whether a row is loaded; ⑤ press "Load into simulator": the parameters go into the left column through the same checks as a share link (out-of-range values clamped, illegal values back to default, every change listed), the address bar becomes this case's share link, and after a 900 s steady run the "Sanity assessment" appears.
The four blocks: A internal contradictions (from the model); B brief targets vs model values (pPUE, return temperature, junction margin, CDU count, redundancy; WUE and free-cooling hours from the Annual profile); C fatal items with "How to fix it" cards (apply in one click, then it re-runs); D catalog match for the CDU named in the brief (a match is checked with its folded catalog capacity; several matches let you pick). All are lumped-parameter model values for comparing options and pre-selection only.
Privacy: the PDF is sent to Anthropic's Claude API and, by default, is not used for training; this site does not store the PDF itself, only the extraction result for 90 days. Remove contract prices and personal data before uploading. Scanned PDFs (no text layer), encrypted PDFs and files over 50 pages are stopped in the browser and never sent.
Note: this is extraction by a large language model and can miss or misread values — check the table row by row before loading. The offline copy opened from disk cannot use this feature (the button reads "Needs the online version").
21. FAQ
- Results differ from the vendor's selection software? Expected. This tool uses first-order scaling (±15–20%); vendor curves take precedence; if the gap is large, read the rating wording first.
- Why does pPUE differ by only 0.002 between In-Row and central? Most secondary-side pressure drop is in cold plates, hoses and quick disconnects; the header is a small share. Pump power drops, pPUE barely moves — the number is right.
- All four-gate items went grey after a change? Any input change voids the record-type items; confirm them again.
- VR200 at 8 racks per row with a 1.3 MW CDU says "sizing does not hold"? A row is about 1.65 MW and In-Row is strictly one CDU per row; pick the 2.5 MW band or reduce racks per row.
- Can I trust the seconds-to-limit on flow loss? It depends on the "cold plate and branch holdup" input in the left column (default 2 L/rack, class E); the order of magnitude is right, the seconds need site calibration.
- Why is the annual profile labelled "coarse"? No typical weather year for that city; it is synthesised from seasonal values. Pick one of the 24 cities with a typical year for a full result.
- Catalog dropdowns are greyed when opened locally? The catalog loads from the same origin online; offline, everything else still works.
- 3D does not open? three.js loads on demand from cdnjs; offline or blocked it is unavailable; the schematic and the model are unaffected.
- Why does the catalog show no prices? The catalog takes public specifications only, not prices; on the TCO tab you enter CapEx bounds yourself.
- Can I cite it in a proposal? Yes, as a "model estimate" with the share link and the assumption list; never as an equipment performance commitment, and never as OCP / ASHRAE certification.
22. Privacy and anonymous usage statistics (from 2026-09-25)
Anonymous session statistics: which pages and tabs were opened, which parameter names were changed and how often, and whether share / report / catalog were used. No operating-point values, no IP, no email. The session id is a random 16-digit number generated by your browser and kept in the current tab only (closing the tab ends it; no tracking across sessions or devices); device is only phone / computer, location only country. The offline copy opened from disk and any non-https page send nothing.
Also recorded: the referring site's domain (not the full URL), interface language, country (Cloudflare's two-letter code) and device type (mobile / desktop). Email addresses and phone numbers in feedback text are replaced with [email] / [phone] in the browser. The site also runs Cloudflare Web Analytics (cookie-free page-view counts).
Turn it off: run localStorage.setItem('lcs_track_off','1') in the browser console; this browser stops sending. Remove the item to turn it back on.
Feedback: "Feedback" in the header opens a small window: 👍 / 👎 plus a note (≤ 500 characters), sent anonymously to the author. Please do not include contact details (anything that looks like an email address is replaced with [email]). If you want a reply, use the LinkedIn link in the same window.
Leads (report / enterprise requests), catalog submissions, access codes and the AI brief reader keep their own records as described in sections 16, 18, 19 and 20; the statistics do not duplicate them.