The job queue for a research GPU cluster: sixty-four accelerators across eight nodes, eleven jobs running, nine waiting, and one that will never start.
The second brand specimen, and the opposite end of the corpus from tesla-brand. Where that reference publishes nine colours and four type sizes, this one publishes thirty-one colours and an eighteen-step ramp with front matter naming every token.
Where the style comes from
Transcribed from awesome-design-md/design-md/nvidia/DESIGN.md, read 2026-09-16, and emitted by:
python3 tools/brand_tokens.py nvidia-brand > systems/nvidia-brand/tokens.css
python3 tools/brand_tokens.py nvidia-brand --check
python3 tools/brand_tokens.py --self-test
| Decision | Transcribed |
|---|---|
| Ramp | #ffffff #f7f7f7 #cccccc #a7a7a7 #898989 #757575 #5e5e5e #1a1a1a #000000 |
| Chromatic | one: NVIDIA Green #76b900, pressed #5a8d00 |
| Link | #0046a4 on the white canvas, the brand green on the black one |
| Error | #e52020 |
| Type | 48 · 36 · 24 · 20 · 17 · 16 · 15 · 14 · 12 · 11 px |
| Weights | 400 and 700, and hierarchy is built from those two plus size |
| Radius | 2px, everywhere. "Never go to 0, never go past 4" |
| Spacing | 8px base, 64px section rhythm |
| Hold | ~1280px |
| Depth | no shadow on any card or content surface |
| Modes | a black canvas and a white one, alternating down one page |
| Motion | not published; hover states are a stated known gap |
Three findings, each asserted by --check rather than described.
- The reference already solved the link problem, and differently in each mode. An inline link is
#0046a4on the white canvas and the brand green on the black one, and its Don't list then forbids that blue anywhere else: "not a button color, not a chrome color", and forbids substituting any other green for the brand green in a CTA. Solvingaccent-textwould have invented a dark green for the light theme, which is exactly the substitution the brand rules out by name. Both values are pinned per theme and both were measured against the worst ground in their own.
- The brand red cannot quite carry a sentence on the recessed band.
#e52020reads 4.61:1 on white and 4.30:1 on the soft surface, so the page emitserror-textsolved at the same hue,#e1181bin light,#f5352fin dark. The fill keeps the published red; only the text value moves. This is the general case the generator'ssemanticsmechanism was added for.
- Black text on the brand green reads 8.71:1; white on it reads 2.4:1. The reference's own
on-primaryis#000000, and the solver independently picks it. A specimen that put a white label on that green would be unreadable and would look, at a glance, like every other product.
The direction
- Archetype. Console.
- Data stance. Table-first.
- Density. Compact: 15px rows, 8px vertical padding.
- The one bold decision. The job that can never run is separated from the jobs that are merely waiting. Every scheduler reports both as pending, which is true and useless: one of them starts tonight and the other will still be there on Friday. Job 4471 asks for sixteen accelerators on a single node and the largest node has eight, so no arrangement of the queue places it. The screen says that, says what would make it run, and puts the control to change it beside the sentence.
The second state
The unplaceable job. It is the state a table-first console makes hardest, because a table's whole promise is that the rows are comparable, and this row is not comparable to the eight above it, it is in a different category wearing the same word.
This specimen is allowed red and the Tesla one is not, which is the most useful thing the two of them say side by side. Same family, same problem, opposite palettes: this reference publishes an error colour and a semantic layer, so the band uses it, and still does not rest on it, because the heading names the cause and the row carries the words "Cannot be placed".
Deviations from CONTRACT.md
- §5, the two modes read as a preference. The black and the white are two chapters of one page in the reference, alternating down it, not a user setting. Binding them to
prefers-color-schemeis this page's decision, taken because a queue is watched for a whole shift by somebody who gets to choose. The values are the reference's; the binding is not, and the design-notes panel says so on the page. - §5, motion. The reference publishes no duration and no easing, and its own Known Gaps section says hover states are undocumented "by system policy". The two rungs in
tokens.cssare the portfolio's, chosen to be unremarkable and to keep every exit inside the 250ms cap. The emitted file states this on its face rather than in a note somebody has to find. - §2, the typeface. NVIDIA-EMEA is proprietary. Inter is substituted, and this is the one substitution here the reference itself recommends, it names Inter as the closest open-source pairing, "x-height and stroke contrast match within ~2% at body sizes", and Arial as the documented fallback.
Findings kept, with reasons
slop_scan·default-ai-palette, near-black plus one bright accent. The rule's own text allows it where the brief asked for it, and here the brief is the brand: this is the reference's declared palette, not a default reached for because the axis was free. Changing it would make the specimen a specimen of nothing.audit_page·apple-touch-icon-missing,manifest-missing: portfolio-wide, as on every other page here.
What the gates changed
Three findings were fixed rather than kept, and two of them improved the page:
side-tab×3. The band's 4px coloured edge and the held row's inset bar both said in colour what the heading and the state word already say in words, so both were removed. The nav's active underline was drawn as an inset box-shadow and is now a real bottom border, same pixels, and now the linter and the reader see the same thing.- A
.cardrule that nothing used. The band now rides on it, which is how the corner square, the system's identity tag, reached the page at all.
Uplift, 2026-10-29
The archetype and the bold decision stand; the way the page proves the decision changed.
- The floor map. The hero is now the cluster itself: eight node rows of eight squares, with the selected job's request drawn above them on the same sixteen-column grid. A request of sixteen on one node is a row of squares wider than any node, its last eight dashed red over a hatched region that stands for slots no node has. Occupied squares are white or grey (alternating by job on a node), idle ones are hollow, the selected job is the brand green, and a projected job is dashed. Green is spent on selection only, as the reference's "active state".
- One model.
index.htmlcarries a small conservative-backfill scheduler (arrival order, each job at the earliest moment it fits for its whole time limit, never moving a job already placed). Every figure, table row, bar and square reads from it. The node assignments were rebuilt so that no node holds more than eight and the running jobs sum to sixty-one; the earlier table put twelve GPUs on one node. Each queued row now says why it waits and when it starts, and no queued job fits at time zero. - The interaction. The verdict card edits job 4471: a GPU slider, a one-node switch, two presets. Dragging to eight, or dropping the constraint, moves the look-ahead slider to the start time and the job lands in green on its node, with the cost to other jobs stated (six start later, the most by 1 h 04 m). Resubmit commits it to the queue; Restore puts the original back. The look-ahead slider scrubs six hours of projected floor.
- Queue. A "Next 6 h" bar per row (solid while running, outlined when projected, dashed red when nothing is projected), working State and Group filters with a real empty state ("Nothing is stuck"), and on a phone a list in place of the table, with the request on two rows of eight so the floor stays at thirty-five-pixel squares.
- Notes. Hairline cards each carrying the corner square, with the two measured pairs (black on green 8.71:1, red on black 4.56:1) as swatches.
- Added: a light/dark control for the lower page, a skip link.
- Kept finding:
launch_checkpolicy-missingreads the edit form as a data-collecting form; nothing is collected.
Round four
- On a phone the order is now headline, the four figures, the verdict card, then the floor map and its key, so the stuck job is on the first two screens rather than below sixty-four cells. From 60rem up the two columns are unchanged.
- The look-ahead slider's ticks (Now, 2 h, 4 h, 6 h) sit on the span the thumb's centre actually travels, with a small tick mark each, and the readout moved up to the label's row so it is no longer a second "Now" floating past the end of the scale.
- The white queue chapter speaks the dark chapter's language: a hairline cell grid, the green rule under the column heads, the green corner square on the table (and on each phone card), a green mark for a running job and a green edge on the selected row.
Round eight
- The 64 squares were all white and identical beyond the job id (real). Cause:
paintFloorgave every running square the same fill (alternating white and grey by position on a node), and the job showed only as an id on its first square. Now a job is a slab and its workload is a texture, in the page's own black and white, with no new colour: a job's squares after the first reach back across the gap (.t.jn::before) and keep a hairline of the ground between them, so where one job ends is a shape. Vision is the plain fill, Fluids is ruled, Genomics is dotted (g-v,g-f,g-g, taken from the job'sgroupthrough oneGROUPmap), each cut out in--nv-chapter-bg. The id sits on a chip of the square's own fill so the texture never runs under a digit. The key names all three. Selected (green) and projected (dashed) squares keep their own treatment. - Declined: no per-job tint and no second accent. NVIDIA's discipline is black, white and green, and green already means the selected job.
Round nine
Scores on the ten heuristics, before and after (0 to 4): status 3 to 4, real world 3 to 3, control 3 to 3, consistency 3 to 3, error prevention 3 to 3, recognition 2 to 3, efficiency 2 to 3, minimalism 3 to 3, recovery 3 to 3, help 2 to 3. Total 28 to 32. Below 3 after: real world (the cluster, its group names and the A100 line are invented, so a real operator's own vocabulary is not tested), control (a pick has no undo beyond picking again), consistency (the pick bar is the one surface that is neither chapter nor lower page), error prevention (the slider and switch cannot be undone one step at a time, only restored whole).
Fixed, most serious first:
- A pick in the queue changed something a screen away (real, severity 2). Cause: the floor map is the answer to a pick, and from the queue it is out of sight, so a pressed row showed nothing but its own highlight. Now a black answer bar rides the bottom of the queue after the first pick ("4496 starts in 3 h 00 m on fen-07"), with one button that goes up to the floor. The rows are rebuilt on every pick, which dropped keyboard focus to the page; focus now returns to the same button and is kept clear of the bar.
- The figures were a screen away from the control that moves them (severity 2). From 60rem up, on a screen at least 50rem tall, the headline column is sticky, so the four figures stay beside the verdict card while the sliders are worked.
- Touch targets (severity 2). On a coarse pointer the queue chips, the mode switch, the header and footer links and the two presets are 40px, the sliders 44px with a 20px thumb (the look-ahead ticks follow the thumb width). The presets became two lines of "Or" and a button.
- The phone queue card broke a name at its hyphen and wrapped its times (severity 2). Under 26rem the name takes its own row under the chip and the state, and the four figures are four content-width columns that do not wrap.
- The red caption sat under the red squares on a phone (severity 1). It is now above them, as the node caption is above its squares.
- "sixty-four" broke across lines in the headline (severity 1). Both compounds are held together.
- Row baselines in the table disagreed by two to five pixels (severity 1). Cells now share a baseline; the bar cell hangs from the top.
- The hour gridlines vanished on the dark page (severity 1). The bar sits over a short ruler of hour ticks in the stroke tone, so no tick crosses an outlined bar.
- The key was missing the green outline of the request and named the grey squares "next job" (severity 1). It now says "Running, 2nd on node" and "What it asks for".
- "As edited" read as if it were done (severity 1). It is now "If resubmitted".
- A filter changed the table silently (severity 1). A live count says "Showing 4 of 20 jobs".
- The table scrolled sideways at 1024px (severity 1). Its natural width is now 974px.
- "GPU-HOURS" had a stretched hyphen (severity 0). Header cells use normal figures.
- Design notes were seven boxes of text. Seven small drawings in the page's own black, white and one green sit over the lines: the black chapter over the white one, sixty-one squares and three hollow, the 64px gap, the request wider than a node, the two modes, the two weights and the 160 ms curve. The two contrast pairs now line up on one baseline.
- The page ended on a grey line. It now closes in black with the wordmark, as the reference's own pages do. "Nodes" in the header goes to the floor.
- Small links (severity 0). The header and footer links, the presets and the skip link are 24px or more in height under any pointer. The node labels have room for "fen-08" at 375px (the label column is 3.25rem), the look-ahead block no longer clips its own caption, and the sticky column names its top inset.
Checked and found right: 61 of 64 (eight nodes of eight, fen-08 holds five), eleven running and nine queued, 19 m to the next whole node (fen-02, job 4408), 3 h 12 m waited (192 minutes), the 4471 landing on fen-02 at +19 m when halved. The page carries no dates or weekdays.
Round ten
A second, colder read. Scores on the ten heuristics, before and after (0 to 4): status 4 to 4, real world 3 to 3, control 2 to 3, consistency 3 to 3, error prevention 3 to 3, recognition 3 to 3, efficiency 3 to 3, minimalism 3 to 3, recovery 2 to 3, help 2 to 3. Total 28 to 31 (round nine's 32 did not survive a read that tried to undo a commit and to click what looked clickable).
Fixed, most serious first:
- Resubmit was a one-press commit with no way back and no receipt (severity 3). The queue it changes is a screen away and the buttons only flipped enabled and disabled. The card now says what happened under the buttons ("Resubmitted: 8 GPUs, one node. Job 4471 starts in 19 m on fen-02. 6 queued jobs start later, the most by 1 h 04 m. Cannot be placed is now 0."), with Undo and "See it in the queue", focus on Undo, and the line brought into view. Undo goes one step back (the request before the last resubmit); "Restore the original" still goes all the way. Pressing Restore or Undo no longer leaves focus on a button that has just disabled itself.
- The job that needs a person was row twelve of twenty (severity 2). A line above the list says "1 job cannot be placed. 4471 ... asks for 16 on one node and the largest has 8" with a Go-to link that scrolls to the row and puts the keyboard on it (and clears a filter that hides it).
- Two nav labels and the key's neighbours looked clickable and were not, and the reverse (severity 2). "Images" and "Accounting" were links that only scrolled to the queue; they are now plain dim text. The whole queue row (and phone card) picks the job, not just its id chip; the id chips on the floor squares pick the job too (mouse only; the queue is the keyboard path).
- Figures with nothing beside them to say what they are made of (severity 2). Each of the four now has one line: "3 idle, all on fen-08", "Longest wait 3 h 12 m, job 4471", "Job 4471", "fen-02, when 4408 ends". All four are computed from the model.
- The look-ahead showed a floor the figures did not describe (severity 2). Away from Now, a line under the slider says how many of 64 are in use at that moment ("56 of 64 in use at that moment, as scheduled"), and the slider's spoken value carries it.
- The remedy was the last row of the verdict (severity 1). "Would run if" is now the first row, in bold.
Inventions: "Undo" as a one-step history of the one editable job (nothing in the brief names it); the figure sub-lines' wording; the receipt sentence. No new numbers: all come from the scheduler.
Left: the cluster's vocabulary is still invented (severity 1); the floor squares are clickable by mouse only (severity 1, the queue is the equivalent).
Round eleven (2026-11-10)
- The "Dark page / Light page" button changed nothing on the screen it was pressed on (severity 2). It sat in the black top bar, and the black chapter is pinned in both schemes (round one deviation: the preference moves only the page under it), so pressing it at 1440 or 390 repainted only its own label. Reproduced: chapter
rgb(0, 0, 0)before and after, body white to black. The page does have both renderings, so the control stays; it moved into the Queue header, under the Queue paragraph, where the repaint is the screen in view. The label now names what it repaints, "Light queue and notes" or "Dark queue and notes", and the aria-label that disagreed with the visible text is gone. It is restyled from the body tokens (--nv-ink,--nv-ink-3border,--hover), not the chapter ones. The scheme is still set before paint by the head script fromlocalStoragethenprefers-color-scheme; the 1ms reduced-motion rule still covers its transition. - "Images" and "Accounting" read as nav and did nothing (severity 2). Round four had turned them into dim text; dim text in the same row and type as Queue and Nodes still reads as a nav item, and nothing announced it as unavailable. The brief names no such sections and the page opens neither, so they are removed from the top bar along with the
.soonrule. The nav is Queue and Nodes, both of which scroll to a section on this page.
Inventions: the two label strings, and the control's place in the Queue header. No product fact, number or page added. Left: the two severity 1 notes (the phone sticky bar stays on the previous job after "Go to 4471"; no jump link from the lede to the verdict on a phone).
Round thirteen (2026-11-10)
- The three running fills could not be read without the legend (reviewer: the one thing still pushed). Reproduced at 1440 and 360: Vision (plain), Fluids (ruled) and Genomics (dotted) were named only in the key, which sits in the headline column at 1440 and a full screen below the map on a phone, past the look-ahead slider. A cell said nothing about whose it was.
- Fix, two parts. (1) At 30rem and up the id chip on the first square of a run of three or more squares now carries the group in regular weight after the id ("4415 Genomics"); the cell's own fill and its own name agree at the cell, with no tooltip and nothing to hover, so it reads the same by touch. The chip rides above the next squares (a z-index on that cell, a max-content chip width) and a probe at 960, 1024, 600, 480, 1440 and 1920, now and two hours ahead, found no chip past the end of its run. (2) A one-line key, "Running / Vision / Fluids / Genomics", sits directly under the node rows at every width, with swatches 20 by 16 drawn from the same fills as the squares; this is the path under 30rem, where the chips stay hidden because a square is 30px, and for the runs of one or two squares. The same three entries left the headline-column key so no fill is explained twice. At 360 the line holds on one row (swatch gap and column gap tightened).
- Inventions. None of product fact. Design choices: the threshold of three squares for writing the group on a chip, and the word "Running" as the key's lead-in. Left: the grey second-on-node runs of one or two squares (4455, 4463) carry no group word at the cell; the key line names the three fills, and the grey tone is still explained in the headline-column key.