NITROBENCH 4 · ODI 2029.2

Benchmarks

Everything below is measured, dated and reproducible, including the results that go against us. NITROBENCH 4 is our own harness — that is a conflict of interest, so we publish the harness, the profiles and the raw run files, and we print the third-party index next to ours where they disagree.

01The headline

Sustained cache throughput, NX5/X
1,180
MB/s · thirty minutes unbroken

Against 870 MB/s from a retail Mk V on the same profile on the same day. A 36% lead at a typical resident set, and a lead we lose entirely past 22 residents.

Fold ratio
1.72×
41.3 effective on 24 physical blocks.
Credits / 100 MB/s
949
Mk V measures 1,931 on the same figure.
Profiles in the set
41
Nine of them stress cases we expect to lose.
Runs per figure
12
Median reported. Spread published in the run files.

02Throughput

Sustained cache throughput — NITROBENCH 4 mixed-set profile
  • Nitrosoft, generation 5
  • Cybrscape, measured by us
  • Nitrosoft, generation 4
NX5/X1,180 MB/s
Mk V870 MB/s
NX5/T690 MB/s
NX4/T520 MB/s
Mk IV430 MB/s
NX5/S410 MB/s

Cybrscape units measured 14 June 2029: two retail Mk V, two retail Mk IV, unmodified firmware. Our measurements of retail hardware, not their published specification. If they measure differently we would like to see the profile.

Figures
Sustained cache throughput
UnitMB/s
NX5/X1,180
Mk V870
NX5/T690
NX4/T520
Mk IV430
NX5/S410
Credits per 100 MB/s sustained — lower is better
  • Nitrosoft, generation 5
  • Cybrscape, measured by us
  • Nitrosoft, generation 4
NX4/T654 cr
NX5/X949 cr
NX5/T1,072 cr
NX5/S1,195 cr
Mk V1,931 cr
Mk IV2,233 cr

The NX4/T wins this outright because the tooling is paid for and we have stopped amortising it. That is not engineering, it is accounting, and it is the honest reason a previous-generation deck tops this chart.

Figures
Credits per 100 MB/s sustained
Unitcr (lower is better)
NX4/T654
NX5/X949
NX5/T1,072
NX5/S1,195
Mk V1,931
Mk IV2,233

03The crossover

Where we stop winning

Throughput held against resident count — NX5/X against a retail Mk V
  • NX5/X
  • Mk V
03006009001,20026101418222630WAREZ RESIDENT AT ONCECROSSOVER250505MB/s

A wide fold table costs lookup time. Up to about eighteen residents we are ahead by a wide margin; at 22 the curves cross; at thirty the Mk V holds 505 MB/s against our 250. If your working set is routinely wider than twenty residents, buy theirs. We would rather lose one sale than acquire a support case that was never going to end well.

Figures
Throughput held against resident count
ResidentNX5/X (MB/s)Mk V (MB/s)
21,180870
61,176868
101,168862
141,140848
181,010812
22720744
26430640
30250505
Our advantage at 14 residents
+292MB/s
1,140 against 848. The typical mixed set sits here.
Crossover
22residents
720 against 744. Inside the run-to-run spread either way.
Their advantage at 30 residents
+255MB/s
505 against 250. Not close, and not our deck's job.

04Residency

Physical against effective blocks — all five generations
  • Physical blocks
  • Effective blocks, folded
G1 · NX11.12×
1213.4
G2 · NX21.19×
1619.0
G3 · NX31.33×
2026.6
G4 · NX41.46×
2435.0
G5 · NX51.72×
2441.3

Memory is the scarce resource in every deck ever built. Nothing in this chart changes that. What changes is the ratio between the two bars.

Figures
Physical against effective blocks
UnitPhysicalEffectiveRatio
G1 NX11213.41.12x
G2 NX21619.01.19x
G3 NX32026.61.33x
G4 NX42435.01.46x
G5 NX52441.31.72x

Why the ratio is the honest metric

A block count is a number about a board. A fold ratio is a number about your work. Two decks with the same twenty-four blocks will hold different amounts depending entirely on how repetitive your resident set turns out to be.

Our figures use a fixed mixed set — nineteen residents, three Barrier probes, two Ward wrappers, a Sentry sweeper and thirteen utilities — specified page-for-page in the harness. A set of nineteen genuinely unique residents folds to about 1.08× on our hardware, which is to say barely at all. We publish that case too; it is profile 34 and it is one of the nine we expect to lose.

The incumbent’s Mk V measures 1.08× on the same set with 28 physical blocks, giving 30.2 effective. Four more blocks of silicon than ours, eleven fewer blocks of work.

05Third party

The index we did not write

The Open Deck Index is a consortium figure. We submit hardware to it, we do not run it, and we do not get to choose the profiles. Index 100 is the median deck in their sample.

We lead on throughput and residency by a wide margin, lead narrowly on ICE traversal, and lose on trace discipline, acoustics and service turnaround. Two of those three are the same two things we lose on our own harness. The third — service turnaround at 64 against 161 — is the worst number on this site.

We are not going to contextualise a 64. It means our service department is well below the median deck supplier and a long way below the incumbent’s. What we are doing about it is on the support page, with the targets we currently miss.

Open Deck Index 2029.2 — index, 100 = median deck, higher is better
  • NX5/X
  • Mk V (measured)
Cache throughputNX ahead
168124
Resident set densityNX ahead
181112
ICE traversalNX ahead
143131
Trace disciplineMk V ahead
129152
AcousticMk V ahead
71148
Service turnaroundMk V ahead
64161

Published by the consortium. We submit hardware; we do not run the tests and we do not see the profiles in advance.

Figures
Open Deck Index 2029.2
MetricNX5/XMk V
Cache throughput168124
Resident set density181112
ICE traversal143131
Trace discipline129152
Acoustic71148
Service turnaround64161

06Head to head

NX5/X against the incumbent Mk V

Ten rows. We win five, they win five, and the five we lose stay in the table.
MeasureNX5/XMk VAhead
Physical blocksThey have more memory than we do.2428Cybrscape
Effective blocks (folded)And we hold more in it.41.330.2Nitrosoft
Sustained throughput1,180 MB/s870 MB/sNitrosoft
Credits per 100 MB/s9491,931Nitrosoft
ICE traversal, median14 ms17 msNitrosoft
Trace floor at idleTheirs is quieter on the wire. We are working on it.1.6 /min1.4 /minCybrscape
Noise under sustained loadNot close. Ours is a loud deck.41 dB(A)21 dB(A)Cybrscape
Warranty12 months36 monthsCybrscape
Median first support replyTheir service department is better than ours.14 h2 hCybrscape
Generations shipped since 202652Nitrosoft

Cybrscape is a good company that builds a good deck and has a better service department than ours. We compete with them on throughput, on residency and on credits per unit of work, and we expect to keep losing on warranty, noise and support until we fix those — which are engineering and hiring problems, not marketing ones.

07Method

How every figure was taken

Conditions

  • Ambient 24 °C ± 0.5, still air, enclosure closed and screwed shut.
  • Thirty-minute unbroken run. The reported figure is the mean of the final twenty-five minutes, so warm-up and any throttle are inside it.
  • Twelve runs per figure, median reported, full spread in the run file.
  • Stock firmware, factory settings, no fold-table overrides, no thermal override.
  • Mains supply through a regulator; draw measured at the wall.
  • Noise at one metre, on-axis, in a room with a 17 dB(A) floor.

Profiles

Forty-one profiles in NITROBENCH 4. The headline mixed set is profile 12: nineteen residents, three Barrier probes, two Ward wrappers, one Sentry sweeper, thirteen utilities, specified page-for-page. Nine profiles are deliberate stress cases we expect to lose, including profile 34 — nineteen genuinely unique residents, where our fold ratio collapses to 1.08×.

The competitor units

two retail Mk V, two retail Mk IV, unmodified firmware, bought at retail through an ordinary supply desk, measured 14 June 2029. Unmodified firmware, factory settings, same room, same day, same profile set. These are our measurements and we label them that way everywhere they appear; they are not Cybrscape’s specification and should not be quoted as such.

Conflict of interest

NITROBENCH 4 is ours. We wrote the harness, chose the profiles and ran the hardware, and a supplier’s own benchmark is worth exactly as much as its method is inspectable. So: the harness ships in the SDK under nitro bench, the profile definitions ship with it, and the signed run files for every figure on this page are in the SDK archive. Run it yourself. If you get a different number we would rather hear about it than not.

What we will not publish

Figures for hardware we have not measured ourselves. Extrapolations. Peak-burst numbers without the sustained number beside them. Anything from a run we could not reproduce at least nine times out of twelve.

08Cadence

Release cadence — sustained throughput at launch, MB/s
G1NX1MARCH 2026120 MB/s
G2NX2DECEMBER 2026215 MB/s
G3NX3SEPTEMBER 2027390 MB/s
G4NX4JULY 2028700 MB/s
G5NX5APRIL 20291,180 MB/s
Figures
Release cadence
GenerationDeckShippedBlocksFoldMB/s
G1NX1March 2026121.12x120
G2NX2December 2026161.19x215
G3NX3September 2027201.33x390
G4NX4July 2028241.46x700
G5NX5April 2029241.72x1,180

G1: Bench only. 40 units, all hand-finished, all replaced free. G2: First unit anyone paid for. Hotfloor 1.4 introduced block folding. G3: Eviction prediction lands. Support queue triples; see /support. G4: Still shipping as the NX4/T. Firmware supported to July 2033. G5: Same twenty-four blocks as G4. That is the entire point of it.