8424

marks · 3DMark Steel Nomad DX12

RTX 4080 Overclock: Kenneth Kohlberg's 8424 Steel Nomad Record

Tarija, Bolivia

3DMark Steel Nomad DX12 — 8424 marks — RTX 4080 Founders Edition, air cooled, stock vBIOS.

First out of 136,085 results, and ahead of every RTX 4080 SUPER result.

Validation HWBOT submission

The numbers

8424 Highest RTX 4080 score in UL's database. AMD Ryzen 5 7600X3D, GPU 3030 MHz core / 1626 MHz memory, 13 July 2026.
136,085 RTX 4080 results in the database
6576 Average RTX 4080 score
6 RTX 4080 results at or above 8000 marks
98 Marks of margin over the RTX 4080 SUPER record — 1.18%
8326 Best RTX 4080 SUPER result ever: jamez74, AMD Ryzen 9 5950X, 2850/1640, 24 December 2024
348,471 Combined AD103 results, both cards
17 AD103 results above 8000 marks. Two of them are this user's: 8424 and 8221.

Second place on the RTX 4080 is tymcguire22, 8273 marks at 2940/1641. The RTX 4080 SUPER has 212,451 results and an average of 6608, on 10240 CUDA cores against the 4080's 9728 and memory binned at 23 Gbps against 22.4.

The result

8424 marks is the fastest RTX 4080 result in UL's 3DMark database, out of 136,085 RTX 4080 results. The average for the card is 6576, placing this run 28% above the mean.

Only six RTX 4080 results in the entire database have passed 8000 marks. Two of them belong to this card (8424 and 8221).

Desktop screenshot: the 3DMark Steel Nomad result page marked VALID RESULT with a score of 8 424 on an NVIDIA GeForce RTX 4080 and an AMD Ryzen 5 7600X3D, next to GPU-Z, CPU-Z and EVGA Precision X1.
The validated run, with GPU-Z, CPU-Z and Precision X1 open alongside it. GPU-Z reports the memory at 1626 MHz; both fans are at 100%. Open full resolution →

It also exceeds the SUPER

8424 marks is higher than the best RTX 4080 SUPER result ever recorded in the same database: 8326 marks, set by jamez74 in December 2024 and not beaten since — eighteen months, across 212,451 SUPER results.

That is the harder comparison, because the SUPER is the better card. It carries 10,240 CUDA cores against the 4080's 9,728, a 5.3% advantage, and its memory is binned at 23 Gbps against 22.4. It is also the variant enthusiasts bought for this purpose: 212,451 results against 136,085, some 76,000 more attempts at the same chip.

RTX 4080RTX 4080 SUPER
CUDA cores9,72810,240
Memory bin22.4 Gbps23 Gbps
Results in database136,085212,451
Average score65766608
Highest recorded score8424 (Kohlberg)8326 (jamez74)
Date of that result13 July 202624 December 2024

Core clock is not the lever

The card reaches 3100 MHz on the core. The record was set at 3030.

A Founders Edition has a power budget that cannot be raised, so core clock and memory clock compete for the same watts. Every MHz of core taken is a MHz of memory given up.

The RTX 4080 results in the database show no relationship between peak core clock and score. Runs at 3255, 3150, 3135 and 3120 MHz scored 7549, 7521, 7577 and 7534. The 3030 MHz run scored 8424. The figure 3DMark reports is the peak clock the driver observed, not the clock held for the length of the run.

2900 3000 3100 3200 3300 7400 7600 7800 8000 8200 8400 2940 MHz · 8273 marks 3015 MHz · 8139 marks 3045 MHz · 8131 marks 2940 MHz · 8127 marks 3045 MHz · 7953 marks 3030 MHz · 7890 marks 3105 MHz · 7887 marks 3000 MHz · 7876 marks 2970 MHz · 7819 marks 3030 MHz · 7804 marks 3045 MHz · 7798 marks 3015 MHz · 7758 marks 2895 MHz · 7712 marks 3045 MHz · 7700 marks 3015 MHz · 7689 marks 2940 MHz · 7682 marks 3045 MHz · 7677 marks 3075 MHz · 7646 marks 3030 MHz · 7624 marks 3060 MHz · 7615 marks 3030 MHz · 7608 marks 3045 MHz · 7605 marks 3075 MHz · 7601 marks 2940 MHz · 7598 marks 2970 MHz · 7598 marks 2985 MHz · 7593 marks 3030 MHz · 7592 marks 3000 MHz · 7589 marks 3120 MHz · 7587 marks 3075 MHz · 7585 marks 3135 MHz · 7577 marks 2880 MHz · 7577 marks 3090 MHz · 7572 marks 3045 MHz · 7571 marks 3015 MHz · 7568 marks 3105 MHz · 7566 marks 3075 MHz · 7566 marks 3090 MHz · 7564 marks 3105 MHz · 7558 marks 3255 MHz · 7549 marks 3030 MHz · 7545 marks 2955 MHz · 7544 marks 3030 MHz · 7543 marks 3120 MHz · 7534 marks 2940 MHz · 7533 marks 3045 MHz · 7532 marks 3060 MHz · 7528 marks 3030 MHz · 7526 marks 3045 MHz · 7525 marks 2985 MHz · 7525 marks 3030 MHz · 7522 marks 3150 MHz · 7521 marks 3045 MHz · 7521 marks 3045 MHz · 7519 marks 3015 MHz · 7519 marks 3045 MHz · 7518 marks 3075 MHz · 7515 marks 3075 MHz · 7513 marks 2970 MHz · 7513 marks 2970 MHz · 7511 marks 3030 MHz · 7508 marks 3105 MHz · 7507 marks 3045 MHz · 7498 marks 3030 MHz · 7497 marks 3105 MHz · 7496 marks 3120 MHz · 7495 marks 3090 MHz · 7495 marks 3030 MHz · 7493 marks 3045 MHz · 7491 marks 2985 MHz · 7491 marks 3030 MHz · 7483 marks 3030 MHz · 7483 marks 3030 MHz · 7482 marks 3030 MHz · 7480 marks 3045 MHz · 7480 marks 3060 MHz · 7479 marks 2985 MHz · 7479 marks 3105 MHz · 7479 marks 3015 MHz · 7477 marks 3105 MHz · 7477 marks 3030 MHz · 7477 marks 3045 MHz · 7474 marks 2985 MHz · 7474 marks 3105 MHz · 7474 marks 3015 MHz · 7474 marks 3090 MHz · 7474 marks 3015 MHz · 7473 marks 3060 MHz · 7473 marks 3000 MHz · 7471 marks 3045 MHz · 7470 marks 3105 MHz · 7470 marks 3030 MHz · 8424 marks 3060 MHz · 8221 marks 3060 MHz · 7971 marks 3030 MHz · 7860 marks 3045 MHz · 7693 marks 3045 MHz · 7680 marks 3075 MHz · 7538 marks 3060 MHz · 7484 marks 3060 MHz · 7478 marks GPU core clock (MHz) Score (marks)
  • Kenneth Kohlberg
  • Everyone else
Peak GPU core clock against score, for the 100 highest RTX 4080 results in the database. Peak clock does not predict score: the fastest run in the set, at 3255 MHz, scored 7549, while the record sits at 3030.

Finding the memory ceiling

GDDR6X uses PAM4 signalling, which carries two bits per symbol. That doubles transfer efficiency but leaves a weak signal-to-noise margin and makes the bus sensitive to PCB quality and temperature. NVIDIA compensates with Error Detection and Replay (EDR): a CRC on the memory bus that catches transmission errors and retries each transfer until it succeeds.

EDR inverts the usual method for finding a memory ceiling. Because errors are corrected rather than displayed, artifacts and crashes stop being useful indicators. The score itself becomes the instrument: it climbs, flattens as replays begin consuming bandwidth, then declines — with the card appearing perfectly stable throughout. The usable ceiling is the top of that curve, not the point of failure.

On this card the curve turns at 1626 MHz. Past it, performance regressed instead of improving, so the clock was brought back down. 1626 is not where the card stops working; it is where it stops getting faster.

Beyond the curve there is no gradual decline. One further step produces a black screen and a full system shutdown requiring the PSU to be cycled. A second, thermal failure mode terminates the run and returns to the benchmark menu — a driver timeout, not corruption; the record run was made in winter because lower ambient widens the stable window on a PAM4 bus. Both failure modes are loud. Neither produces the pattern that should make a result suspect: a completed run, a valid score, and visible artifacts.

Method

Three stages, in order. Tuning both clocks simultaneously reveals nothing about which one moved the score.

  1. Core alone, memory at stock, stepped to failure. Ceiling: 3100 MHz.
  2. Memory alone, core at stock, stepped while watching the score curve rather than the screen. Ceiling determined by performance, not by crash.
  3. Balance point. With both ceilings known, find the split of the power budget that maximises the score rather than either clock. Result: 3030 / 1626. Seventy MHz of core traded for memory clock, and the score rose.

Thermal setup: the fans run at 100% for the length of every 3DMark run. The phase-change pad on the die was cured before use: twenty-five cycles, each about two hours hot, then left to cool to 35 °C before the next one begins.

Configuration

GPURTX 4080 Founders Edition, stock vBIOS, 3030 / 1626 MHz
GPU coolingStock FE cooler, phase-change pad on the die
CPURyzen 5 7600X3D @ 4750 MHz, Curve Optimizer -55 all-core via PBO2 Tuner
CPU coolingDeepcool AK620, Thermal Grizzly Kryonaut
MemoryG.Skill Flare X5 DDR5 32 GB, rated 6000 CL36, running 6400 CL32-36-36-36
MotherboardASUS ROG Strix X670E-E Gaming WiFi
PSUASUS ROG Strix 1200 W
Open PC case: the RTX 4080 Founders Edition mounted horizontally above the power supply shroud, a twin-tower air cooler above it, and case fans front and rear.
The card as it ran: Founders Edition, stock cooler, air only.

In Steel Nomad the overall score equals the graphics score, so the CPU contributes nothing to this number. Worth stating, because results of this kind are assumed to require a halo CPU. A six-core produced this one.

About

Kenneth Kohlberg is an overclocker based in Tarija, Bolivia, competing on HWBOT as KennethK. All results are on retail hardware, air cooled. Contact: kennethkohlberg77@gmail.com

HWBOT profile