PC requirements & benchmarks

GPU, memory and CPU requirements for DKB Recomp, plus measured RTX 5060 Ti, RTX 3070 Ti Laptop and RX 9070 XT benchmarks.

  • GPU: NVIDIA RTX rec.
  • RAM: 16 GB min / 32 GB rec.
  • Lazy heap: ~9.3 GiB commit
  • DLSS & HDR10: Supported
Macro view of a handheld mainboard PCB and multi-core GPU silicon package
GPU silicon pipeline (Ampere to Maxwell to SPIR-V)
Advertisement

Every hardware requirement below comes directly from the Lucas-blip-png/dkb-pc-recomp README and release notes for v0.3.0:

Operating system
Windows 10 / 11 x64 (supported); Linux x86-64 with kernel 6.7+ (preview); macOS 14.4+ on Apple Silicon (preview).
Graphics card (GPU)
NVIDIA GPU with a recent driver strongly recommended (tested on GeForce RTX 5060 Ti and RTX 3070 Ti Laptop). AMD and Intel GPUs are experimental due to device-lost faults (#3, #84, #99, #122, #124, #142, #193, #204).
System memory (RAM)
16 GB RAM plus a large page file minimum (commit limit ≥ 24 GiB); 32 GB RAM is comfortable. Default first-level commit is ~13.5–14.5 GiB, or ~9.3 GiB with TOTK_HEAP_LAZY=1 on Windows.
Processor (CPU)
Modern x86-64 or Apple Silicon CPU. In tools/lowend_bench.py, the first level held 60 fps on 16 and 8 logical processors, 58.9 fps on 4, and 55.6 fps on 2 logical processors.
Controller
SDL3-compatible gamepad recommended (Xbox, DualSense, Switch Pro, 8BitDo, GameSir), or keyboard and mouse with direct mouse camera enabled.
Game dump
Your own decrypted copy of Donkey Kong Bananza base 1.0.0 or update 3.2.0 (v458752) as a .nup file or ExeFS + RomFS folders.

Measured GPU & scene benchmarks

Upstream release notes and contributor profiling PRs (#161, #195, #207, #211) report the following real-world frame rates:

Advertisement
GPU & SystemGame DumpResolution / SceneMeasured Frame Rate
GeForce RTX 5060 Ti (Desktop)1.0.0 & 3.2.01080p – 3840x2160 (4K)60 fps held for 150 s at start & 90 s at 4K; heavy scene 51.8 fps avg (20.6 fps 1% low)
GeForce RTX 3070 Ti Laptop1.0.01080p31–35 fps in open areas (CPU/Vulkan replay bound)
Radeon RX 9070 XT (16 logical CPUs, Windows)3.2.0Hills region (28k draws)40.9 fps (8 threads) → 44.9 fps (6 threads) → 46.0 fps with TOTK_DYN_UBO=1 (#211)
Radeon RX 6700 10 GB + i5-11400F (16 GB RAM)1.0.0 / 3.2.0Default (30 fps cap)30–60 fps cleanly through sublayer 200; drops to 16–17 fps at sublayer 200 (#207)

CPU core scaling (tools/lowend_bench.py)

In v0.3.0, upstream added tools/lowend_bench.py (documented in docs/dkb-lowend.md) to pin the game to the first N logical processors on the opening level:

Logical ProcessorsFirst-Level FPSReplay Threading Behavior
16 logical processors60.0 fps3 helper threads started automatically (#185)
8 logical processors60.0 fps3 helper threads started automatically (#185)
4 logical processors58.9 fps2 helper threads started automatically (#185)
2 logical processors55.6 fpsSingle-thread / baseline replay path

Memory commitment & TOTK_HEAP_LAZY=1

Without lazy commitment, DKB Recomp commits roughly 13.5 GiB at startup and 13.75 GiB after 40 minutes (growing ~6 MB/min), peaking near 14.5 GiB on the first level. On 16 GB RAM systems, Windows must have a page file large enough to keep total commit limit above 24 GiB. In v0.3.0, setting the environment variable TOTK_HEAP_LAZY=1 on Windows commits guest heap pages on first touch (backed by spin-lock patches in 1.0.0 and 3.2.0, #126), dropping first-level commit from 14.5 GiB to ~9.3 GiB. You can inspect memory allocation with TOTK_MEM_REPORT=N.

DLSS, HDR10, 4K & Frame Generation in v0.3.0

  • Temporal DLSS (experimental, NVIDIA RTX only): Feeds sub-pixel jitter, depth and camera-only motion vectors to NVIDIA NGX. DLSS output matches the game’s 3D resolution directly, and the HUD displays the upscaler’s true input/output dimensions. (Note: if combining DLSS with multi-threaded pass slicing, track PR #209.)
  • Native HDR10 (experimental, off by default): Configured via hdr, hdr_peak_nits and hdr_game_nits. On update 3.2.0 with hdr = 1, setting DKB_HDR_NATIVE=1 lets Bananza render its own HDR10 pipeline and scales highlights past the stock 1.36x paper-white tonemap up to your display’s peak nits (#150).
  • 4K & Ultrawide: 3840x2160 and 5120x1440 dynamically grow the render-target pool (#152), joined by 3840x1600 and 5120x2160 presets (#168).
  • Frame generation (FRAME_GEN=2x): Displays an interpolated frame half an interval before each real frame in mailbox or immediate present mode, automatically skipping generated frames on hitches or crop transitions (docs/dkb-framegen.md).

DKB Recomp recommends a gamepad on Windows, Linux and macOS via its SDL3 Horizon host, with dedicated Controller and Mouse & Keyboard settings tabs plus Bananza action descriptions. These controllers pair cleanly on PC:

ControllerBest For in DKB RecompHardware AdvantageLink
GameSir Cyclone 2Tunnel punching & 3D camera controlTMR sticks, Hall linear/microswitch triggers, 1000 Hz wireless dongleView on Amazon
8BitDo Ultimate 2 WirelessDesktop Windows & Linux setupZero-drift TMR sticks, charging dock, Switch/XInput mode switchView on Amazon
Sony DualSenseLow-latency USB & SDL3 promptsResponsive face buttons, crisp D-pad, native SDL3 HID supportView on Amazon
Xbox Wireless ControllerPlug-and-play Windows 10/11Native Windows XInput & Linux xpadneo pairing with zero setupView on Amazon

As an Amazon Associate, this site earns from qualifying purchases.

Advertisement
Stepped limestone tiers of the Temple of the Inscriptions surrounded by dense jungle canopy at Palenque
Stepped stone tiers and jungle canopy at Palenque
Desktop gaming PC workstation with dual monitors and tempered glass chassis
Windows x64 desktop workstation for Vulkan 1.3 and DLSS
Open Apple laptop in three-quarter side profile on a studio background
macOS ARM64 preview target (MoltenVK and GfxBackend)

Common questions

Why are AMD and Intel GPUs listed as experimental?

All initial development and shader validation took place on an NVIDIA RTX 5060 Ti and RTX 3070 Ti Laptop. Several AMD RDNA2/RDNA3/RDNA4 cards report VK_ERROR_DEVICE_LOST (-4) or ring gfx timeouts (#3, #84, #99, #122, #198, #204), though v0.2.0 and v0.3.0 fixed the AMD map cyan wedges (#125), bloom NaN bands (#132), and added TOTK_FIX_VS_FAULT=all and TOTK_CRUMB_BIND=1 diagnostics.

How many worker threads should I set for DKB_REPLAY_THREADS?

In contributor testing on an 8-core / 16-thread CPU (#211), 6 threads outperformed 8 threads (44.9 fps vs 41.9 fps, and 46.0 fps with TOTK_DYN_UBO=1) because leaving two cores free reduces thread contention during driver command recording.

Does running above 60 fps affect game physics?

In v0.3.0, TOTK_FPS_LIMIT=unlimited was added (#113) and the smash radial blur that stayed stuck above 60 fps on 3.2.0 was fixed (#158, #159). However, Issue #201 notes that climbing walls or grates can fail when the framerate actually exceeds 60 fps—cap at 60 fps when climbing sections require it.