Scenes captured with handheld and wearable cameras are still shipped to a CUDA workstation for 3D Gaussian Splatting (3DGS) training. Training on the device that captured them, or on its companion phone or laptop, requires more than translating kernels: GPUs and drivers differ in parallel features, memory limits, and command submission, and these differences decide whether a training schedule executes and returns the correct update. We present Cross3DGS, a cross-vendor, cross-platform Vulkan trainer that keeps one optimizer and chooses its execution schedule per device in two steps: rules on device limits, workload size, and observed driver failures exclude the schedules known not to run correctly, and a fixed preference order selects one of the remaining feasible schedules. The same core completes standard-quality training (30k updates, one million Gaussians) on every configuration we measured, spanning NVIDIA, Intel, Arm, Qualcomm, AMD, and Apple GPUs and including Android devices and an iPhone, whereas the existing portable trainers VkSplat and Brush do not run on some of the mobile GPUs. Averaged over 11 scenes on a workstation, Cross3DGS has higher quality than VkSplat with no significant difference in training time, and Brush reports about three times the training time at comparable quality. The final quality of Cross3DGS is nearly identical across devices. A mobile application turns a handheld capture into a viewable model entirely on the device.
Figures & tables
Figure 2: Existing 3DGS trainers and Cross3DGS as software stacks. Existing trainers fix their kernel sets per backend or vendor, or pick among them by timing; Cross3DGS selects a schedule for each device in two steps and is the only one of the three trainers we ran to complete standard-quality training on all eight devices (Figure 4 ).
resident (all in GPU memory), streamed (one per update)
–
images exceed one storage buffer
device fault
radix sort
2
subgroup, portable
–
–
corrupted sort keys
projection and Adam
2
fused, split
buffer bindings
SH coefficients exceed one storage buffer
–
Table 1: The six options of a schedule, the number of values of each, and the values. The catalog contains every combination. Within each option, values are listed in the order the common preference order ranks them. The last three columns name what moves a device to a later value; the driver failures were observed on Qualcomm Adreno740 .
Figure 3: Choosing a schedule for a device. Step 1 drops every schedule in the catalog that fails a feasibility rule; the rules read the device limits Cd , the workload W , and the driver rules E . Step 2 picks one of the remaining feasible schedules with a fixed preference order. The gradient methods in the catalog are checked offline against the reference update. Table 2 shows both steps on the eight measured devices.
vendor
NVIDIA
Intel
Intel
Arm
Qualcomm
AMD
Apple †
Apple †
product
PRO6000 [ 49 ]
Iris Xe [ 30 ]
UHD730 [ 31 ]
Mali G1-Ultra [ 6 ]
Adreno740 [ 51 ]
RX9070XT [ 2 ]
M5Pro [ 4 ]
A14 [ 3 ]
launch year
2025
2023
2024
2025
2022
2025
2026
2020
host
Windows workstation
Windows laptop iGPU
Ubuntu desktop iGPU
Android phone
Android tablet
Ubuntu desktop
macOS laptop
iOS phone
memory (GB)
96
16
16
12
12
16
48
6
Device limits Cd
subgroup width
32
32
32
16
64
64
32
32
max shared memory (KiB)
48
32
64
32
32
64
32
32
Table 2: The eight measured configurations, their device limits Cd , the number of the 256 schedules still feasible after each kind of feasibility rule is applied in turn, and the selected schedule c⋆ for the standard-quality setting, one row per option of Table 1 . Adreno740 also stores its SH coefficients in two buffers. Memory is dedicated GPU memory for discrete GPUs and unified system memory otherwise. Host devices: Iris Xe is integrated in a Core i7-1370P and UHD730 in a Core i3-14100, whose launch years are given; G1-Ultra is in a Xiaomi 17T Pro, Adreno740 in a Xiaomi Pad 6S Pro 12.4, M5Pro in a MacBook Pro, and A14 in an iPhone 12 Pro. Daggers mark configurations held out from the development of the rules and the preference order.
Figure 4: Standard-quality time, quality, and execution coverage on bicycle . Right: three-run median training seconds (log scale, device- and method-local timing scopes) against mean-view PSNR; color marks the GPU, as in the text, and marker shape the method. The shaded band is the eight-device Cross3DGS PSNR span. Left: check = measured, dash = unsupported by the evaluated version (VkSplat: no Android or iOS host, ≥ 32-lane subgroups; Brush: no iOS build), cross = invalid execution (Brush exceeds the 128-MiB storage-buffer range of Adreno740 at one million Gaussians; VkSplat hangs the UHD730 GPU). Daggers mark the configurations held out from the development of the rules and the preference order.
feasible
same GPU
CPU double
configuration
methods
∣Δg∣
∣ΔAdam∣
∣Δg∣
∣ΔAdam∣
viol.
PRO6000
8
3.0e−8
6.3e−7
1.6e−5
3.0e−7
0
Iris Xe
4
9.3e−9
2.4e−7
1.6e−5
1.2e−7
0
UHD730
5
2.8e−8
6.2e−7
1.6e−5
1.2e−7
0
G1-Ultra
4
1.4e−8
1.2e−7
1.6e−5
1.2e−7
0
Adreno740
3
9.3e−9
1.2e−7
1.6e−5
1.2e−7
0
Table 3: One-step update consistency (Eq. 3 ) on the six development configurations. Same GPU : every feasible gradient method against exact pair gather, 100 random small scenes, within 10−5+10−4∣qk∣ . CPU double : the selected schedule against a double-precision implementation with finite differences and one Adam update, 20 cases, within 10−6+10−3∣gk∣ and 10−5+10−2∣Δθk∣ . Entries are the largest absolute differences; viol. counts values outside tolerance.
configuration
bicycle
garden
PRO6000
25.30/0.767/0.196
27.01/0.852/0.108
Iris Xe
25.28/0.766/0.196
26.99/0.852/0.108
UHD730
25.30/0.767/0.196
27.02/0.852/0.108
G1-Ultra
25.30/0.767/0.195
27.00/0.852/0.108
Adreno740
25.29/0.767/0.196
26.97/0.852/0.108
RX9070XT
25.30/0.767/0.196
27.03/0.852/0.107
Table 4: Cross3DGS standard-quality quality on two scenes and eight configurations: PSNR/SSIM/LPIPS medians of three runs. Daggers mark held-out configurations.
Figure 5: Standard-quality renders and error maps on bicycle for three devices whose deployed gradient accumulation differs (Table 2 ) and for the two baselines on PRO6000 ; each label reports its run’s mean-view PSNR, with the GPU’s color and the method’s marker of Figure 4 . The view is the PRO6000 per-view PSNR median, selected before inspection; one fixed crop is shared by every column, and error maps use a fixed [0,0.25] mean-absolute RGB range.
Cross3DGS
VkSplat
Brush
scene
P ↑ /S ↑ /L ↓
s
P ↑ /S ↑ /L ↓
s
P ↑ /S ↑ /L ↓
s
Mip-NeRF 360
bicycle
25.30 /.767/.196
115.1
25.11/.760/.201
114.5
25.15/ .769 / .188
425
garden
27.01/.852/.108
110.0
26.95/.848/.114
106.3
27.21 / .855 / .098
387
stump
26.84 / .786 / .163
110.0
26.63/.779/.171
113.4
26.42/.776/.166
397
bonsai
32.35/ .947 /.120
171.8
32.56/ .947 /.119
175.2
32.78 / .947 / .117
440
Table 5: Standard-quality 11-scene comparison on PRO6000 : three-run medians, P/S/L = PSNR ↑ /SSIM ↑ /LPIPS ↓ , s = training-only seconds (Brush: its own integer timing scope). Bold marks the best displayed value per row. “Macro” averages the three per-dataset scene means; its time is the median scene time.
Figure 6: Standard-quality bicycle training on the three mobile devices, screen on; 1-minute means. (a) GPU temperature, the hottest of eight sensors on Adreno740 . (b) GPU clock. (c) Step time. (d) G1-Ultra in the shaded areas of (b) and (c), enlarged and overlaid: step time in orange (left), GPU clock in black (right). (e) Battery energy since the cable came out; A14 from its level; G1-Ultra mean power before and after its clock drop. Vertical lines: G1-Ultra clock drops, A14 on the charger.
Figure 7: A plush toy reconstructed on A14 , G1-Ultra , and Adreno740 , each from its own capture, and two outdoor scenes on G1-Ultra , at the demonstration preset unless labeled standard quality: held-out target, reconstruction, and mean-absolute RGB error in [0,0.25] . Plush-toy frames are shown as recorded.
demonstration preset
standard quality
plush toy
plush toy
plush toy
shrine
statue
shrine
statue
device
A14
G1-Ultra
Adreno740
G1-Ultra
G1-Ultra
G1-Ultra
G1-Ultra
updates/cap.
3k/250k
3k/250k
3k/250k
3k/250k
3k/250k
30k/1M
30k/1M
registered
100/100
100/100
100/100
100/100
99/100
100/100
99/100
train/held-out
87/13
87/13
87/13
87/13
86/13
87/13
86/13
seed points
62,127
95,814
66,912
54,676
36,784
54,676
36,784
Table 6: Each stage of the application for the seven rows of Figure 7 (SH3, MCMC); 3DGS times include the live preview. Each indoor plush-toy session runs every stage on its device from that device’s stored capture. The three captures differ, so these columns do not compare devices. The outdoor shrine and statue sessions do the same from G1-Ultra captures; each standard-quality run re-trains from its demonstration run’s structure from motion and shares its times.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
setting
value
updates
30,000, one training image per update
view sampling
uniform with replacement, seeded
loss
0.8L1+0.2(1−SSIM) ; SSIM with an 11×11 Gaussian window ( σ=1.5 ), C1=0.012 , C2=0.032 , zero padding, mean over the full image
SH degree
3; the active degree is min(3,⌊t/1000⌋) at update t
capacity
1,000,000 Gaussians
initialization
SfM points; scale 0.1× the root-mean-square distance to the three nearest points; opacity 0.5; random unit quaternion; SH DC from the point color, higher SH zero
Appendix
Table 7: Cross3DGS standard-quality settings, identical on every configuration. Ri and si are the rotation and scales of the source Gaussian.
method
reported time
included
excluded
Cross3DGS
sum over training calls of the host wall time ( steady_clock ) from issuing execute (one update) or executeBatch (eight) to its return
all 30,000 updates; the host waits for each submission (eight updates when batched, one pass when fenced) before issuing the next
scene loading, pipeline and shader setup, validation, render and PLY export, evaluation, and viewer readback
VkSplat
sum over updates of the host wall time ( perf_counter ) around each train_step call
all 30,000 updates
the same
Brush
the training time its command-line interface reports, in whole seconds
what its training loop counts
not controllable from outside
Appendix
Table 8: Timing scopes. The warm-up affects only step-time percentiles, never the total: the first 20 updates in the command-line trainer (desktop and iOS), and the first 100 in the Android trainer and the capture-to-view applications. On Android, the application’s training time also includes pipeline upload and readback.
PSNR (dB)
training time (s)
scene
seed
Cross3DGS
VkSplat
Brush
Cross3DGS
VkSplat
Brush
bicycle
20260815
25.31
25.11
25.18
114.2
114.4
425
20260816
25.30
25.11
25.15
115.3
114.8
425
20260817
25.26
25.11
25.14
115.1
114.5
426
garden
20260815
27.09
26.98
27.21
110.3
106.3
386
20260816
26.99
26.92
27.27
109.9
105.8
388
Appendix
Table 9: Every run of the 11-scene comparison on PRO6000 (Table 5 ). Brush times are its own whole-second scope (Table 8 ).
kind
option
rule
a schedule fails when
value instead
device limit Cd
gradient method
SUBGROUP-TILED
it uses the tiled assignment and the subgroup width is below 8 or shared memory is below 49,152 bytes
a later gradient method
gradient method
ATOMIC-NATIVE
it merges with native atomics and the device lacks float32 atomic add
CAS or fixed reduction
projection and Adam
BINDING-FUSED
it fuses projection and Adam and a stage may bind fewer than 26 storage buffers
split
radix sort
RADIX-SHARED
the radix histogram, (max(2048,256⌈256/s⌉)+512)×4 bytes for subgroup width s , exceeds shared memory
none: training is refused
pair dispatch
DISPATCH-INDIRECT-FEATURE
it dispatches pairs indirectly and the runtime’s indirect-dispatch flag is off
direct
workload W
training images
DESCRIPTOR-TARGET-RANGE
it keeps images resident and width × height × views × 4 bytes exceeds the storage-buffer range
streamed
Appendix
Table 10: The 13 feasibility rules of the catalog. Kind is the input a rule reads: device limits Cd , workload W , or driver rules E ; option is the schedule option of Table 1 that it constrains. Indirect dispatch is core in Vulkan 1.0, so the flag is always on and DISPATCH-INDIRECT-FEATURE has never removed a schedule. Table 11 gives the failure behind each driver rule.
option
rule
cause: the excluded value that was run
result
gradient method
ATOMIC-SCATTER-QCOM
a gradient pass merging with CAS
scatter failure: the pass timed out and the device was lost ( VK_ERROR_DEVICE_LOST )
command submission
SUBMISSION-BATCHED-QCOM
batched submission
watchdog timeout: the run did not finish and was stopped by a 480 s watchdog
training images
TARGET-RESIDENT-QCOM
all training images resident, the current target copied with vkCmdCopyBuffer
device fault: the second update faulted, the device was lost, and the tablet restarted
radix sort
RADIX-SUBGROUP-QCOM
subgroup radix sort
corrupted sort keys: of 17,624 sorted entries, 3,727 out of order and 8,580 with wrong keys or values
Appendix
Table 11: Driver failures on the Adreno740 tablet (driver 0676.76.1). Each row ran the option excluded by its rule in Table 10 .
configuration
vendor identity
storage-buffer bindings
PRO6000
0x10DE
1,048,576
Iris Xe
0x8086
200
UHD730
0x8086
201,326,592
G1-Ultra
0x13B5
8,388,608
Adreno740
0x5143
16,777,216
RX9070XT
0x1002
8,388,606
Appendix
Table 12: Rule inputs not shown in Table 2 , as reported by each driver. Every configuration allows the 26 bindings that fused projection and Adam need at the standard-quality setting.
configuration
operating system
GPU driver
Vulkan API (device)
PRO6000
Windows 11 (10.0.26200)
NVIDIA 596.72
1.4.329
Iris Xe
Windows 11 Home (10.0.26200)
Intel 32.0.101.7085
1.4.323
UHD730
Ubuntu 24.04.5, Linux 7.0.0-38
Mesa ANV 25.2.8 (0ubuntu0.24.04.4)
1.4.318 a
G1-Ultra
Android 16, HyperOS 3.0.305 b
Arm r54p1 b
1.3.305 b
Adreno740
Android 16, HyperOS 3.0.305
Qualcomm 0676.76.1 c
1.3.128
RX9070XT
Ubuntu 24.04.5, Linux 7.0.0-38
Mesa RADV 25.2.8 (0ubuntu0.24.04.4)
1.4.318 a
Appendix
Table 13: Software of the standard-quality runs (Figure 4 ). a Recorded on the same host under the previous Mesa package (0ubuntu0.24.04.2). b The driver and API were queried on 2026-08-10, before an over-the-air update to the HyperOS version listed, which the runs used. c Queried on 2026-11-14 under the same HyperOS version, before the runs.
rank
common
AMD
Intel
1
tiled + native
per-splat + CAS
tiled + native
2
tiled + CAS
exact pair gather
pixel + native
3
pixel + native
per-splat + native
per-splat + native
4
pixel + CAS
pair reduction
pixel + CAS
5
per-splat + native
pixel + CAS
exact pair gather
6
per-splat + CAS
pixel + native
pair reduction
Appendix
Table 14: Gradient-method preference orders. “Pixel” is pixel/subgroup and “native” the native atomic merge.
configuration
selected method
fastest method
slowdown
M5Pro
pixel/subgroup + native
pixel/subgroup + native
1.00
A14
pixel/subgroup + native
pixel/subgroup + native
1.00
Appendix
Table 15: Held-out timing screens at the trained one-million-Gaussian SH3 state of bicycle : 2k updates, three runs per method. Slowdown is the selected method’s step time divided by the fastest feasible method’s.
Figure 8: Screens from the plush-toy sessions of Table 6 (100 frames, demonstration preset): G1-Ultra (top), Adreno740 (middle), and A14 (bottom). Capture panels were saved during recording; the other panels during replay. The G1-Ultra capture screen is turned upright. SIFT, matching, and optimization show the frames on their recorded side; mapping and the viewer orient gravity down.
3DGS only
capture-to-view
temp.
temp.
device
policy
GPU
step
GPU
battery
CPU
GPU
SfM
3DGS
GPU
battery
G1-Ultra
spec
2.70 / 3.50 / 4.21
1.72 b
AnTuTu
2.70 / 3.50 / 4.21
1.72
default
0.65
1
45.3
32.2
1.08 / 1.47 / 1.74
0.57
27.1
7.8
53.6
40.9
Game Turbo
0.56
0.86
44.6
33.0
1.17 / 2.11 / 2.17
0.56
18.9
7.9
58.0
43.9
Appendix
Table 16: Power policies (Sec. I ). Clocks are means (GHz): CPU during SfM matching, slowest cluster first; GPU during 3DGS. Step is speed relative to default; stage times are minutes. End temperatures (°C) average the last 15 s and both runs for 3DGS only, or the last 30 s for capture-to-view. GPU temperature is the hottest reported sensor (one on G1-Ultra , eight on Adreno740 ), so the devices’ readings differ in scope. iOS reports thermal state instead; a dash is unreported. Spec and AnTuTu give maxima: kernel CPU limits matching vendor figures, and observed benchmark clocks, respectively. a An Instruments trace places A14 in the middle of three GPU performance states for about 72% of training, the rest at maximum; iOS reports no clock. b Not vendor specifications: GPU clocks from Comparigon [ 17 , 18 ] ; A14 CPU clocks from The Apple Wiki [ 63 ] (2.99 GHz) and WikiChip [ 67 ] (1.82 GHz).
GPU (°C)
battery (°C)
device
power
Fig. 6
peak
at launch
peak
time (h)
energy (Wh)
mean (W)
battery (%)
GPU clock
Adreno740
battery
a, b, c, e
57.0
32.9
37.4
2.82
16.9
6.0
100 → 54
0.55 GHz throughout
external
a, b, c
57.9
32.9
36.7
2.83
–
–
91 → 84
0.55 GHz throughout
G1-Ultra
battery
a–e
55.5
29.9
41.2
2.45
7.2
3.0
100 → 77
0.60 GHz from 19.8k updates
external
a–d
56.1
30.3
41.3
2.44
–
–
100 → 100
0.60 GHz from 16.1k updates
A14
battery
c, e
–
–
–
2.80
9.2 a
4.2 a
100 → 15
not reported by iOS
Appendix
Table 17: Standard-quality bicycle runs of Figure 6 . Peaks during training are the highest 1-minute GPU mean and highest battery reading. Sensors follow Table 16 . Time covers training; energy and mean power cover unplug to training end, including the screen. Neither Android charger kept up; G1-Ultra drew battery power even at 100%. a Energy over the 2.17 h before charging at 25.5k updates: battery-level drop times the rated 10.78 Wh, an upper bound for an aged battery. b Powered from a computer’s USB port.
GPU (°C)
battery (°C)
device
at launch
peak
at launch
peak
thermal state
total (min)
A14
–
–
–
–
nominal → serious
45.9
G1-Ultra
45.2
66.5
32.4
41.8
normal
35.3
Adreno740
42.8
70.8
32.7
37.7
normal
37.5
Appendix
Table 18: The plush-toy sessions of Table 6 , on external power. Peak is the highest temperature during the session, reached during SfM on both Android devices; the GPU temperature is that of Table 16 . iOS reports no temperature. The captures differ, so the rows do not compare devices.
We present VkSplat, a high-performance, cross-vendor 3D Gaussian Splatting (3DGS) training pipeline implemented fully in Vulkan compute, addressing performance and compatibility limitation of existing training pipelines. With various optimizations, we achieve 3.3× speed and 33% VRAM reduction over CUDA+PyTorch baseline, maintaining quality, and demonstrating compatibility across GPU vendors. To the best of our knowledge, this is the first fully-Vulkan-based 3DGS training pipeline that achieves state-of-the-art performance. Code: https://github.com/harry7557558/vksplat
Although 3D Gaussian Splatting (3DGS) supports real-time rendering, training relies on offline geometry, long-horizon densification, and memory-intensive differentiation. We cast post-capture static-scene training as resource-constrained finite-horizon optimization, where terminal quality depends jointly on the initial representation and reference-consistent gradient execution. To the best of our knowledge, PocketGS is the first publicly documented pipeline to complete geometry construction, Gaussian initialization, and 3DGS optimization on a commodity smartphone. Its I, G warm start combines refined poses, a filtered dense prior, and surface-aligned anisotropic Gaussians. Its T operator replays compositing interactions and scatters gradients to canonical parameters, matching a numerical reference at the reported precision. Controlled 500-step experiments on LLFF, Mip-NeRF 360, and MobileScan show high-quality reconstruction at the device horizon. Across MobileScan, every scene remains below 3 GB peak memory and post-capture processing averages 255.2 s, enabling an on-device capture-to-render workflow. Separate 30k experiments characterize native-schedule behavior.
Training 3D Gaussian Splatting (3DGS) at billion-primitive scale is fundamentally memory-bound: each Gaussian primitive carries a large attribute vector, and the aggregate parameter table quickly exceeds GPU capacity, limiting prior systems to tens of millions of Gaussians on commodity single-GPU hardware. We observe that 3DGS training is inherently sparse and trajectory-conditioned: each iteration activates only the Gaussians visible from the current camera batch, so GPU memory can serve as a working-set cache rather than a persistent parameter store. Building on this insight, we introduce TideGS, an out-of-core training framework that manages parameters across an SSD-CPU-GPU hierarchy via three synergistic techniques: block-virtualized geometry for SSD-aligned spatial locality, a hierarchical asynchronous pipeline to overlap I/O with computation, and trajectory-adaptive differential streaming that transfers only incremental working-set deltas between iterations. Experiments show that TideGS enables training with over one billion Gaussians on a single 24 GB GPU while achieving the best reconstruction quality among evaluated single-GPU baselines on large-scale scenes, scaling beyond prior out-of-core baselines (e.g., approximately 100M Gaussians) and standard in-memory training (e.g., approximately 11M Gaussians).
Chonghao Zhong, Linfeng Shi, Hua Chen +4
Hong Kong University of Science and Technology · Great Wall Motor · Tsinghua University +1