Uploaded February 2024 | Updated November 2026, 6 hours ago
This is wild: FluidX3D can "SLI" together an 🔵 Intel Arc A770 + 🟢 Nvidia Titan Xp, pooling 12GB+12GB of their VRAM for one large 450 million cell CFD simulation. Top half computed+rendered on A770, bottom half computed+rendered on Titan Xp. They seamlessly communicate over PCIe. Performance is ~1.7x of what either GPU could do on its own.
OpenCL shows its true power here - a single implementation works on literally all GPUs at full performance, even at the same time. I have specifically designed FluidX3D for cross-vendor multi-GPU, to allow combining any GPUs as long as VRAM capacity and bandwidth are similar.
Now that I have some new hardware, I can finally demonstrate this in practice. This setup is turbulence created by a sphere at Re=1M. 532×1600×532 resolution in 2×12GB VRAM, 64k time steps, 1.5h for compute+rendering.
Find the FluidX3D CFD software on GitHub: github.com/ProjectPhysX/FluidX3D
How does cross-vendor multi-GPU work?
Each GPU computes only half of the simulation box. VRAM capacity and bandwidth are similar (A770: 16GB@560GB/s, Titan Xp: 12GB@548GB/s) such that the compute time for both domains is similar. Where the two GPU domains (each 8.6 GB in size) touch, some data has to be exchanged. These layers (8.5 MB in size) are first extracted within each GPU's VRAM into transfer buffers. The transfer buffers are copied from VRAM to CPU RAM over PCIe (A770: PCIe 4.0 x8 (~8GB/s), Titan Xp: PCIe 3.0 x8 (~4GB/s). The CPU waits for all transfer buffers to arrive, and then only swaps their pointers. Afterwards, transfer buffers are copied back over PCIe to the GPUs, are inserted back into the domains within VRAM, and each GPU can again compute LBM on its own domain. Because OpenCL only needs the generic PCIe interface and not proprietary SLI/Crossfire/NVLink/InfinityFabric, this works with any combination of Intel/Nvidia/AMD GPUs.
Here is the simplified multi-GPU communication architecture illustrated:
.-----------------------------------------------------------------.
| GPU 0 |
| LBM Domain 0 |
'-----------------------------------------------------------------'
| selective /|\
\|/ in-VRAM copy |
.-------------------------------------------------------.
| GPU 0 - Transfer Buffer 0 |
'-------------------------------------------------------'
| PCIe /|\
\|/ copy |
.-------------------------. .-------------------------.
| CPU - TB 0 | | CPU - TB 1 |
'-------------------------'\ /'-------------------------'
pointer X swap
.-------------------------./ \.-------------------------.
| CPU - TB 1 | | CPU - TB 0 |
'--------------------------' '--------------------------'
/|\ PCIe |
| copy \|/
.-------------------------------------------------------.
| GPU 1 - Transfer Buffer 1 |
'-------------------------------------------------------'
/|\ selective |
| in-VRAM copy \|/
.-----------------------------------------------------------------.
| GPU 1 |
| LBM Domain 1 |
'-----------------------------------------------------------------'
|
domain synchronization barrier
|
-----------------------------------------------------→ time
Timestamps
0:00 intro
0:04 simulation
#FluidX3D #CFD #OpenCL #Intel #IntelArc #Nvidia #FluidSimulation #turbulence
This is wild: FluidX3D can "SLI" together an 🔵 Intel Arc A770 + 🟢 Nvidia Titan Xp, pooling 12GB+12GB of their VRAM for one large 450 million cell CFD simulation. Top half computed+rendered on A770, bottom half computed+rendered on Titan Xp. They seamlessly communicate over PCIe. Performance is ~1.7x of what either GPU could do on its own.
OpenCL shows its true power here - a single implementation works on literally all GPUs at full performance, even at the same time. I have specifically designed FluidX3D for cross-vendor multi-GPU, to allow combining any GPUs as long as VRAM capacity and bandwidth are similar.
Now that I have some new hardware, I can finally demonstrate this in practice. This setup is turbulence created by a sphere at Re=1M. 532×1600×532 resolution in 2×12GB VRAM, 64k time steps, 1.5h for compute+rendering.
Find the FluidX3D CFD software on GitHub: github.com/ProjectPhysX/FluidX3D
How does cross-vendor multi-GPU work?
Each GPU computes only half of the simulation box. VRAM capacity and bandwidth are similar (A770: 16GB@560GB/s, Titan Xp: 12GB@548GB/s) such that the compute time for both domains is similar. Where the two GPU domains (each 8.6 GB in size) touch, some data has to be exchanged. These layers (8.5 MB in size) are first extracted within each GPU's VRAM into transfer buffers. The transfer buffers are copied from VRAM to CPU RAM over PCIe (A770: PCIe 4.0 x8 (~8GB/s), Titan Xp: PCIe 3.0 x8 (~4GB/s). The CPU waits for all transfer buffers to arrive, and then only swaps their pointers. Afterwards, transfer buffers are copied back over PCIe to the GPUs, are inserted back into the domains within VRAM, and each GPU can again compute LBM on its own domain. Because OpenCL only needs the generic PCIe interface and not proprietary SLI/Crossfire/NVLink/InfinityFabric, this works with any combination of Intel/Nvidia/AMD GPUs.
Here is the simplified multi-GPU communication architecture illustrated:
.-----------------------------------------------------------------.
| GPU 0 |
| LBM Domain 0 |
'-----------------------------------------------------------------'
| selective /|\
\|/ in-VRAM copy |
.-------------------------------------------------------.
| GPU 0 - Transfer Buffer 0 |
'-------------------------------------------------------'
| PCIe /|\
\|/ copy |
.-------------------------. .-------------------------.
| CPU - TB 0 | | CPU - TB 1 |
'-------------------------'\ /'-------------------------'
pointer X swap
.-------------------------./ \.-------------------------.
| CPU - TB 1 | | CPU - TB 0 |
'--------------------------' '--------------------------'
/|\ PCIe |
| copy \|/
.-------------------------------------------------------.
| GPU 1 - Transfer Buffer 1 |
'-------------------------------------------------------'
/|\ selective |
| in-VRAM copy \|/
.-----------------------------------------------------------------.
| GPU 1 |
| LBM Domain 1 |
'-----------------------------------------------------------------'
|
domain synchronization barrier
|
-----------------------------------------------------→ time
Timestamps
0:00 intro
0:04 simulation
#FluidX3D #CFD #OpenCL #Intel #IntelArc #Nvidia #FluidSimulation #turbulence



![PhysX3D [GRAVITY SIMULATION] 25k body galaxy @20x playback speed
This is PhysX3D simulating a galaxy containing 25000 stars. For calculation brute force and RK4 integration is being used on a notebook GPU (Nvidia GTX 960M, 1.54TFlops). The GPU calculates roughly 3.5 Billion particle interactions every second.
This simulation took 12 minutes to calculate.
More information at: http://www.projectphysx.de
Music: Puddle of Infinity - Open Sea Morning PhysX3D [GRAVITY SIMULATION] 25k body galaxy @20x playback speed](https://i.ytimg.com/vi/TmBf2DI-3cE/mqdefault.jpg)



![PhysX3D [GRAVITY SIMULATION] Der zweite Mond der Erde – 2016 HO3
Die Erde hat einen zweiten Mond, 2016 HO3 wird er bezeichnet. Er ist sehr weit außen, 70 Mal so weit wie der Erdmond. Er ist auch sehr klein, nur ca. 70 Meter im Durchmesser, und deshalb nicht einmal mit einem Amateurteleskop sichtbar. Aber es gibt ihn, das hat die NASA bestätigt, und er umkreist die Erde, was ihn zu einem Mond macht. Aufgrund der großen Distanz zur Erde ist seine Umlaufbahn jedoch nicht stabil, sodass er diese in etwa 325 Jahren verlassen wird und zum Erdtrabanten in einer Hufeisenumlaufbahn wird. Klingt komisch, ist es auch. Ich hab’s genauer unter die Lupe genommen und seine Bahn für die nächsten 1300 Jahre vorausberechnet.
Mehr Informationen auf: http://www.projectphysx.de
Songs:
M83 – Outro
M83 – Fearful Odds
M83 – Oblivion PhysX3D [GRAVITY SIMULATION] Der zweite Mond der Erde – 2016 HO3](https://i.ytimg.com/vi/V7mHQ0L4Wkc/mqdefault.jpg)

![FluidX3D [FLUID SIMULATION] Real Time Prototype Demo
Simulating a liquid is usually reserved for supercomputers. However, in the last few years, the amount of calculating power provided by consumer graphics cards skyrocketed to the point where well optimized code today can result in a successful real time simulation. As a rule of thumb, one GPU is about as fast as 100 multi-core CPUs for well parallelizable code.
This is an optimized OpenCL implementation of the barebones Lattice Boltzmann Method (LBM) in its simplest 3D form with the D3Q15 velocity set, optimized to run parallelized on the GPU.
All the shown demos run in real time on an old laptop GPU (GTX 960M with 640 CUDA cores @1200MHz).
The blue curves represent integrated streamlines through the velocity field, which itself is updated by the GPU approximately 50 times per second.
The maximum lattice size is limited by the amount of VRAM to 256x256x256, at which the program fills 2.6GB of graphics memory.
For more information about FluidX3D and my other projects, visit
http://www.projectphysx.de
Music: Death & Rebirth - Lotus Eater FluidX3D [FLUID SIMULATION] Real Time Prototype Demo](https://i.ytimg.com/vi/Vv-ZoJudJq0/mqdefault.jpg)
