#fluidx3d — Public Fediverse posts

Directly found the first issue im #FluidX3D: #OpenCL __attribute__((opencl_unroll_hint(x))) is not supported on AMD's old #GPU driver. Guaranteeing legacy hardware compatibility means testing on legacy hardware. I'm not giving software rot a chance! 🖖🧐
https://github.com/ProjectPhysX/FluidX3D/commit/b4481ae5bdc96fb062a6db3092af3470e50ccab0

Directly found the first issue im #FluidX3D: #OpenCL __attribute__((opencl_unroll_hint(x))) is not supported on AMD's old #GPU driver. Guaranteeing legacy hardware compatibility means testing on legacy hardware. I'm not giving software rot a chance! 🖖🧐
https://github.com/ProjectPhysX/FluidX3D/commit/b4481ae5bdc96fb062a6db3092af3470e50ccab0

#FluidX3D #CFD v3.7 brings faster Q-criterion isosurface rendering with #OpenCL local memory optimization! 🖖🤠
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.7

Instead of 32 velocities for each #GPU thread, now an 8x8x8 workgroup loads & reuses 11x11x11 velocities in L1$, a 12x VRAM BW reduction.

Fascinating insight: Which thread loads which cell from VRAM to L1$, and which thread renders which grid cell within the workgroup, can be very different!
https://github.com/ProjectPhysX/FluidX3D/blob/master/src/kernel.cpp#L2827-L2956

PS: plugged X-wing Gif in #GitHub preview 🖖😜

#FluidX3D #CFD v3.7 brings faster Q-criterion isosurface rendering with #OpenCL local memory optimization! 🖖🤠
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.7

Instead of 32 velocities for each #GPU thread, now an 8x8x8 workgroup loads & reuses 11x11x11 velocities in L1$, a 12x VRAM BW reduction.

Fascinating insight: Which thread loads which cell from VRAM to L1$, and which thread renders which grid cell within the workgroup, can be very different!
https://github.com/ProjectPhysX/FluidX3D/blob/master/src/kernel.cpp#L2827-L2956

PS: plugged X-wing Gif in #GitHub preview 🖖😜

Newest #IntelArc #GPU family member is here, the Panther Lake Arc B390... and it... purrs? 🖖 🥺 🐈‍⬛
My OpenCL-Benchmark on the B390 measures ~7.4 TFlops FP32 and ~120GB/s memory bandwidth. hw-smi also works with the B390.
#FluidX3D benchmarks here: https://github.com/ProjectPhysX/FluidX3D#single-gpucpu-benchmarks
And the #OpenCL infos:
- Arc B390: https://opencl.gpuinfo.org/displayreport.php?id=6718
- Core Ultra X7 358H: https://opencl.gpuinfo.org/displayreport.php?id=6717

Newest #IntelArc #GPU family member is here, the Panther Lake Arc B390... and it... purrs? 🖖 🥺 🐈‍⬛
My OpenCL-Benchmark on the B390 measures ~7.4 TFlops FP32 and ~120GB/s memory bandwidth. hw-smi also works with the B390.
#FluidX3D benchmarks here: https://github.com/ProjectPhysX/FluidX3D#single-gpucpu-benchmarks
And the #OpenCL infos:
- Arc B390: https://opencl.gpuinfo.org/displayreport.php?id=6718
- Core Ultra X7 358H: https://opencl.gpuinfo.org/displayreport.php?id=6717

#FluidX3D #CFD has reached ⭐ 5000 Stargazers on #GitHub! 🖖🥳
Grid refinement update is still in development, I haven't forgotten... ⬜◻️◽▫️
https://github.com/ProjectPhysX/FluidX3D

#FluidX3D #CFD has reached ⭐ 5000 Stargazers on #GitHub! 🖖🥳
Grid refinement update is still in development, I haven't forgotten... ⬜◻️◽▫️
https://github.com/ProjectPhysX/FluidX3D

Finally Intel #GPU support on Linux too. Watch all the metrics go brrr in multi-GPU #FluidX3D #CFD workload! Will #opensource soon™️

Hardening against the myriads of broken counters in all those bugged APIs was a long shot. 🖖🫠

____________ | Windows | #Linux |
CPU / RAM | ✅️️WinAPI | ✅️️/proc |
#Nvidia GPU | ✅️️NVML | ✅️️NVML |
#Intel GPU | ✅IGCL | ✅SYSMAN |
#AMD GPU | ✅️️️️ADLX | ✅️️️️AMDSMI |

Finally Intel #GPU support on Linux too. Watch all the metrics go brrr in multi-GPU #FluidX3D #CFD workload! Will #opensource soon™️

Hardening against the myriads of broken counters in all those bugged APIs was a long shot. 🖖🫠

____________ | Windows | #Linux |
CPU / RAM | ✅️️WinAPI | ✅️️/proc |
#Nvidia GPU | ✅️️NVML | ✅️️NVML |
#Intel GPU | ✅IGCL | ✅SYSMAN |
#AMD GPU | ✅️️️️ADLX | ✅️️️️AMDSMI |

#FluidX3D #CFD v3.6 is out! This release accumulates a number of small improvements over the last months. Most notably, better interactive graphics support on #macOS with XQuartz. Have fun! 🖖😎🌊🍏
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.6

#FluidX3D #CFD v3.6 is out! This release accumulates a number of small improvements over the last months. Most notably, better interactive graphics support on #macOS with XQuartz. Have fun! 🖖😎🌊🍏
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.6

Finally, the bars for #Intel hardware in the #FluidX3D #mermaid gantt performance chart are BLUE, as nature intended it.
https://github.com/ProjectPhysX/FluidX3D?tab=readme-ov-file#single-gpucpu-benchmarks

Finally, the bars for #Intel hardware in the #FluidX3D #mermaid gantt performance chart are BLUE, as nature intended it.
https://github.com/ProjectPhysX/FluidX3D?tab=readme-ov-file#single-gpucpu-benchmarks

Some experimentation with ```mermaid ...``` charts in #GitHub #markdown. Turns out you can hack the formatting on the quadrantChart to turn it into an xy-scatter plot with individual point size/coloring/labeling.🖖🧐
First plot is datasheet memory bandwidth vs. FP32 TFlops/s, second plot is #FluidX3D performance vs. bandwidth, for lots of #GPU​/​#CPU hardware.

Some experimentation with ```mermaid ...``` charts in #GitHub #markdown. Turns out you can hack the formatting on the quadrantChart to turn it into an xy-scatter plot with individual point size/coloring/labeling.🖖🧐
First plot is datasheet memory bandwidth vs. FP32 TFlops/s, second plot is #FluidX3D performance vs. bandwidth, for lots of #GPU​/​#CPU hardware.

Here's me demoing #Intel Arc Pro B60 #GPU workstations at #SC25 in St. Louis, runnig SolidWorks and #FluidX3D! 🖖😎
https://www.youtube.com/watch?v=Z8yxiyXTi7I

Here's me demoing #Intel Arc Pro B60 #GPU workstations at #SC25 in St. Louis, runnig SolidWorks and #FluidX3D! 🖖😎
https://www.youtube.com/watch?v=Z8yxiyXTi7I

Making progress on >top secret #FluidX3D update< but still a long way to go 🖖🧐

Making progress on >top secret #FluidX3D update< but still a long way to go 🖖🧐

#SC25 in wonderful St. Louis was a blast! I showcased this FluidX3D-on-Arc-Pro workstation demo there, met so many friends, had exciting conversations with all sorts of #HPC enthusiasts. And I finally saw the Cardinals - the bright red/brown birds that became the symbol of the city's famous Baseball team. 🐦

This is Flynn's #Tron Light Cycle simulated in #FluidX3D #CFD, on 4x #Intel #ArcPro B60 #GPU​s with 24GB VRAM each, at a massive 1.8 Billion grid cells resolution. 🖖🤠
https://www.youtube.com/watch?v=5kZ3MmNOLoE

#SC25 in wonderful St. Louis was a blast! I showcased this FluidX3D-on-Arc-Pro workstation demo there, met so many friends, had exciting conversations with all sorts of #HPC enthusiasts. And I finally saw the Cardinals - the bright red/brown birds that became the symbol of the city's famous Baseball team. 🐦

This is Flynn's #Tron Light Cycle simulated in #FluidX3D #CFD, on 4x #Intel #ArcPro B60 #GPU​s with 24GB VRAM each, at a massive 1.8 Billion grid cells resolution. 🖖🤠
https://www.youtube.com/watch?v=5kZ3MmNOLoE

That's one small step for the cow - one giant leap for #CFD! 🐄
The eagle-eyed among you will spot what I'm working on with #FluidX3D. 🖖🤠

That's one small step for the cow - one giant leap for #CFD! 🐄
The eagle-eyed among you will spot what I'm working on with #FluidX3D. 🖖🤠

Small but notable quality-of-life update for #FluidX3D - a function to calculate the center of mass for an imported .stl triangle mesh. This makes it a lot easier to set up for example 3-bladed rotors im #CFD.
https://github.com/ProjectPhysX/FluidX3D/commit/51203b425f24c0e31241f7f193fce9e07b972a06

Small but notable quality-of-life update for #FluidX3D - a function to calculate the center of mass for an imported .stl triangle mesh. This makes it a lot easier to set up for example 3-bladed rotors im #CFD.
https://github.com/ProjectPhysX/FluidX3D/commit/51203b425f24c0e31241f7f193fce9e07b972a06

8x AMD Instinct #MI355X (288GB @8TB/s) take back the lead over 8x Nvidia #B200 (180GB @8TB/s) in #FluidX3D #CFD, achieving 362k MLUPs/s (vs. 219k MLUPs/s). Thanks to Jon Stevens from Hot Aisle to run the benchmarks! 🖖😊

In single-GPU, both perform about the same, but in 8x #GPU config, MI355X is 65% faster. The difference comes from PCIe bandwidth - MI355X does 55GB/s, B200 only 14GB/s. #Nvidia leaves a lot of perf on the table by not exposing #NVLink P2P to #OpenCL.

https://github.com/ProjectPhysX/FluidX3D

8x AMD Instinct #MI355X (288GB @8TB/s) take back the lead over 8x Nvidia #B200 (180GB @8TB/s) in #FluidX3D #CFD, achieving 362k MLUPs/s (vs. 219k MLUPs/s). Thanks to Jon Stevens from Hot Aisle to run the benchmarks! 🖖😊

In single-GPU, both perform about the same, but in 8x #GPU config, MI355X is 65% faster. The difference comes from PCIe bandwidth - MI355X does 55GB/s, B200 only 14GB/s. #Nvidia leaves a lot of perf on the table by not exposing #NVLink P2P to #OpenCL.

https://github.com/ProjectPhysX/FluidX3D

#FluidX3D #CFD v3.5 upgrade is out, for free! I have implemented multi-​#GPU support for the particles extension - now all model extensions are supported with multi-GPU. Also: faster particle force spreading, more documentation, updated FAQs, and an #OpenCL bug fix for ARM GPUs. Have fun everyone! 🖖😎

Full release notes on #GitHub 👉 https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.5

#FluidX3D #CFD v3.5 upgrade is out, for free! I have implemented multi-​#GPU support for the particles extension - now all model extensions are supported with multi-GPU. Also: faster particle force spreading, more documentation, updated FAQs, and an #OpenCL bug fix for ARM GPUs. Have fun everyone! 🖖😎

Full release notes on #GitHub 👉 https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.5

#FluidX3D is now 3 years out in the wild! This fun hobby project has turned into the largest #CFD software on #Github, with ⭐4.6k Stars. Community feedback has been overwhelming. 🖖🤯

Since last year I've enabled using full multi-TB memory capacity on CPUs/iGPUs, added #GPU-accelerated force/torque summation, optimized vtk export, and lots of bug fixes, keeping the # of known bugs at 0.

More updates are planned, as well as a roadmap to commercial use. Let's keep growing!

https://www.youtube.com/watch?v=6enTZjO9fbQ

#FluidX3D is now 3 years out in the wild! This fun hobby project has turned into the largest #CFD software on #Github, with ⭐4.6k Stars. Community feedback has been overwhelming. 🖖🤯

Since last year I've enabled using full multi-TB memory capacity on CPUs/iGPUs, added #GPU-accelerated force/torque summation, optimized vtk export, and lots of bug fixes, keeping the # of known bugs at 0.

More updates are planned, as well as a roadmap to commercial use. Let's keep growing!

https://www.youtube.com/watch?v=6enTZjO9fbQ

My keynote talk at #IXPUG #ISC25 is finally online! 🖖🤠
GigaScale #FluidX3D #CFD Simulations from Communication and IO Perspective
https://www.youtube.com/watch?v=uAIkpcX5EFc&list=PLA-vfTt7YHI2HEFrpzPhhQ8PhiztKhHU8

My keynote talk at #IXPUG #ISC25 is finally online! 🖖🤠
GigaScale #FluidX3D #CFD Simulations from Communication and IO Perspective
https://www.youtube.com/watch?v=uAIkpcX5EFc&list=PLA-vfTt7YHI2HEFrpzPhhQ8PhiztKhHU8

PCIe experimemts: What is better, R-B-G or R-G-B? 🖖🤔🌈

I forgot I had the newer ProArt Z790 not Z690. Both have 5.0 x8/x8, yet the 3rd PCIe slot is a 4.0 x4 not 3.0 x4.

GPUs are
R - AMD 7700 XT, 4.0 x16
G - Nvidia Titan Xp, 3.0 x16
B - Intel B580, 4.0 x8

R: 4.0 x8 (16 GB/s)
B: 4.0 x8 (16 GB/s)
G: 3.0 x4 (4 GB/s)

R: 4.0 x8 (16 GB/s)
G: 3.0 x8 (8 GB/s)
B: 4.0 x4 (8 GB/s)

R-G-B is less total BW but better balanced. #FluidX3D #CFD perf is ~indifferent though, indicating 4GB/s is sufficient.

PCIe experimemts: What is better, R-B-G or R-G-B? 🖖🤔🌈

I forgot I had the newer ProArt Z790 not Z690. Both have 5.0 x8/x8, yet the 3rd PCIe slot is a 4.0 x4 not 3.0 x4.

GPUs are
R - AMD 7700 XT, 4.0 x16
G - Nvidia Titan Xp, 3.0 x16
B - Intel B580, 4.0 x8

R: 4.0 x8 (16 GB/s)
B: 4.0 x8 (16 GB/s)
G: 3.0 x4 (4 GB/s)

R: 4.0 x8 (16 GB/s)
G: 3.0 x8 (8 GB/s)
B: 4.0 x4 (8 GB/s)

R-G-B is less total BW but better balanced. #FluidX3D #CFD perf is ~indifferent though, indicating 4GB/s is sufficient.

#FluidX3D #CFD v3.4 is out! With an AMD #GPU finally in my posession, I can now test FluidX3D locally on #AMD​/​#Intel​/​#Nvidia GPUs within the same PC, to guarantee full compatibility. This allowed me to identify and fix 2 critical bugs that were coding mistakes on my side yet somehow only exposed with AMD's driver. Be sure to update! 🖖🤠
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.4

#FluidX3D #CFD v3.4 is out! With an AMD #GPU finally in my posession, I can now test FluidX3D locally on #AMD​/​#Intel​/​#Nvidia GPUs within the same PC, to guarantee full compatibility. This allowed me to identify and fix 2 critical bugs that were coding mistakes on my side yet somehow only exposed with AMD's driver. Be sure to update! 🖖🤠
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.4

Finally I can "SLI" AMD+Intel+Nvidia #GPU​s at home! I simulated this crow in flight at 680M grid cells in 36GB VRAM, pooled together from
- 🟥 #AMD Radeon RX 7700 XT 12GB (RDNA3)
- 🟦 #Intel Arc B580 12GB (Battlemage)
- 🟩 #Nvidia Titan Xp 12GB (Pascal)
My #FluidX3D #CFD software can pool the VRAM of any combination of any GPUs together via #OpenCL.
#Krähenliebe #birds #crow
https://www.youtube.com/watch?v=1z5-ddsmAag

Finally I can "SLI" AMD+Intel+Nvidia #GPU​s at home! I simulated this crow in flight at 680M grid cells in 36GB VRAM, pooled together from
- 🟥 #AMD Radeon RX 7700 XT 12GB (RDNA3)
- 🟦 #Intel Arc B580 12GB (Battlemage)
- 🟩 #Nvidia Titan Xp 12GB (Pascal)
My #FluidX3D #CFD software can pool the VRAM of any combination of any GPUs together via #OpenCL.
#Krähenliebe #birds #crow
https://www.youtube.com/watch?v=1z5-ddsmAag

Live long and prosper, Jensen Huang! 🖖🧐
#GTCParis #FluidX3D

Live long and prosper, Jensen Huang! 🖖🧐
#GTCParis #FluidX3D

Paris is amazing! Had a great time at #Teratec25 demoing #FluidX3D #CFD on #Intel #Xeon6, and meeting so many #HPC people there! 🖖😁

Paris is amazing! Had a great time at #Teratec25 demoing #FluidX3D #CFD on #Intel #Xeon6, and meeting so many #HPC people there! 🖖😁

#FluidX3D #CFD v3.3 is out! .vtk export now converts and writes data in chunks, to reduce memory footprint and time for large memory allocation. This is helpful especially with large CPU simulations where the simulation itself fills almost the entire RAM capacity. Also bug fixes. Have fun! 🖖😎
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.3

#FluidX3D #CFD v3.3 is out! .vtk export now converts and writes data in chunks, to reduce memory footprint and time for large memory allocation. This is helpful especially with large CPU simulations where the simulation itself fills almost the entire RAM capacity. Also bug fixes. Have fun! 🖖😎
https://github.com/ProjectPhysX/FluidX3D/releases/tag/v3.3

Battle of the giants: Nvidia #Blackwell B200 takes the lead in FluidX3D CFD performance

#Nvidia #B200 just launched, and I'm one of the first people to benchmark 8x B200 via Shadeform, in a WhiteFiber server with 2x #Intel #Xeon6 6960P 72-core CPUs. 🖖😋

8x Nvidia B200 go head-to-head with 8x #AMD #MI300X in the #FluidX3D #CFD benchmark, winning overall (with FP16S storage) at 219300 MLUPs/s (~17TB/s combined VRAM bandwidth), but losing in FP32 & FP16C storage. 8x MI300X achieve 204924 MLUPs/s.

Battle of the giants: Nvidia #Blackwell B200 takes the lead in FluidX3D CFD performance

#Nvidia #B200 just launched, and I'm one of the first people to benchmark 8x B200 via Shadeform, in a WhiteFiber server with 2x #Intel #Xeon6 6960P 72-core CPUs. 🖖😋

8x Nvidia B200 go head-to-head with 8x #AMD #MI300X in the #FluidX3D #CFD benchmark, winning overall (with FP16S storage) at 219300 MLUPs/s (~17TB/s combined VRAM bandwidth), but losing in FP32 & FP16C storage. 8x MI300X achieve 204924 MLUPs/s.

My #IWOCL 2025 Keynote presentation is online! 🖖🧐
Scaling up #FluidX3D #CFD beyond 100 Billion cells on a single computer - a story about the true cross-compatibility of #OpenCL
https://www.youtube.com/watch?v=Sb3ibfoOi0c&list=PLA-vfTt7YHI2HEFrpzPhhQ8PhiztKhHU8&index=1
Slides: https://www.iwocl.org/wp-content/uploads/iwocl-2025-moritz-lehmann-keynote.pdf

My #IWOCL 2025 Keynote presentation is online! 🖖🧐
Scaling up #FluidX3D #CFD beyond 100 Billion cells on a single computer - a story about the true cross-compatibility of #OpenCL
https://www.youtube.com/watch?v=Sb3ibfoOi0c&list=PLA-vfTt7YHI2HEFrpzPhhQ8PhiztKhHU8&index=1
Slides: https://www.iwocl.org/wp-content/uploads/iwocl-2025-moritz-lehmann-keynote.pdf