Internal benchmark

Distributed rendering, 38x faster

We rendered a 120-frame Full HD Blender scene. On a laptop CPU, one frame took 2 minutes 28 seconds, which puts all 120 frames at about 296 minutes. Split across four GPU nodes on Symbola's GPU/CPU Grid, the full job finished in 7.76 minutes, end to end. We also moved a 10 GB file through Bandwidth Pool at three times the speed of a direct download.

38.15x
faster rendering
97.4%
less wall-clock time
3.02x
faster 10 GB download

GPU/CPU Grid: distributed Blender rendering

The job: 120 frames at 1920 × 1080 in Blender. In the baseline, a laptop rendered on its CPU alone; we timed the first frame and multiplied by 120. In the Symbola run, GPU/CPU Grid split the frame range into four chunks of 30, rendered them in parallel on four GPU nodes and reassembled the output on the laptop.

RunSetupWall-clock time
BaselineLaptop, 10th-gen Intel Core i5, CPU only. Frame 1 measured at 148 s, total estimated for 120 frames296 min (4.93 h), estimated
Symbola4 laptops across India, each with an NVIDIA RTX 4050 Laptop GPU (6 GB VRAM) and 8 GB RAM, 30 frames per node7.76 min, measured end to end
SettingValue
SoftwareBlender 4.4.3
Render engineCycles
WorkersUbuntu, rendering in Docker; frames 1–30, 31–60, 61–90 and 91–120, one range per node
Scene“Candy Bounce”, a Blender Geometry Nodes demo
Output120 frames, 1920 × 1080, 24 fps, PNG
Samples60 per frame
Run date6 January 2026

The 7.76 minutes covers everything from submitting the job on the laptop to having all 120 frames back on it. GPU rendering itself took 61 to 101 seconds per node. The rest went to packaging the frames, merging them in order and moving them back and forth. In this test those files travelled over ordinary connections, not through Bandwidth Pool. Routing them through it is our next milestone.

Watch the recorded rendering run.

Bandwidth Pool: 10 GB transfer

The job: download a 10 GB file from Datapacket. The baseline used a single direct connection with curl. In the Symbola run, Bandwidth Pool streamed it over three encrypted paths through three worker nodes at the same time.

RunSetupAverage speedTime
BaselineDirect download, 1 route430.50 Mbps176 s
Symbola3 nodes, 3 parallel paths1,301.28 Mbps61.48 s

Watch the recorded transfer run.

Energy: a projection, measured in the pilot

These figures are a projection from the measured run times and the typical power draw of the hardware involved.

MachineAssumed power while working
Baseline laptop, Core i5, CPU only45 to 65 W
Each of the four RTX 4050 laptops100 to 150 W while rendering
Client laptop during the Symbola run45 to 65 W
RunEstimated energy
Baseline laptop, about 296 min0.22 to 0.32 kWh
Symbola, 7.76 min0.016 to 0.086 kWh
DifferenceAbout 60% to 95% less

The low Symbola figure counts only the time the four nodes spent rendering. The high figure assumes all four ran at full power for the whole 7.76 minutes, including transfers. The full working is in section 10.5 of the whitepaper. We measure the runs with a meter in the pilot and publish the figures here.

How we measured

Both runs were recorded on video and logged by our own benchmark framework, which captures telemetry from every node involved. The telemetry consoles on the homepage demo show the logged output alongside each recording.

Want the raw logs or the full hardware list? Ask us and we will send them.

Limits, stated plainly

  • We ran these tests ourselves. They have not been reproduced by a third party yet, and we would welcome anyone who wants to.
  • The rendering baseline is a laptop rendering on CPU, and its total is estimated from one measured frame. The Symbola nodes render on GPUs, so most of the 38x comes from GPU hardware rather than from distribution alone. A comparison against a single GPU is next.
  • The energy figures are estimates based on typical power draw, not meter readings.
  • Rendering separate frames is a naturally parallel job. Workloads that need constant communication between nodes will see smaller gains.
  • Both runs were finished jobs, a render and a file transfer, not live streams. Continuous real-time workloads are the next milestone on the roadmap, and we will benchmark them the same way.

The products behind the numbers

Read how GPU/CPU Grid splits and schedules compute jobs, and how Bandwidth Pool spreads a transfer over several paths.

Have a workload you want to test on Symbola?