Posted on

GPU propagation engine

5G cell

Our fast GPU engine is perfect for modelling wireless coverage

We have developed the next generation of fast radio simulation engines for urban modelling with NVIDIA CUDA technology and Graphics Processing Units (GPUs).

The engine was made to meet demand across many sectors, especially FWA, 5G and CBRS for speed and accuracy.

As well as fast viewsheds, it enables a new automated best-site-analysis capability, which will accelerate site selection and improve efficiency whilst keeping a human in the loop.

Designed for 5G

5G networks are much denser than legacy standards due to the limited range of mmWave signals, necessary for high bandwidth data. The same limitation means these signals are very sensitive to obstructions, and Line of Sight (LOS) coverage is essential for performance.

With 1 metre accuracy and support for LiDAR, 3D clutter and custom clutter profiles, you can model rural and urban areas with high precision.

Area coverage

The simplest mode is a fast “2.5D” viewshed (with a path loss model) which creates a point-to-multipoint heatmap around a given site using LiDAR data. Ours has better Physics than some of the “line of sight” eye candy on the market and doesn’t have trouble with Sub-GHz frequencies which are harder to model accurately.

This is up to 50 times faster than our multi-threaded CPU engine, SLEIPNIR.

GPU demo January 2022

In this mode we can do diffraction and material attenuation with our custom clutter classes.

Best site analysis

Best Site Analysis (BSA) is a monte-carlo analysis technique across a wide area of interest to identify the best locations for a transmitter. This can be done quickly with a new /bsa API call. The output will identify optimal sites, and just as important, inefficient sites.

This feature is powerful for IoT gateway placement, 5G deployments and ad-hoc networking where the best site might presently be determined by a map study based on contours as opposed to a LiDAR model.

Best Site Analysis on ATAK

High speed

Our GPU engine is up to 50 times faster through the API than the current (CPU) engine SLEIPNIRTM

By harnessing the power of high performance graphics cards, we are able to complete high resolution LiDAR plots in near real time, negating the need for a “go” button, or even a progress bar. This speed enables API integration with vehicles and robots which will need to model wireless propagation quickly to make better decisions, especially when they’re off the grid. It was designed around consumer grade cards like the GeForce series but supports enterprise Tesla grade cards due to our card agnostic design.


Our implementation is efficient by design. We want speed to model wireless coverage but not if it requires kilowatts of power. During testing we worked with older GeForce consumer cards and were able to model millions of points in several milliseconds with less than 50W of power. Or in other words, the same power as flicking a light bulb on and off.

Any fool can buy large cards and waste electricity, but we’re proud to have a solution which is fast and economical. This also means it can be run on a laptop as it’s available now as our SOOTHSAYER product.

Open API

The GPU engine is an “engine” parameter in our /area API so you can use it from any interface (or your own custom interface) by setting the engine option in the request body. The OpenAPI 3.0 compliant API returns JSON which contains a PNG image so for existing API integrations using our PNG layers there will be no code changes required to enable it.

Self-hosted GPU server

Instead of buying milk every month you can buy the cow. We also sell SOOTHSAYER which is a self-hosted server with our GPU engine onboard. It supports NVidia GPU cards from Maxwell architecture onwards and most enterprise Hypervisors like ESXi and Proxmox. You get to use your existing LiDAR data too, so you’re not buying it twice.

To see how easy it is to setup a GPU card with SOOTHSAYER we’ve made a video:



Using GPU cards to model Physics, including EM propagation, is an established concept dating back 20 years, despite businessmen claiming otherwise. What is novel here, is making this exciting technology accessible to users priced out of premium tools.

Staying true to CloudRF’s accessible and affordable principles, we’ve included it in our service as an optional processing engine.

CloudRF is a member of the NVIDIA inception program

Posted on

Keeping motorsport smooth

A motorsport customer invited us to a track day to observe a peculiar RF problem… High resolution ‘dashcam’ video feeds have become standard in motorsport with multiple cameras present on vehicles and drivers. Unlike a consumer dashcam, these real-time video feeds use TV broadcasting radio links to relay a signal from the vehicle to the video processing facility via track-side receivers. The problem is Motorbikes with video feeds were experiencing RF difficulty on bends despite being close to high gain receiving antennas. This issue was investigated with the CloudRF API which revealed the following findings:

  • Lean angle has a direct impact on signal quality
  • Adjacent riders will attenuate signals
  • A bike at full tilt will experience antenna polarisation loss
  • Mast heights must be wavelength x (distance/4) to compensate for dynamic losses
  • Low masts, like tight boots, are terrible!

Doppler shift

The system must contend with several challenges; firstly the doppler effect caused by a shifting emitter. This effect is negligible at slower speeds but can cause reception issues as the vehicle increases speed and the frequency appears off tune. A superbike moving at 100mph would result in as much as a 5% shift in frequency which depending on the location of the receiver could be an increase or decrease of 1MHz. The effect can be modelled and managed with frequency tracking receivers designed to overcome this high speed issue.

Lightweight RF links

Weight is critical in maintaining a competitive edge so the RF links employed are low power emitters, using the vehicle’s native electrics. Licensing restrictions also limit the maximum allowed power output. The standard used in this case is ISDB-T which is an MPEG-4 H-264 high definition video stream which at full rate employs QAM-64 modulation. The H-264 quality High Definition (HD) video feeds people are accustomed to require a high signal-to-noise ratio of at least 20dB which is achieved by careful deployment of track-side high gain antennas and dedicated broadcasting spectrum at 2.3GHz. The system uses 7MHz of bandwidth (broken down into sub-carriers) which has an absolute noise floor of -105dBm. Adding the necessary 20dB SNR gives -85dBm which with the addition of 10dB of environmental (7dB) and receiver (3dB) noise gives a target threshold of -75dBm. The antenna on the motorbike is a shark fin, vertically polarised design mounted on the tail of the bike behind the rider. For the purposes of this investigation the antenna has been modeled with 1dBi gain and and an ERP of 18dBm / 65mW, equivalent to just under a consumer WiFi router. The video broadcast unit is concealed nearby within the bike’s tail with minimal cabling between the antenna for tidiness and maximum efficiency. The track-side antennas would be directional antennas with at least 10dBi of forward gain. These would be positioned at key points on the race track for maximum benefit. The siting of these antennas is where CloudRF is used to test options.


Using GPS data from races, it was found that there could easily be 8 motorbikes in a tight group on a bend. As the bikes all take the bend, several changes occur which all impair RF propagation, resulting in disruption to the smooth HD feed:

1. Rider attenuation

A significant change to consider is the increase in environmental attenuation caused by the crowd of riders. At 2.3GHz the human body will absorb 3dB of RF power. Assuming there are 3 bikes between the rider and the receiver this could be a substantial +9dB of attenuation – comparable to a brick wall (7dB) at this wavelength.

2. Antenna tilt

As bikes lean on a bend so do their vertically polarised antennas. As an antenna deviates from its optimal polarity (vertical) to horizontal it loses power up to 3dB at full tilt (90 degrees). If a bike is at half tilt, polarisation loss will cost the RF link 1.5dB.

3. Antenna height

Coupled with tilt, the height of the antenna above the earth will reduce from ~100cm to as little as ~50cm. This will reduce its effectiveness as more of the key fresnel zone will be attenuated by the earth. Using a fresnel zone calculator the fresnel zone radius for a 2.3GHz link over 300m is 3 metres. Elevating the track side antennas on masts is one way to overcome this issue but when one end of the link is so near the earth the (tower) elevation must be much higher than the fresnel radius if it is to clear the earth completely as these profile images demonstrate. Modelling using the Irregular terrain model which considers fresnel attenuation shows substantial loss caused by minor reductions in the (bike) antenna height. As you can see in the path profile below, the curved fresnel zone clips the earth which introduces attenuation.

Modelling the problem with the CloudRF API

The customer wondered if the issue might be identified through Monte-Carlo simulations whereby random inputs, in this case bikes, were placed on the track and the coverage mapped for comparison. This type of simulation is possible through the coverage API with custom client scripts and can help identify where to site receive antennas around a given track. After much deliberation it was realised that the benefit of area modelling would be limited in contrast to focusing on the impact of a bend on a single bike which could be adjusted for different lean angles and simulated crowds. For this study the Path Profile (PtP) API was used to focus on a short 300m straight line path between a bike and a mast, with variation to the inputs. The bike’s height was adjusted to simulate lean based upon a starting height of 100cm (Superbike tail height average) down to a minimum of 50cm when at full tilt. The impact of adjacent riders was simulated by adjusting the receiver gain downwards, in this case by 9dB to simulate 3 other riders. The significance of receiver height was demonstrated by adjusting the mast to clear the fresnel zone at this distance.


The following data was generated using the ITM model with transmission heights ranging from 1m (Bike is upright) to 50cm (full lean). The ITM model considers the effect of the obstruction of the fresnel zone which is the cone of power around the path of a signal. Measurements are based upon a mast 300m away, on flat earth, with a 9dBi sectorial panel antenna. The zone grows deeper as it travels so mast height must consider this as well as line of sight clearance.

3m mast

A 3m mast is higher than most vehicles and ground clutter but only for line of sight. At 300m the fresnel zone is 3.12m so this mast height is only high enough up to about 250m before power is lost as the fresnel zone is attenuated by the earth. Results show that without obstructions 3m is borderline as bikes lean and as soon as a bike is obstructed by another it falls below the target threshold, regardless of lean angle.

3 metre mast, unobstructed

3 metre mast, obstructed

6m mast

A 6m mast is a major improvement. Being well clear of the fresnel zone makes it able to handle a full 60 degree lean at 300m. Results show that without obstructions 6m is good for all scenarios and if a bike is obstructed by others it only falls below the target threshold by 5dB which could be recovered with a higher gain antenna or by siting the antenna closer to the bend.

6 metre mast, unobstructed

6 metre mast, obstructed

The results reveal the following common findings:
  • Lean angle has a direct impact on signal quality with a full 60 degree lean adding more than 6dB of attenuation
  • Adjacent riders can introduce substantial attenuation with 3dB per rider
  • A bike at full tilt will lose another 1.5dB in antenna polarisation loss
  • Receiver height must be at least twice the maximum fresnel zone distance to budget for these issues
  • Receiver distance must be sufficient to maintain the double fresnel clearance so a distant mast is OK providing it is high enough
  • Low masts, like tight boots, are terrible!

Ideal mast height

The ideal mast height is relative to the frequency. At 2.3GHz the wavelength is 0.13m which based on the 300m distance used must be multipled by 24 to clear the fresnel zone making the minimum mast height 3.1m. As tests have shown, this height is insufficient to handle dynamic losses from leaning and other riders so should be doubled. Based on data, the recommended minimum height for a mast covering bikes on a bend is wavelength multiplied by distance/4 which gives the following table.
Distance mHeight m

Scripts and data

Scripts and data used to generate this study are available here. To use them you will need to enter your CloudRF API credentials into the CSV files and run them as follows: python3 pathprofile_3m.csv For plotting to PNG charts you will need Gnuplot: gnuplot unobstructed.gnuplot