一位雷达工程师对新款特斯拉HW4雷达的分析

A radar engineer's analysis of the new Tesla HW4 radar.

Not long after Tesla made the controversial decision to remove radar from its cars in 2021, rumors started circulating about an in-house radar program (further validated byFCC filings in 2022). This program has always been shrouded in mystery, especially given the very public statements Elon Musk was making in parallel about focusing exclusively on Tesla Vision: “Our AI-based software architecture has been increasingly reliant on cameras, to the point where radar is becoming unnecessary earlier than expected. As a result, our FSD team is fully focused on evolving to a vision-based autonomous system.” (Tesla Q1 2021 Earnings Call).

Fast forward to 2023, and as details of the new HW4 architecture started to emerge, more and more signs pointed to a return of radar to the Tesla architecture. Now that HW4-equipped vehicles have started rolling-out, it wasn’t long beforedetailed photos of the in-house developed Tesla Radar emergedon Twitter, thanks to @greentheonly.

As a radar geek, nothing is more fun than analyzing a new product in the space, and seeing how other teams have approached the challenge of designing an in-house radar purpose-built for enabling L4 autonomy, as our team does here at Ghost. What follows is an unassuming act of educated speculation – with no more data available than what one can see by looking at pictures of the device. As such there are details that are obvious, but many more that are simply educated guesses – and bound to be off in places.

So to get to the big question up-front, ‘why did they bother?’ While the device we see here is clearly more capable than the Continental radar Tesla shipped years ago (4D, and likely provides much more software control, time synchronization, and optimization for the Tesla system), it also appears to fall very short of the specs of the new generation of long-range 4D imaging radars that have been introduced by the automotive industry. It also lacks the compute power or bandwidth and central processing architecture to lean-in to novel neural network-based radar data processing approaches that are on the cutting edge of software-defined radar (like we’re building here at Ghost). The main advantage here isn’t hardware, specs, or raw capabilities, it seems to be control – enabling deep optimization and synchronization of the radar with the overall autonomy system –an approach we very much believe in at Ghost!

With that, let’s dive into the details…

Sensitivity

The radar has 6 TX and 8 RX channels, giving a maximum virtual channel count of 48. I suspect signal processing limitations will practically limit them to ~20 virtual channels at a time (more on this later). For reference, previous gen radars typically had 8 virtual channels and the forthcoming Mobileye radarhas over 2,000(though it is much debated how much of this they can support in signal processing).

Table 1: Observed Radar Specifications

The radar chips have RF traces consistent with TI’s AWR2243, a chip that’s been commercially available for several years. Steady and proven, it uses FMCW waveforms. While FMCW has always made sense as the low-cost entry point for automotive radars, I thought we might see a newer waveform like phase coding or full arbitrary in this radar. Those would offer relief from the sensitivity and unambiguous velocity constraints of time-division MIMO as well as better interference mitigation. The lack of a commercially available radar chip to support those waveforms likely drove Tesla to stick with FMCW. Although the 2243 supports 76-81 GHz, the radar bandwidth will be limited by the antenna to about 2 GHz. Their current FCC license limits operation to 76-77 GHz, so it’s unlikely to see this expanded using this design. Speaking of the antennas, they are end-fed printed patches; although the patches have a unique element structure, this lossy, narrow-banded antenna technology is common in prior-generation automotive radars. I’d expect about 13 dBi of gain out of each element. The color of the RF material suggests Isola, a very popular choice in modern automotive radars. With this material and the estimated length of the traces that would be 5 dB TX and 3 dB RX. So, what does this all mean for sensitivity?

Equation 1. Radar Range Equation‍

Applying the numbers above to this equation gives the plot in Figure 1. I would prefer to include the effects of rain and water droplet accumulation in this plot. I didn’t attempt to because it is dependent on details of the design I don’t have. Better to plot what I can defend & let you speculate.

Figure 1. Detection of a motorcycle in rain, a key requirement for some automotive radars, will likely be < 100 m for the Tesla radar.

MIMO Topology

Antenna topology is always a challenge trade between performance, cost, and signal processing approach. There are many possible combinations that Tesla could be using, but I suspect the two combinations in Figure 2 are in the mode list.

Figure 2. Candidate antenna topologies for a 1) fine azimuth resolution (blue) and 2) az/el modes (orange)

A design detail to point out is that the second and third (from the left) RX elements have different element patterns and spacing than the others. This clever design allows Tesla flexibility with their 8 RX elements; one set of choices they could make is described in Figure 3 and Table 1.

Figure 3. Element 2 & 3 have wider element patterns and narrower spacing

Table 2. RX array configurations

A key note about the analyses: they are based on the extracted geometry in Figure 9. If those dimensions are off, predictions about element spacings and beamwidths would be off.

Now, back to topologies. The first (blue) is a mode designed to minimize azimuth measurement error by increasing azimuth extent. No elevation measurement is available in this mode. The spacing of the virtual elements (~1.2-wavelengths) will introduce grating lobes. This could lead to ambiguous azimuth measurements; when the measurement errors are this large, some self-driving systems would count this as a false-alarm or missed detection. To eliminate this problem in HW would require more elements to fill in the spacing (adding cost & surface area) or shrinking the spacing with the current number of elements (thus increasing azimuth measurement error). I empathize with the engineers making this trade-off! Given the topology that they chose, sidelobe mitigation must be a primary concern for their signal processing team. There are ways to mitigate the effects. Whatever they do, it’s probably very clever, because 1.2-wavelength spacing and level-4 radar performance don’t typically go together.

Figure 4. Squares are the virtual arrays.

Another concern that the Tesla engineers must solve with this mode is elevation-dependent changes to the signal. The radar is mounted behind a facia (Figure 6) which isn’t flat and the elements are near the edge of the board/radome, meaning that signals arriving from positive elevation angles have a different refraction path than signals arriving from negative elevation angles. With no ability to measure elevation in this mode, and somewhat limited signal processing resources, those differences could directly manifest as angle error. This mode might have good accuracy at a particular elevation angle (likely el = 0), but the angle performance could fall off at other angles. This would be important if trying to measure the elevation of bridges or sign boards.

The second (orange) is a mode designed to provide both azimuth and elevation measurements of target location. When combined with range and Doppler, this makes a “4D radar”. In a classic az/el radar mode, elements would be positioned in a rectangular grid with extent based on desired angle accuracy and spacing of 0.5-wavelengths to eliminate grating lobes (Figure 5). To match the extents that Tesla chose, that would take 34 elements, which isn’t acceptable to cost. Instead, they can make a virtual array that covers the full extent along the diagonal only. The benefit is that it achieves the same angular resolution as the full array (which always looks great on a spec sheet), but it comes at the cost of severe sidelobes that create false-alarms and miss detections (this is harder to quantify, so it never shows up on a spec sheet). I built a simple simulation of the hypothetical Tesla virtual arrays (Figure 6) to show what the principle antenna cuts might be.

Figure 5. Both the array you want (left) and the array you can afford (right) have the same azimuth and elevation extent but very different sidelobes.

Figure 6. Grating lobes are a particular type of sidelobe. All sidelobes have the effect of creating ambiguity: the radar sees a signal, but it can’t tell if it is a strong signal coming through a sidelobe or a weak signal coming through the main lobe. This can result in “phantom” targets appearing in front of the car.

Table 3. Summary of performance of potential MIMO modes

Radome

Figure 7. The mounting location behind the facia presents calibration challenges but alleviates some environmental & aesthetic challenges.

Another notable feature of Tesla’s system is that the radome does not appear to be exposed directly to the external environment; it is behind a facia. This has the benefit of reducing some of the environmental loads applied to the radome. It’s not clear from these pictures how they will handle water/snow/ice accumulation on the facia. Detection range and angle accuracy fall off quickly in those conditions. This mounting location also has implications on calibration: I don’t think the radar can be calibrated until installed on the car since the facia will affect calibration. I don’t know if they put the car on a turn-table or have a wrap-around antenna chamber; either way, it sounds like a cool (possibly expensive) measurement facility! If instead they assumed that the facia on every car will be the same and have the same geometry, their angle accuracy would be poor.

Signal processing

Tesla appears to have chosen to leave the radar signal processing on the radar head. This contrasts with SW-defined radars that can scale the size of the compute without changing the design of the radar head. It also means Tesla must find compute components that will survive the environment the radar needs to be in; for previous generation radars, this excluded the use of FPGAs. Tesla broke through this barrier using Xilinx’s recently released automotive grade Zynq (Figure 8 ) with 1 Gb of external DDR3 memory from Micron. It’s impossible to know exactly what signal processing concepts Tesla is applying here, so I can only make “professional guesses” based on textbook radar signal processing and my personal experiences. Based on the antenna topology, I predict Tesla needs to process 19 virtual elements when measuring azimuth only and 16 virtual elements when measuring azimuth and elevation. The logic cells (85k), LUTS (53k) and BRAM (4.9 MB) of the processing resources spread over this many channels would support basic range, Doppler, beamforming, detection, and search-based angle measurement algorithms. It’s unlikely they could support adaptive processing for interference mitigation, advanced calibration & measurement techniques, nor radar machine learning. They can likely support fine resolution (~1 foot) modes with limited range swaths and medium resolution modes (~1 meter) with large range swaths.

Figure 8. Placing computational resources on the radar head reduces the bandwidth back to the self-driving compute system, but it makes it harder to scale radar compute

The connector only has six pins. Assuming redundant power and ground, that only leaves two for interface. Automotive ethernet is the most likely communication protocol. The Marvell Q1010 family chip is shown in another Twitter photo, and that chip supports 100BASE-T1, or 100 Mbps.

Summary

When Tesla removed last-generation automotive grade radar off its cars and quietly setup an in-house development program, the radar community suspected it would be a short matter of time before a new radar appeared. Now that we have this first look at the Tesla Radar, it’s clear that Tesla didn’t prioritize raw hardware performance with their design; instead they wanted a product they could get in cars quickly that they could control. By owning the software that runs the signal processing algorithms, Tesla can take total responsibility for radar performance, optimize it for coordination with their vision-centric system, and make modifications to suit their system on an ongoing basis. The hardware selections that I see here reflect a design that prioritized time-to-market and cost over performance. With an active team now designing in-house radars, one has to imagine that this is just the beginning, and we’ll see progression of this program over time as Tesla evolves their entire system towards level 4 autonomy. At Ghost, we agree with several of the core strategies that Tesla is employing (a focus on software and optimization of radar hand-in-hand with the overall autonomy system), but we’ve gone far deeper down the software-defined approach, leveraging a central compute architecture that gives much more flexibility for radar data processing…you canget a first look at the Ghost Radar program here.

Appendix: geometry

The math presented above is based on the geometric extraction below, which is all scaled to the package size of the AWR2243 (10.4 mm)

Figure 9. Extracted geometry is based on the published dimensions of the hypothesized radar chip (AWR2243)