TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

Inside a 77 GHz Radar Chip: Front End, ADC, FFT and MIMO

Murugavel Ganesan
by
0
Share Facebook X LinkedIn WhatsApp Email
Automotive 77 GHz radar module with the radome removed, showing the radar chip and patch antenna arrays on the circuit board

A 77 GHz car radar can be built from separate parts, with a millimetre wave front end on one chip and the processor that does the heavy maths on another. It can also be one chip that holds the transmitters, receivers, ADCs, FFT engine, DSP, safety microcontroller and network interface. TI's AWR2944 puts four transmitters, four receivers, a C66x DSP and an Arm Cortex-R5F on 45 nm RFCMOS, and NXP's SAF85xx family does a similar job on 28 nm RFCMOS.

Integration changes where the design decisions sit, but it does not remove them. Bandwidth still sets range resolution, the ADC still caps maximum range, chirp timing still caps speed, and the antenna count still sets angle resolution. This post walks through each building block, then works one illustrative design from requirement to memory budget. In that example, a 300 MHz sweep gives about 0.5 m range resolution out to about 200 m, and the real lesson is what four-transmitter MIMO does to the speed limit. Every number in the example is an assumption chosen for teaching, not a product specification.

Key takeaways

  • Range resolution depends only on the swept bandwidth that is actually sampled: c/2B, so 300 MHz gives about 0.5 m and 4 GHz gives 3.75 cm.
  • The ADC sampling rate and chirp slope together set maximum range. A faster ADC buys range or a steeper, shorter chirp.
  • MIMO multiplies virtual antennas (transmitters times receivers) and so improves angle resolution, but time-division MIMO divides the unambiguous velocity by the number of transmitters.
  • On-chip memory is a first-class design limit, because every range FFT of a frame has to be stored before the Doppler FFT can start.

The signal chain in one paragraph each

The chips discussed here use frequency modulated continuous wave (FMCW) operation. TI's white paper "The fundamentals of millimeter wave radar sensors" by Cesar Iovescu and Sandeep Rao (revised July 2020) is a clear primary reference for the maths that follows, and the post uses its notation.

Synthesiser and transmitters

A synthesiser generates a chirp, a tone whose frequency rises linearly with time. The chirp has a start frequency, a bandwidth B, a duration and a slope S. The transmit chains amplify it and drive the antennas. Its phase noise limits how well weak targets near strong ones can be seen. TI lists phase noise for the AWR2944 at 1 MHz offset as -96 dBc/Hz in the 76 to 77 GHz range.

Receivers and mixers

Each receive chain has a low noise amplifier and a mixer that multiplies the received echo by the transmitted chirp. The echo is a delayed copy of the chirp, so the mixer output is a constant tone, called the IF signal, whose frequency is 2Sd/c for a target at distance d. Range becomes frequency. That is why an FFT can measure distance.

IF filter and ADC

The IF signal is filtered and digitised. Because the IF frequency grows with distance, the ADC rate caps the range the radar can see. Since the IF tone is 2Sd/c, an ADC that captures IF frequencies up to Fs limits range to d_max = Fs c / 2S. Each receive channel needs its own filter and ADC, which is why the TI application report "MIMO Radar" (SWRA554A, revised July 2018) notes that adding real receivers has limits: each one is a full chain with an LNA, mixer, IF filter and ADC.

Range FFT, Doppler FFT and angle FFT

A first FFT across the samples of one chirp turns IF tones into range bins. A frame of many chirps is then processed with a second FFT across chirps, which reads the small phase change each target causes from chirp to chirp and turns it into velocity. A third FFT across antennas gives angle. A detection step picks out real targets from this cube of data, and the microcontroller then clusters and tracks them.

Processors, memory and interfaces

The FFTs run on a hardware accelerator, detection and angle work on a DSP, and tracking and the vehicle interface on a safety microcontroller. On the AWR2944, TI lists a hardware accelerator (HWA 2.1) for FFT, log magnitude and memory compression, a C66x DSP at 360 MHz, a lock-step capable Cortex-R5F at 300 MHz, 4 MB of on-chip RAM, CAN-FD and Ethernet. Hardware integrity is listed as "up to ASIL B targeted". For readers who want the wider picture of how vehicle controllers reach those safety targets, our post on automotive microcontrollers and ISO 26262 covers lock-step cores and the hardware metrics behind ASIL ratings.

From a board of parts to one die

The word "discrete" in radar usually means the front end and the processor are separate chips. A front end transceiver produces raw ADC data and ships it over a fast link to a processor that does the FFTs. The integrated approach keeps the whole chain on one die and sends out only a point cloud or object list.

TI's two generations show the direction of travel. The AWR1642, which TI listed as active as of 10 November 2026, integrates two transmitters, four receivers, a 600 MHz C674x DSP, a 200 MHz Cortex-R4F, 1.5 MB of memory and a 12.5 Msps ADC on 45 nm RFCMOS. The AWR2944 doubles the transmitters to four, raises the ADC ceiling to 37.5 Msps, and adds a newer accelerator, more memory and Ethernet. NXP's product page describes the SAF85xx as a 77 GHz RFCMOS radar one-chip SoC, and its product video names the process as 28 nm. As of 10 November 2026 the page lists four transmitters and four receivers, Arm Cortex-A53 and Cortex-M7 cores, a BBE32 vector DSP with an SPT radar accelerator, and a Gigabit Ethernet port.

Device (vendor figures)ProcessTX / RXMain processing on dieMax ADC rate
TI AWR164245 nm RFCMOS2 / 4C674x DSP 600 MHz, Cortex-R4F 200 MHz, accelerator12.5 Msps
TI AWR294445 nm RFCMOS4 / 4C66x DSP 360 MHz, Cortex-R5F 300 MHz, HWA 2.137.5 Msps
NXP SAF85xx28 nm RFCMOS4 / 4Cortex-A53 and Cortex-M7 cores, BBE32 vector DSP, SPT acceleratorNot stated on the product page

These rows come from sources with different levels of detail, so they show what each part contains, not how the parts compare in a car.

Integration saves board area and parts, and the raw ADC stream never leaves the die. In the example below, four receivers at 20 Msps with 16-bit samples produce 1.28 Gbit/s while sampling. Keeping that on chip avoids a high-speed link. Where raw data does have to leave the sensor, for example to a central computer, interface choice becomes a real design topic, which our post on carrying radar and LiDAR data over MIPI CSI-2 explores.

A worked example: from requirement to memory

Every input in this section is an illustrative assumption, not a specification of any product. The formulas come from the TI white paper and application report cited below, and every result was computed with a short script using c = 299,792,458 m/s and a 77 GHz carrier, so the wavelength is about 3.89 mm. Values are rounded.

AssumptionValue
Chirp slope S15 MHz/µs
ADC sampling window per chirp20 µs
ADC rate20 Msps (400 samples per chirp)
Ramp time, including settling before sampling24 µs
Chirp period, including idle time30 µs
Antennas4 TX, 4 RX, time-division MIMO, half-wavelength spacing
Chirps per transmitter per frame128

Step 1: range resolution

The bandwidth that counts is the part of the sweep that is sampled: 15 MHz/µs times 20 µs is 300 MHz. Range resolution is c/2B, about 0.50 m. The full ramp occupies 360 MHz, which fits inside the 76 to 77 GHz band with room to spare. Settling time uses spectrum without buying resolution.

Step 2: maximum range

d_max = Fs c / 2S = 20 Msps times c over 30 MHz/µs, about 200 m. At that distance the IF tone is 20 MHz, at the edge of what the ADC captures. If the same chirp ran on a part limited to 12.5 Msps, maximum range would fall to about 125 m. To double resolution to about 0.25 m with the same sampling window, the slope has to double, which doubles the IF frequency at 200 m to 40 MHz and doubles the ADC rate needed. Resolution and range pull against the ADC from opposite ends.

One caution: datasheets quote ADC rates for real and complex sampling differently, and the usable IF bandwidth depends on which mode you use. Read the definition in the datasheet before plugging its number into this formula.

Step 3: maximum velocity, and the MIMO penalty

The unambiguous velocity is λ/4Tc, where Tc is the time between chirps seen by the same transmit and receive pair. With one transmitter and a 30 µs period, that is about 32.4 m/s, or 117 km/h of closing speed. With four transmitters taking turns, each transmitter repeats only every 120 µs, so the limit falls by four, to about 8.1 m/s or 29 km/h.

This is the hidden cost of time-division MIMO. A front radar on a motorway sees closing speeds far above 29 km/h, so a design like this would rely on velocity disambiguation in software, a shorter chirp period, or a different way of separating the transmitters. TI's report describes binary phase modulation (BPM) MIMO, in which all transmitters are active in every slot of a block and are separated with a code. It notes an SNR benefit of 10log10(NTX), because more power is radiated per slot. Choosing the multiplexing scheme is one of the most important calls in the whole design.

Step 4: velocity resolution and frame time

Velocity resolution is λ/2Tf, where Tf is the frame duration. Here 128 chirps per transmitter times four transmitters times 30 µs gives a 15.36 ms frame and a velocity resolution of about 0.13 m/s. Longer frames sharpen velocity but slow the update rate.

Step 5: angle resolution

With N equally spaced antennas at half-wavelength pitch, angle resolution near boresight is about 2/N radians. Four real receivers alone give 0.5 rad, about 29 degrees, consistent with the "about 30º" TI gives for four receivers. The 4 by 4 MIMO virtual array of 16 elements gives 0.125 rad, about 7.2 degrees. Resolution worsens away from boresight, by a factor of 1/cos θ.

Step 6: the memory budget

One consequence drives chip architecture: because the Doppler FFT runs across the chirps of a frame, range FFT data for every chirp in the frame must be stored before the Doppler FFT can start. Keeping 256 range bins as 16-bit complex values (4 bytes each) for 512 chirps across 4 receivers needs 256 times 512 times 4 times 4 bytes, which is 2 MiB. That is half of the AWR2944's listed 4 MB and more than the AWR1642's 1.5 MB. On-chip RAM, and the compression feature TI lists in its accelerator, are not afterthoughts. They decide which frame shapes a chip can run.

Result (illustrative)ValueWhat sets it
Range resolutionabout 0.50 mSampled bandwidth
Maximum rangeabout 200 mADC rate and slope
Max velocity, 1 TX / 4 TX TDMabout 32.4 / 8.1 m/sTime between chirps from the same TX
Velocity resolutionabout 0.13 m/sFrame time
Angle resolution, 4 RX / 16 virtualabout 29 / 7.2 degreesVirtual array length
Radar cube memory2 MiBBins, chirps, channels, word size

When one chip is not enough

Imaging radars chase angle resolution of a degree or two, which needs far more virtual channels than a single 4 by 4 chip can offer. The usual route is to cascade several front ends from one shared clock. TI's public reference design "Imaging Radar Using Cascaded mmWave Sensor" (TIDEP-01012, design guide TIDUEN5A, revised March 2020) cascades four AWR2243 devices, one master and three slaves, sharing a 20 GHz local oscillator. Together they give 12 transmitters and 16 receivers. That is 192 transmit and receive pairs, but the guide gives the azimuth virtual array as 86 elements, because some virtual positions overlap and the master's three transmitters are placed vertically for elevation.

That gap between 192 and 86 is a useful lesson. Virtual channels multiply, but unique, useful positions depend on the antenna layout. With 86 elements at half-wavelength pitch, the 2/N rule gives about 1.3 degrees near boresight. This is a textbook estimate; TI's design guide quotes about 1.4 degrees for the board.

A cascade brings back what integration removed: clock and phase alignment between chips, calibration across devices, and a fast link to a separate processor.

Design choices that matter most

  1. Band and bandwidth. CEPT's ECC/DEC/(04)03, approved on 19 March 2004 and corrected on 6 March 2015, designates 77 to 81 GHz for automotive short range radar, and the register listed it as active as of 10 November 2026. Wide sweeps for fine resolution live there, while long range modes typically use narrower sweeps.
  2. Chirp slope versus ADC rate. For a fixed chirp time and ADC rate, short range modes can use a steep slope and wide bandwidth, while long range modes need a shallow slope and so a narrower bandwidth, as the d_max formula above implies.
  3. MIMO scheme. Time division is simplest but costs velocity range. Code-based schemes recover transmit power and timing at the price of decoding work and sensitivity to phase errors.
  4. Where processing lives. An object list keeps the vehicle network light. Raw or lightly processed data lets a central computer fuse radar with other sensors, which our post on cameras and sensor fusion in ADAS discusses from the camera side.
  5. Safety target. A part targeting ASIL B may need system measures to support a higher-rated function.

Our SoC lifecycle on The Semiconductor Hub shows where these architecture and safety decisions fall in a chip programme.

Sources

The worked example uses illustrative numbers, and its results are textbook estimates rather than measured performance. Vendor figures are the companies' own specifications and claims, as published on the dates shown.

Murugavel Ganesan
About the Author
Murugavel Ganesan
Product Director | SerDes · IVN · Edge AI · Compute | Ex Intel · Infineon

A semiconductor product leader working across product architecture, definition, strategy, and silicon and software execution, from concept through volume production. He has worked with teams across Asia, Europe and the USA, including at Intel, Infineon and ArrayComm.

His work covers ASICs, SoCs, FPGAs and chiplets for wireless, wireline, automotive, IoT and AI, including PCIe, CXL and DDR interfaces alongside RF and mixed-signal design.

Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!