B210/AD9361 - Hidden DC offset degrades reception on first tunes

[gibbi]

Issue Description

When initializing a B210, the DC offset is calibrated before the streamer is started. While additional DC compensation can seemingly reduce the offset, the physical offset at the ADC is persisting unless further tuning operations with more than a 200 MHz frequency change trigger an internal recalibration.

This effect can be severe at higher frequencies and gain settings. Offsets of up to 0.7 full scale have been observed. This significantly limits ADC headroom and dynamic range in cases where the initial tune persists for the lifetime of the application.

Setup Details

This issue can be reproduced with UHD versions from 4.9 down to 3.10 (likely all previous releases). The host OS appears to be irrelevant (Ubuntu 24.04 was used).

Only a B210 with terminated ports is required. A signal generator is used for further verification.

Steps to Reproduce the Problem

• Attach a signal generator and inject a CW signal at 5.801 GHz that is close to clipping on channel A.
• Record using rx_samples_to_file at 5.8 GHz with maximum gain:
  • with DC compensation enabled (File A)
  • without DC compensation (File B)

Workaround

• Modify rx_samples_to_file to tune from 5.8 GHz to 5.0 GHz and back after get_streamer() is called (this forces an internal recalibration).
• Record using mod_rx_samples_to_file at 5.8 GHz with maximum gain and DC compensation disabled (File C).

Expected Behavior

• File A should exhibit no DC spike after a few milliseconds.
• File B should exhibit a reasonably small DC offset (e.g., < 0.1 FS), and the CW signal should not clip.

Actual Behavior

• File A exhibits severe artifacts (see attached plot) when the initial DC offset exceeds 0.5 FS. This appears to be hardware-dependent, so on most B210 units File A may not appear suspicious.
• File B consistently exhibits a high DC offset on channel A (0.3–0.7 FS), and the CW signal is clipping.
• In File C, the forced recalibration consistently reduces the DC offset to a reasonable level (< 0.1 FS), and the CW signal is not clipping.

Additional Information

I can provide multiple B210 units with severe initial DC offset; all are brand new. The offset appears to decrease with each successive tune, but remains persistent if no retuning is performed.

[gibbi]

I think I understand the problem now.

The RF-DC compensation in the AD9361 is an analog DC compensation stage. It has a prescaler DAC_FS.

The tracking algorithm only updates when specific events occur (UG761: Table 58):

• gain changes
• very low signal levels
• exit from RX mode

It is possible to read back the applied values (BB and RF). While observing them, I had a hard time ever getting it to apply sensible values. Gain changes would update the value, but it always feels sluggish and “behind”. BB tracking always has to do the heavy lifting. Only extensive gain changes on the same frequency reliably got tracking to a point where BB tracking compensation was below 0.1 FS.

The UHD behavior is not helping the already poor tracking by changing DAC_FS from 2 → 3 when passing over 4 GHz during tuning. The worst case is a small jump (<100 MHz) across this gap, as it changes the compensation voltage dramatically without re-calibration. Tuning without adjusting RF-DC is essentially always a bad idea, as the adjustments from tracking in most cases never happen. Depending on the USRP this can leave the user with a significant analog DC in most operating scenarios.

The artifacts above are what happens when the BB compensation fights an offset greater than 1, causing it to flip.

The only reliable way I found is to force the analog compensation to precalibrated values per gain/frequency. I always use DAC_FS = 2, as its range was sufficient on every device tested. The values scale almost linearly, so only a few calibration points are needed.

Proposed fix

• Pre-calibrate a LUT using a UHD utility
• Always use DAC_FS = 2
• Disable chip-internal BB/RF calibrations
  • enables much faster initialization and frequency jumps
• Add an API for readback of tracking variables for verification

Note: This might cause problems for users of AGC, since RF-DC compensation may still be required there.