GaN PA for 5G massive MIMO - thermal derating in practice?

GaN PAs promise high efficiency, but thermal derating under real massive MIMO traffic is another story. How much output power do you actually lose at temperature?

GaN PA thermal derating for massive MIMO usually fails in the field when you derate from datasheet Tcase alone and ignore the real stack: junction → package → cold plate / heat pipe → ambient + airflow.

What tends to hold up:
1. Measure (or tightly model) Rth_j-c and Rth_c-a under your actual duty / TDD pattern, not CW brochure numbers. Massive MIMO traffic is bursty; average dissipation can look fine while peak junction spikes still age the device.
2. Derate from junction temperature and lifetime (FIT / HTOL curve), then translate to a board-level Pout / PAE / ACLR budget. Case-temp-only derates miss package and TIM variance across a panel of PAs.
3. Watch memory / trap recovery after high-power bursts: a "thermal" ACLR fail is sometimes trapping + bias settle, not only temperature.

Useful post: your TDD duty, ambient, measured case vs estimated Tj, and whether ACLR or PAE hit the wall first. Share cooling (passive vs liquid / heat pipe) - that usually predicts how aggressive a derate you actually need.

— rfforum 编辑部 / Editorial Team

NVMe_Dev_Henry, in my experience the headline output-power loss is the wrong first number to chase. I measure output power, gain, PAE and ACLR together while stepping the case temperature under the real TDD pattern, not CW, because gain falls and the knee voltage rises with temperature, and the linearity limit usually arrives before a large power loss is visible. A PA that still delivers its rated power at temperature, but only with ACLR outside the limit, is not usable at that power.

I estimate junction temperature from dissipated power and the junction-to-case thermal resistance, using measured DC power minus RF output power, and I log case temperature with a thermocouple at the package as well as the heatsink. Each temperature step is held until the case is stable, which can take ten minutes or more under a bursty pattern, because a short CW soak underestimates the peak junction temperature during bursts. I then plot ACLR, PAE and saturated power against estimated junction temperature, and the derating curve comes from whichever limit is hit first.

Two practical points matter in an array. First, the middle devices run hotter than the edge devices, so a derating based on a single device on a cold plate is optimistic. Second, after a high-power burst I watch how quickly ACLR recovers; a slow recovery over seconds points to trapping or bias settling rather than pure heating, and no amount of extra heatsinking fixes that. I set the operating power from the junction-temperature and lifetime limit, then leave margin for thermal-interface and airflow variation across the panel.

We scoped the waveform on this once and could see the anomaly clearly. Worth a look.

@RF_Bench_Notes — logging Pout, gain, PAE and ACLR together under the real TDD pattern while stepping case temperature, and estimating Tj from dissipation rather than chasing headline power loss first, is the GaN thermal derating order we wanted next to the MIMO PA threads.

Editorial takeaway: freeze pattern, case soak, and ACLR-at-temperature on the same sheet; a CW ambient power number hides the linearity limit that actually gates the product.

— rfforum 编辑部 / Editorial Team