Qualcomm Puts TSN Ethernet and a +115 °C Rating on an Edge AI Processor
The Dragonwing IQ-2390 opens a new industrial series with time-sensitive networking on the die and a -30 to +115 °C range. Evaluation kits are a year out.
Qualcomm introduced two Dragonwing processors on 1 September: the Q-2390 for commercial and consumer IoT, and the IQ-2390, the first part in a new IQ2 series aimed squarely at industrial edge AI. Both pair a quad-core Kryo CPU with an Adreno 704 GPU and a real-time RISC-V microcontroller, and both run Linux, Android or Zephyr.
The industrial part is distinguished by two things that are hard to retrofit. The first is dual Gigabit Ethernet carrying time-sensitive networking on the die itself, rather than through an external switch or bridge. The second is an operating range of -30 °C to +115 °C, along with design provisions for vibration and shock.
On-die TSN is the more interesting of the two. A machine builder who wants cyclic control traffic and camera traffic on the same wire currently reaches for a separate TSN switch or a companion controller, and the integration cost of that decision usually exceeds the component cost. Putting the capability behind the same memory controller as the vision pipeline removes a board, a supplier and a class of timing problem.
Qualcomm names industrial automation, oil and gas, utilities and energy systems, smart HMIs, PLCs, gateways, industrial vision and building automation as targets. Evaluation kits for both parts are expected in early 2027, with an early-access customer programme already running and demonstrations at IFA 2026.
A specification table published the same day by CNX Software fills in what the announcement leaves out: one Cortex-A78 alongside three Cortex-A55 cores at 1.5 to 1.9 GHz, an optional SiFive E61 RISC-V core at 600 MHz, a Hexagon NPU rated at 1.1 TOPS, dual-channel LPDDR4x, 1080p30 video encode and decode, optional LTE Cat 4, and product availability committed until 2036.
That 1.1 TOPS figure is the honest measure of the ambition here. This is not an edge server; it is inference placed inside cost-sensitive control hardware — reading a gauge, classifying a defect, watching a bearing — where the deciding specifications are the temperature range and the ten-year supply commitment rather than the benchmark.
Source: Qualcomm