Intel has introduced Starfire, a space-grade system-on-chip for U.S. government use that brings its 18A process into a place where bit flips are more than an annoyance. According to Intel’s Starfire product sheet, the chip combines CPU cores, an AI accelerator, and graphics in one Foveros package, with samples planned for the third quarter of 2026.
The part comes in two versions. Intel lists a 10 W low-power model rated for up to 45 TOPS and a 35 W performance model rated for up to 75 TOPS. Both are specified to operate from minus 55 degrees Celsius to 125 degrees Celsius and to last more than 10 years.
Both versions use the same basic floor plan: four performance cores made on Intel 18A, four low-power efficiency cores, a three-tile neural processing unit also made on 18A, and a four-core Xe GPU with 64 execution units made on Intel 3. The lower-power Starfire runs its performance cores at 1.0 GHz, its efficiency cores at 850 MHz, and its GPU between 800 MHz and 1.0 GHz. The performance version pushes the performance cores to 3.1 GHz, efficiency cores to 2.1 GHz, and the GPU to 2.0 GHz.
Intel says both models include 12 lanes of PCIe Gen4 and support either LPDDR5 or DDR5 memory. That puts Starfire closer to a compact AI-capable computer than the old radiation-hardened controllers that have handled many spacecraft jobs for years.
Why the node choice matters
Starfire’s split between Intel 18A for compute and Intel 3 for graphics echoes Intel’s approach with Clearwater Forest, the 288-core Xeon design that combines 18A compute tiles with Intel 3 base tiles. The choice is technically aggressive for space. Smaller transistors store less charge per bit, which can make advanced chips more vulnerable to radiation-induced upsets. Intel is betting on its 18A design, including RibbonFET transistor technology and hardening work, rather than staying on older, more forgiving lithography.
That is the part of the pitch that still needs proof. Intel’s documentation lists radiation data for total ionizing dose, single-event latch-up, and single-event effects as characterization still in progress. In plain English: the company has announced the chip, but the document does not say Starfire is radiation-qualified yet. Intel also notes that specifications may change.
Aimed above legacy spacecraft processors
The comparison point for this market remains BAE Systems’ RAD750, a radiation-hardened PowerPC processor that has flown for roughly two decades. Public specifications put that chip at 110 MHz to 200 MHz, with 10.4 million transistors on 150 nm or 250 nm manufacturing. It has flown on missions including the Mars rovers, Kepler, and Fermi, among more than 150 spacecraft.
Newer space processors have started to move beyond that baseline. BAE’s multicore RAD5545 and a NASA-backed Microchip processor project aim to increase spaceflight computing throughput, with NASA’s program targeting 100 times the performance of current chips. Starfire’s advertised 75 TOPS and dedicated neural processor aim at a different workload: running AI inference in orbit, closer to where data is collected, rather than shipping every decision back through a narrow communications link.
Intel Government Technologies is handling Starfire. Intel is pitching domestic manufacturing and pricing it describes as market-competitive. The company also points to Intel Foundry’s U.S.-based leading-edge logic manufacturing, Trusted Foundry status, and its ties to Pentagon programs including RAMP-C and SHIP. The manufacturing story carries its own caveat: Intel’s 18A yields have been reported as unlikely to reach industry-standard levels until 2027.
This story draws on original reporting from Tom's Hardware.
