Chip selection guide

April 28, 2026 · View on GitHub

This guide maps mission profiles to the built-in ChipProfile constants. It is intended as a starting point for space_ml_sim.analysis.trade_study, not a flight-qualification recommendation.

Status. Parametric guidance. Numbers are first-order; for a flight mission run a full trade study using analysis.trade_study and the SPE Monte Carlo (environment.SPEStatisticalModel.sample_mission).

Quick decision tree

Need ≥10 TOPS for AI?
 ├─ Yes →  TID budget over 50 krad?
 │        ├─ Yes → VERSAL_AI_CORE  (130 TOPS, 100 krad, 7 nm, space-grade)
 │        └─ No  → JETSON_AGX_ORIN with shielding   (275 TOPS, 10 krad, COTS)
 │                 or TRILLIUM_V6E with heavy shielding (450 TOPS, 15 krad)
 │
 └─ No  →  Reconfigurable / FPGA?
          ├─ Yes → XQRKU060 (1.5 TOPS, 100 krad, most-flown space FPGA)
          │        or ZYNQ_ULTRASCALE (0.5 TOPS, 30 krad, Q8S OBC class)
          │
          └─ No  →  Need >100 krad?
                   ├─ Yes → SAMRH71 (Cortex-M7, 100 krad)
                   │        or GR740 (LEON4 quad, 300 krad)
                   │        or RAD5500 (1 Mrad, glacial)
                   └─ No  → NOEL_V_FT (open RISC-V, 50 krad)
                            or AURIX_TC4X (auto-grade, ⚠ not space-qualified)

By mission profile

LEO (500 km, 53° / SSO 650 km, 98°)

  • Background TID over 5 yr: 0.5–10 krad
  • SEU regime: sparse, manageable with selective TMR
  • Dominant risk: Single Event Latch-up in the SAA — pick chips with SEL immunity ≥40 MeV·cm²/mg

Recommended: JETSON_AGX_ORIN with 5 mm Al shielding, or VERSAL_AI_CORE for longer missions / heavier ML workloads.

Avoid: RAD5500 (overkill, sacrifices compute for no benefit at LEO).

Sun-synchronous EO (700–900 km, 98°)

  • Higher trapped proton dose than ISS-class
  • 5–50 krad over 5 years behind 2 mm Al
  • Recommended: VERSAL_AI_CORE for AI workloads, XQRKU060 for FPGA pipelines, GR740 for control-plane only.

MEO / GPS-class (20 200 km, 55°)

  • Outer-belt protons + electrons
  • 50–500 krad over 10 yr
  • Recommended: RAD5500 or GR740 for OBC; XQRKU060 only with active scrubbing (uses built-in FRAME_ECC).

GEO (35 786 km, 0°)

  • Electron-dominated environment
  • Recommended: VERSAL_AI_CORE, RAD5500, or XQRKU060. Do not fly bare COTS chips here.

Lunar transfer / cislunar

  • Outside the magnetosphere; GCR background is small but SPE risk is the dominant TID source.
  • Use HeliocentricEnvironment.lunar_transfer() for background and SPEStatisticalModel(solar_phase="max") for the worst-case event budget.
  • For a 6-month CubeSat mission at solar max with 2 mm Al, expect ~5–15 krad of SPE-driven TID at the 95th percentile.

Recommended: VERSAL_AI_CORE (100 krad budget covers SPE p95 comfortably). For lower-power / lower-compute payloads: SAMRH71, GR740, or NOEL_V_FT.

Don't fly bare: JETSON_AGX_ORIN (10 krad budget can be exhausted by a single major SPE).

Mars transit (1.0 → 1.5 AU, 7–9 months)

  • Higher GCR than Earth orbit, plus full SPE statistics.
  • 20–50 krad SPE-driven worst case at solar max.
  • Recommended: RAD5500 for the OBC, VERSAL_AI_CORE for AI payload, or XQRKU060 if FPGA-based.

Venus flyby (~0.72 AU, short)

  • GCR is suppressed by stronger solar modulation
  • SPE risk is still real (events arrive radially from the Sun)
  • For a 1–2 month flyby window: VERSAL_AI_CORE is the cheapest sufficient option; JETSON_AGX_ORIN works only with thick shielding (>5 mm Al) during solar minimum.

Outer planets (Jupiter, Saturn)

  • Out of scope for v0.5. The Jovian magnetosphere dominates and the built-in models do not include trapped-electron belts beyond Earth.

By compute requirement

If you need …Pick
≥200 TOPS for transformer inferenceTERAFAB_D3 (projected) or TRILLIUM_V6E
100–200 TOPS, mission-criticalVERSAL_AI_CORE (qualified for 15 yr)
200+ TOPS, COTS, tolerable riskJETSON_AGX_ORIN + 5 mm Al + active TMR
1–10 TOPS, FPGA pipelinesXQRKU060 or ZYNQ_ULTRASCALE
<1 TOPS, control-plane onlyGR740, SAMRH71, NOEL_V_FT
Reliability above all (1 Mrad budget)RAD5500
Lab / cost-down trade-study placeholderAURIX_TC4X (⚠ not space-qualified)

Working through a trade study

The chip table is data; the decision is a function. Use:

from space_ml_sim.analysis.trade_study import compare_chips
from space_ml_sim.models import ALL_CHIPS
from space_ml_sim.environment import HeliocentricEnvironment, SPEStatisticalModel

env_background = HeliocentricEnvironment.cruise_1au_solar_min()
spe = SPEStatisticalModel(solar_phase="max", shielding_mm_al=2.0)

study = compare_chips(
    chips=ALL_CHIPS,
    background_env=env_background,
    spe_model=spe,
    mission_days=210,
)
print(study.ranked_by_margin())

The trade study folds in TID margin, SEU rate at the chip's published cross-section, and compute headroom for the workload. Use it as a first filter, not the final answer.