Introduction
Cosmic rays are high-energy particles from space that interact with Earth’s atmosphere and generate secondary particles, especially neutrons. When one of these particles passes through a semiconductor memory cell, it can deposit enough charge to change a stored 0 to a 1, or a 1 to a 0. Engineers call this a single event upset (SEU). It is a type of soft error, which means the hardware is not permanently damaged, but the data in that cell may be wrong until the value is refreshed, corrected, or overwritten.
This calculator estimates the probability that at least one bit flip occurs in a memory region over a chosen operating interval. It is designed for intuition, planning, and quick scenario comparison. In other words, it helps answer practical questions such as: “How much more exposed is a large memory array than a small one?” or “How much does the risk rise when the same system runs at higher altitude?”
Because the tool uses a simplified Poisson reliability model, it should not be treated as a certification-grade radiation analysis. Still, the simple model is useful because it captures the first-order idea that rare per-bit events can become noticeable when you multiply them by a very large number of bits and a long runtime.
How to use the calculator
Start with the memory region you want to evaluate, then add the exposure time and a sea-level soft error rate. The altitude field applies an approximate multiplier to represent the fact that higher elevations experience less atmospheric shielding. The result is the probability that the entire memory region sees one or more raw upsets during the selected time.
-
Enter Memory Size (MB) for the memory region you care about. This tool treats “MB” as mebibytes (MiB) in the computation, so 1 MB here means 1,048,576 bytes.
-
Enter Operation Time (hours) for how long the memory is exposed.
-
Enter Altitude Above Sea Level (meters). Higher altitude generally means less atmospheric shielding and a higher particle flux.
-
Enter the Soft Error Rate per bit per hour at sea level. If you have a vendor SER or FIT figure, convert it to a per-bit-hour rate before using it.
-
Select Compute Probability to see the probability of at least one bit flip during the interval.
A good way to learn from the output is to change just one variable at a time. Keep memory size and SER fixed, for example, then compare sea level with a mountain site or a high-altitude airborne environment. That makes the role of each term much easier to see.
Formula and assumptions
The model assumes bit flips occur independently with a constant average rate. Under those assumptions, the number of events in a time window follows a Poisson distribution. The calculator first converts the memory size to bits, then scales the sea-level error rate by an altitude factor, and finally converts the expected number of flips into a probability of one or more flips.
In plain language, N is the number of bits at risk, t is the runtime in hours, h is altitude in meters, and SER is the sea-level soft error rate per bit-hour. The symbol λ is the expected number of raw bit flips in the interval. Once you know λ, the probability of at least one flip is just one minus the chance of zero flips.
- Memory conversion: N = memoryMB × 8 × 1024 × 1024.
- Altitude model: altitudeFactor = e^(h / 6000), where h is altitude in meters.
- Expected flips: λ = SER × N × t × altitudeFactor.
- Result: P(at least one) = 1 − e^(−λ).
When λ is much smaller than 1, the probability is approximately equal to λ. That is why tiny per-bit rates can still matter once a system contains billions of bits or runs for a long time. When λ grows well above 1, the probability of at least one event rapidly approaches 100%.
Worked example
Suppose you have 512 MB of memory operating for 24 hours at 0 m altitude, and you assume a sea-level soft error rate of 1e-12 per bit-hour. The steps are straightforward.
- Bits: N = 512 × 8 × 1024 × 1024 = 4,294,967,296 bits
- Altitude factor at 0 m: e^(0/6000) = 1
- Expected flips: λ = 1e-12 × 4.294967296e9 × 24 ≈ 0.103
- Probability: P = 1 − e^(−0.103) ≈ 0.098 → about 9.8%
The important interpretation is that this is a probability for the whole memory region over the whole day. It is not a statement that any particular bit is likely to fail. It is also not the same as saying the system itself has a 9.8% chance of visible failure, because many systems have mitigation layers between a raw upset and a user-visible error.
Limitations and practical notes
This calculator is intentionally simple. That simplicity makes it excellent for comparison and intuition, but it also means the output should be read with engineering judgment. Several real-world effects are compressed into a few parameters.
- Altitude model is approximate. Real particle flux depends on geomagnetic latitude, shielding, solar conditions, aircraft routing, building materials, and local environment. The exponential factor here is a convenient teaching model, not a full transport simulation.
- SER is highly device-dependent. Different memory technologies, process nodes, voltages, temperatures, packaging, and layouts can have very different upset sensitivity.
- Events are treated as independent. The Poisson model does not represent correlated multi-bit upsets, temporary bursts, or detailed spatial effects inside a memory array.
- At least one flip is not the same as system failure. ECC, parity, retries, scrubbing, checksums, redundant computation, and application-level safeguards can all reduce the impact of a raw upset.
- Memory input should reflect the bytes actually at risk. If only a portion of the memory is powered, active, or relevant to your workload, use that smaller amount.
If you need engineering-grade predictions for avionics, medical devices, scientific instruments, or space systems, use component-specific SER data together with an environment model matched to the actual location and shielding. This page is best understood as an educational estimator and a fast what-if calculator.
Understanding the calculation in real terms
Cosmic rays constantly bombard Earth. Most are deflected by the magnetic field or absorbed by the atmosphere, yet a fraction of secondary particles still reach the surface. For one bit, the odds of a strike that causes an upset are tiny. For a memory subsystem with billions of cells, the picture changes. The probability that something somewhere flips during a long operating interval becomes much more relevant than the risk to one individual bit.
That is why the calculator focuses on the probability of at least one event. In planning conversations, that is usually the first threshold people care about. If nothing happens, great. If one or more events happen, then the next question becomes whether mitigation catches them before they matter. The Poisson model is useful because it gives a clean bridge between a small average event rate and that practical threshold probability.
The soft error rate input is the most important parameter and often the hardest one to choose well. Vendors may publish SER in FIT, where 1 FIT equals 1 failure per 109 device-hours. To turn a FIT figure into the per-bit-hour value needed here, you must know whether the published figure is per device, per megabit, or under some specific test condition. If you are uncertain, treat the calculator as a sensitivity tool and try several plausible SER values instead of one “precise” number.
Altitude matters because the atmosphere acts as shielding. At higher elevations there is less air mass above the device, so more energetic secondaries can reach it. That is why mountain sites and aircraft altitudes often show elevated neutron flux relative to sea level. The exponential model used here deliberately keeps that relationship simple. It is not exact, but it captures the direction and rough scaling that make altitude worth considering in design tradeoffs.
Approximate altitude multipliers in this model
The table below shows the multiplier applied to the sea-level SER using the simplified expression e^(h/6000). These are illustrative values that can help you sanity-check whether the output trend matches your expectations.
| Altitude (m) |
Rate Multiplier |
| 0 |
1.00 |
| 1500 |
1.28 |
| 3000 |
1.65 |
| 6000 |
2.72 |
| 12000 |
7.39 |
Why the raw probability is only the first layer
In practice, designers rarely stop at the raw upset probability. Many systems use error-correcting code (ECC), which adds redundancy so single-bit errors can be corrected automatically and multi-bit errors can often be detected. Memory scrubbing periodically reads and corrects memory before multiple latent upsets accumulate. Higher-level software can add checksums, retry logic, replicated computation, or voting between redundant results.
These mitigation layers change the probability that a raw bit flip becomes an incorrect answer, a crash, or a visible fault. That is why this calculator’s output is best interpreted as an exposure metric. It helps estimate how often the environment and system size create opportunities for trouble. Then you decide whether your ECC scheme, scrub interval, or redundancy plan is enough for the consequence of failure in your application.
If you need more statistical detail, the same Poisson model also gives the probability of exactly k flips in the interval. But for many users, the “one or more” result is the most immediately useful because it aligns with mission planning and scenario comparison. It tells you when rare events stop feeling abstract and start becoming something the architecture should consciously handle.
How to interpret the result
A small percentage does not mean the issue is unimportant. In reliability work, even a 0.01% probability can matter if the consequence of failure is severe or if the scenario is repeated many times. On the other hand, a large raw upset probability does not automatically mean the end user will notice a problem, because corrected errors may never escape hardware or software safeguards.
It is also useful to compare time windows. A device that has a very low chance of a bit flip during one hour may still have a meaningful chance over a month of continuous operation. Likewise, doubling the memory roughly doubles the expected number of opportunities for an upset, and increasing altitude raises the event rate further. This is why data-center memory protection, avionics validation, and long-running embedded systems often pay so much attention to ECC, scrubbing, and fault handling even when single-bit events are individually rare.
In short, think of the result as a probability of raw exposure. If the number is comfortably low for your use case, the simple model may be enough for rough planning. If the number is high, or the application is safety-critical, it is a signal to collect better SER data, model the environment more carefully, and review your mitigation stack in detail.
