The version using the 4-core CPU Intel N97, which is specified by Intel as an embedded CPU, certainly has in-band ECC, which was tested by some reviewers of this SBC.
I do not know whether the 8-core version (H4 Ultra) also enables in-band ECC, as for that CPU Intel does not specify embedded uses, so they may disable the ECC support in the factory.
I see that someone has enabled successfully in-band ECC on the 8-core ODROID H4 Ultra and has run benchmarks with ECC disabled/enabled. Therefore it appears that in-band ECC support exists on all models.
The results of benchmarks with in-band ECC disabled/enabled may be not representative for real workloads. In-band ECC relies on caching the ECC bits in a dedicated ECC cache, in order to avoid excessive memory accesses. The effectiveness of the ECC cache can be very different for the benchmark and for the real workload, leading to misleading results. Usually for the real workload it is likely that the cache hit-rate will be higher, so the performance drop with in-band ECC enabled will be less conspicuous.
alt Hacker News
The version using the 4-core CPU Intel N97, which is specified by Intel as an embedded CPU, certainly has in-band ECC, which was tested by some reviewers of this SBC.
I do not know whether the 8-core version (H4 Ultra) also enables in-band ECC, as for that CPU Intel does not specify embedded uses, so they may disable the ECC support in the factory.
See e.g.:
https://www.cnx-software.com/2024/05/26/odroid-h4-plus-revie...
However, looking right now at:
https://forum.odroid.com/viewtopic.php?f=171&t=48377
I see that someone has enabled successfully in-band ECC on the 8-core ODROID H4 Ultra and has run benchmarks with ECC disabled/enabled. Therefore it appears that in-band ECC support exists on all models.
The results of benchmarks with in-band ECC disabled/enabled may be not representative for real workloads. In-band ECC relies on caching the ECC bits in a dedicated ECC cache, in order to avoid excessive memory accesses. The effectiveness of the ECC cache can be very different for the benchmark and for the real workload, leading to misleading results. Usually for the real workload it is likely that the cache hit-rate will be higher, so the performance drop with in-band ECC enabled will be less conspicuous.