What it is
IHI 0070 is Arm’s architecture specification for the System Memory Management Unit, the SMMUv3 generation. A CPU core translates its own loads and stores with an MMU. A GPU, NPU, display controller, Ethernet controller, storage controller, or DMA engine does not. Those masters issue reads and writes that carry their own addresses. The SMMU sits on the path into shared DRAM and decides which physical page a transaction may touch, or whether the transaction is refused.
The document is a programmer’s model and a behavior specification, not a gate-level design. Two SoCs can both implement it and still differ in how many devices are attached and which optional features exist. For a newcomer it is the IOMMU of an Arm application processor: virtual memory for hardware masters other than the CPU.
Who it applies to
It applies on Arm MPUs and HPC SoCs used as vehicle computers, zonal controllers, and ADAS boards, wherever a non-CPU master can read or write system RAM. Silicon integrators implement the block. Hypervisor and kernel engineers program it. Safety and security engineers depend on that programming when they claim one master cannot overwrite another’s buffers.
A microcontroller with no SMMU in its memory map is out of scope. Many Classic AUTOSAR ECUs never include this function. Whether a board is affected depends on the SMMU that silicon actually implemented, not on the mere presence of an Arm core.
What it covers
For every device transaction, software must be able to say which address space applies and what is permitted.
The source is named by a StreamID. An optional SubstreamID separates several address spaces behind one device, such as queues or contexts on an accelerator. Software places a stream table in memory. Each entry selects a context: stage-1 translation, stage-2 translation, or both, plus the base of the page tables.
Stage 1 is the operating-system or guest mapping, from a device virtual address to an intermediate physical address. Stage 2 is the hypervisor mapping, from that intermediate address to a real physical page, and the check that rejects a DMA outside the guest’s memory. Translation-table formats follow the Arm virtual memory model, so page size, permissions, and memory attributes mean the same thing for a device as for a core.
Software talks to the SMMU through queues. A command queue carries invalidations, configuration updates, and synchronization. An event queue reports faults. The architecture defines how translations may be cached and when software must invalidate them. Optional behavior covers devices that translate addresses themselves and ask the SMMU for permission, and the split between secure and non-secure accesses. The silicon manual states which of these the part actually provides.
| Term | What you use it for |
|---|---|
| StreamID | Identify the device or DMA channel |
| Stage 1 | Apply that software stack’s own virtual map |
| Stage 2 | Limit a guest to its own physical pages |
| Event queue | Report a refused or untranslatable DMA |
How it fits with other documents
IHI 0070 is the device-side companion of Arm virtual memory. CPU translation and virtualization are specified in the Arm Architecture Reference Manual. The SMMU reuses that table model so one hypervisor policy can cover both instruction execution and DMA. The SoC technical reference manual is what says how many SMMU instances exist, which StreamID is wired to which master, and which options were built.
ISO 26262 does not specify an SMMU. A safety argument for an HPC that mixes ASIL and QM work can still treat a correctly configured SMMU as spatial isolation between bus masters. A cybersecurity case can treat it as a control that bounds a compromised accelerator or network interface. Those standards set the risk and the rigor of the argument. IHI 0070 describes streams, stages, and queues, and it does not assign an ASIL or a security level.
Where the SoC connects PCIe or a similar fabric, endpoints may offer their own address-translation services. The interconnect specification defines that protocol. IHI 0070 defines how the memory system behaves when those requests reach the SMMU.
In practice
Read the silicon manual before writing software. Confirm that an SMMU is present, which revision it claims, which masters are attached, and whether stage 2, SubstreamIDs, and stall-on-fault exist. Assign StreamIDs and page tables before any guest may start DMA, and give that configuration a single owner, usually the hypervisor or the host kernel.
Turn translation on only after the stream table is visible to the SMMU. When software later changes a table the SMMU may have cached, it has to issue the architecture’s invalidation commands. Flushing the CPU TLB and leaving the SMMU alone is a typical reason DMA works until the first remap, then corrupts memory or stops.
A fault is delivered on the event queue, not as a data abort on the core. The symptom the rest of the system sees is often a timeout in a camera, network, or inference path, because a buffer was never filled. The StreamID in the event is what ties that timeout to a missing map or to a guest that left its stage-2 region.
Review the finished map with the safety interference analysis and the security boundary. A guest that can rewrite another guest’s stream table has removed the isolation the block was integrated to provide.