Bare-Metal Execution

Definition and Context

In post-silicon hardware validation, bare-metal execution describes a mode in which verification software runs directly on the silicon prototype without the support of an operating system, debuggers, device drivers, or other system software ^[1]. Because early silicon prototypes are typically unstable and lack a complete software stack, bare-metal execution is often the only option available to bring-up engineers, and tools intended to run in this mode must be deployed "in the very early stages of the post-silicon validation effort when the OS cannot yet be run on the system" ^[2].

A tool that operates in this mode is called a bare-metal exerciser. Threadmill, developed at IBM Research – Haifa, is one such tool ^[2].

Characteristics

Bare-metal execution environments used for post-silicon validation exhibit several defining characteristics:

• No operating system services — file systems, device drivers, and standard I/O facilities are unavailable ^[3].
• No debugging infrastructure — interactive debuggers and on-platform observability mechanisms are typically absent or extremely limited ^[1].
• Limited memory and state — some platforms (e.g., on-wafer testers) provide very limited memory for programs and data ^[3].
• High overhead for memory and state operations — loading and offloading memory and state information is expensive because no OS services can mediate the transfers ^[3].
• Real-time execution speed — once running, bare-metal code executes at full silicon speed without OS-induced jitter ^[3].

Design Implications for Tools

The constraints of bare-metal execution drive several design decisions for any tool that must run in this mode:

1. Simplicity — Because hardware failures are hard to debug without observability tools, the exerciser itself must be simple software ^[2].
2. On-platform self-sufficiency — A bare-metal exerciser, once loaded on silicon, generates, executes, and checks tests "indefinitely" without further interaction with the environment ^[4].
3. Lightweight generation engine — Because silicon time is expensive, the on-platform test generation engine must be fast and light, in contrast to more sophisticated pre-silicon generators ^[2].
4. Pre-computed payloads — Heavy computations such as floating-point operand generation are performed off-platform; only their results are embedded in the exerciser image ^[3].
5. Minimal interaction with environment — The tool should avoid operations such as reading files from I/O devices, which are not supported in bare-metal mode ^[2].
6. Lightweight checking — Reference models and complex checking machinery are less appealing because off-platform usage requires costly result offloading and on-line usage consumes precious execution cycles ^[2].

Checking Strategies Under Bare-Metal Constraints

Because traditional reference models are impractical, bare-metal exercisers typically rely on lightweight checking methods:

• Multi-pass consistency checking — Each test-case is executed multiple times and the end-of-test values of selected system resources (memory, registers) are compared for consistency ^[2].
• Injected and natural variability — The effectiveness of consistency checking depends on existing or injected differences between passes, such as variations in thread scheduling or system operation modes ^[2].
• User-specified self-checking procedures — Verification engineers may supply procedures that verify predictable aspects of an abstract scenario ^[2].

Host-Side Execution of Bare-Metal Images

Instruction-accurate instruction set simulators also commonly run bare-metal executables compiled for a target ISA, providing a host-side environment with no operating system on the emulated target. ExactStep, for example, supports RISC-V (RV32IMAC, RV64IMAC), MIPS (mips-i), and ARM-v6m targets and "is used to run bare-metal executables compiled for those ISAs" ^[5]. ExactStep executes one instruction per call to cpu::step(), prioritizing ease of extension for cosimulation, peripheral development, and SystemC bus interfacing over raw execution performance ^[5].

Related Concepts

For more detail on a concrete realization of a bare-metal exerciser, see the article on Threadmill, the tool used in the verification of the IBM POWER7 processor ^[1]. For an instruction-accurate simulator that runs bare-metal images on a host, see ExactStep ^[5].