We present the novel micro-architectural features, supported by an innovative and novel pre-silicon methodology in the design of POWER10. The resulting projected energy efficiency boost over POWER9 is 2.6x at core level (for SPECint) and up to 3x at socket level. In addition, a new feature supporting inline AI acceleration was added to the POWER ISA and incorporated into the POWER10 processor core design. The resulting boost in SIMD/AI socket performance is projected to be up to 10x for FP32 and 21x for INT8 models of ResNet-50 and BERT-Large. In this paper, we describe the novel methodology deployed and used not only to obtain these efficiency boosts for traditional workloads, but also to infuse AI/ML/HPC capability directly into the POWER10 core.

This paper introduces a new static verification technique using tree-grammars. The core contribution is the combination of a structural netlist traversal with parser generation techniques for tree-grammars. Today's commercial static analysis tools offer a rich set of parameterized connectivity checks, but their predefined nature prevents effective checks on the highly customized structures found in high-end processor designs. The method presented here allows to formulate the required connectivity using a tree-grammar, thus combining high checking flexibility with convenient specification. Unlike other grammar based structural verification approaches, this method does not require the complete netlist to be matched against the production rules, which allows short runtimes even on large multi-core chip netlists. Results are presented for the most recent 22nm high-end processor designs.

In recent technology nodes, reliability is considered a part of the standard design ¿ow at all levels of embedded system design. While techniques that use only low-level models at gate- and register transfer-level offer high accuracy, they are too inefficient to consider the overall application of the embedded system. Multi-level models with high abstraction are essential to efficiently evaluate the impact of physical defects on the system. This paper provides a methodology that leverages state-of-the-art techniques for efficient fault simulation of structural faults together with transaction-level modeling. This way it is possible to accurately evaluate the impact of the faults on the entire hardware/software system. A case study of a system consisting of hardware and software for image compression and data encryption is presented and the method is compared to a standard gate/RT mixed-level approach.

Fault simulation is essential in test generation, design for test and reliability assessment of integrated circuits. Reliability analysis and the simulation of self-test structures are particularly computationally expensive as a large number of patterns has to be evaluated.

This paper provides a methodology that leverages state-of-the-art techniques for efficient fault simulation of structural faults together with transaction level modeling. This way it is possible to accurately evaluate the impact of the faults on the entire hardware/software system.

High delay-fault coverage requires rather sophisticated clocking schemes in test mode, which usually combine launch-on-shift and launch-on-capture strategies. These complex clocking schemes make low power test planning more difficult as initialization, justification and propagation require multiple clock cycles. This paper describes a unified method to map the sequential test planning problem to a combinational circuit representation. The combinational representation is subject to known algorithms for efficient low power built-in self-test planning. Experimental results for a set of industrial circuits show that even rather complex test clocking schemes lead to an efficient low power test plan.

The complexity of the test infrastructure and test strategies in systems-on-chip approaches the complexity of the functional design space. This paper presents test design space exploration and validation of test strategies and schedules using transaction level models (TLMs). Since many aspects of testing involve the transfer of a significant amount of test stimuli and responses, the communication-centric view of TLMs suits this purpose exceptionally well.

Structural on-line self-test may be performed to detect permanent faults and avoid their accumulation. This paper improves concurrent BIST techniques based on a deterministic test set. Here, the test patterns are specially generated with a small number of specified bits. This results in very low test latency, which reduces the likelihood of fault accumulation. Experiments with a large number of circuits show that the hardware overhead is significantly lower than the overhead for previously published methods. Furthermore, the method allows to tradeoff fault coverage, test latency and hardware overhead.

An effective technique to save power during scan based test is to switch off unused scan chains. The results obtained with this method strongly depend on the mapping of scan flip-flops into scan chains, which determines how many chains can be deactivated per pattern.

Hardening a circuit against soft errors should be performed in early design steps before the circuit is laid out. A viable approach to achieve soft error rate (SER) reduction at a reasonable cost is to harden only parts of a circuit. When selecting which locations in the circuit to harden, priority should be given to critical spots for which an error is likely to cause a system malfunction. The criticality of the spots depends on parameters not all available in early design steps. We employ a selection strategy which takes only gate-level information into account and does not use any low-level electrical or timing information. We validate the quality of the solution using an accurate SER estimator based on the new UGC particle strike model. Although only partial information is utilized for hardening, the exact validation shows that the susceptibility of a circuit to soft errors is reduced significantly. The results of the hardening strategy presented are also superior to known purely topological strategies in terms of both hardware overhead and protection.

Error correcting coding is the dominant technique to achieve acceptable soft-error rates in memory arrays. In many modern circuits, the number of memory elements in the random logic is in the order of the number of SRAM cells on chips only a few years ago. Often latches are clock gated and have to retain their states during longer periods. Moreover, miniaturization has led to elevated susceptibility of the memory elements and further increases the need for protection. This paper presents a fault-tolerant register latch organization that is able to detect single-bit errors while it is clock gated. With active clock, single and multiple errors are detected. The registers can be efficiently integrated similar to the scan design flow, and error detecting or locating information can be collected at module level. The resulting structure can be efficiently reused for offline and general online testing.

Many STUMPS architectures found in current chip designs allow disabling of individual scan chains for debug and diagnosis. In a recent paper it has been shown that this feature can be used for reducing the power consumption during test. Here, we present an efficient algorithm for the automated generation of a test plan that keeps fault coverage as well as test time, while significantly reducing the amount of wasted energy. A fault isolation table, which is usually used for diagnosis and debug, is employed to accurately determine scan chains that can be disabled. The algorithm was successfully applied to large industrial circuits and identifies a very large amount of excess pattern shift activity.

In this paper, we propose a new programmable deterministic built-in self-Test (BIST) method that requires significantly lower storage for deterministic patterns than existing programmable methods and provides high flexibility for test engineering in both internal and external test. Theoretical analysis suggests that significantly more care bits can be encoded in the seed of a linear feedback shift register (LFSR), if a limited number of conflicting equations is ignored in the employed linear equation system. The ignored care bits are separately embedded into the LFSR pattern. In contrast to known deterministic BIST schemes based on test set embedding, the embedding logic function is not hardwired. Instead, this information is stored in memory using a special compression and decompression method. Experiments for benchmark circuits and industrial designs demonstrate that the approach has considerably higher overall coding efficiency than the existing methods.

Built-in self test is a major part of the manufacturing test procedure for the cell processor. However, pseudo random patterns cause a high switching activity which is not effectively reduced by standard low power design techniques. If special care is not taken, the scan-speed may have to be reduced significantly, thus extending test time and costs. In this paper, we describe a test power reduction method for logic BIST which uses test scheduling, planning and scan-gating. In LBIST, effective patterns that detect additional faults are very scarce after a few dozens of scan cycles and often less than one pattern in a hundred detects new faults. In most cases, such an effective pattern requires only a reduced set of the available scan chains to detect the fault and all don't-care scan chains can be disabled, therefore significantly reducing test power