AI and Workload Capabilities

What are the Power11 processor AI capabilities?

IBM Power11® introduces a set of integrated AI acceleration technologies designed to run enterprise AI workloads efficiently on the CPU, without relying on GPU acceleration. These capabilities are available across the entire Power11 family, including the IBM Power S1112, and are built directly into each processor core, enabling consistent AI performance regardless of system size or deployment location.

On‑chip AI acceleration: Matrix Multiply Assist (MMA)

The core of Power11's AI capabilities is the Matrix Multiply Assist (MMA) engine, a hardware accelerator embedded inside every Power11 core. MMA offloads matrix multiplication operations - the computational foundation of neural networks and classical machine learning - directly to the processor. This enables high‑throughput execution of Deep learning inference, Tensor operations and Mathematical routines used in AI and analytics.

Power11 retains four MMA units per core, allowing substantial parallelism. Compared to Power10, Power11 increases effective MMA throughput through higher core density, improved data‑feeding efficiency, and better memory bandwidth. These enhancements accelerate inference workloads without requiring specialized external hardware.

Because MMA is embedded in the CPU pipeline, latency is extremely low, enabling fast scoring, classification, and decisioning workloads — ideal for platforms such as the S1112, which often operates at the edge or in distributed sites.

Vector and SIMD acceleration through VSX

Power11 also leverages the Vector Scalar Extension (VSX) architecture, a SIMD instruction framework that accelerates vector math across large data arrays. VSX unifies floating‑point and vector registers into the Vector-Scalar Register (VSR) file, supporting:

• Vector arithmetic
• Permute and shuffle operations
• Multiply‑accumulate operations
• Parallel reductions and dot products

These capabilities accelerate numerical kernels used by:

• Tensor algebra
• Signal processing
• Scientific computing
• Classical machine learning (SVMs, regressions)
• Matrix decomposition libraries

VSX complements MMA by handling broader vector workflows, enabling Power11 systems, including the S1112, to run medium‑scale AI and HPC workloads entirely on the CPU.

Mixed‑precision support for efficient AI processing

Modern AI frameworks commonly use reduced precision formats (for example, INT8 and FP16) to increase throughput and lower latency. Power11 includes hardware support for mixed‑precision operations, enabling:

• Higher inferencing throughput
• Reduced power consumption
• Better cache utilization
• Lower memory bandwidth requirements

This makes Power11 particularly well suited to CPU inference scenarios where responsiveness matters more than extremely large model sizes.

AI architectural enhancements for Power11

Beyond MMA and VSX, Power11 introduces broad architectural improvements that benefit AI workloads:

• Up to 25% more cores per chip, increasing parallel throughput
• Higher memory bandwidth via DDR5 OMI memory
• Improved cache hierarchy for faster data reuse
• Better CPU‑to‑CPU interconnects (in multi‑socket systems)
• Improved energy efficiency, supporting AI loads at lower power draw

These enhancements ensure that AI workloads scale well and deliver consistent performance even under sustained load.

On a compact, single‑socket entry server like the S1112, these benefits translate directly into faster inference, more responsive analytics, and the ability to colocate AI workloads alongside IBM i, AIX, or Linux applications.

Compiler, toolchain, and Python ecosystem optimizations

AI workloads depend heavily on optimized mathematical libraries. Power11 introduces enhancements across the software stack to ensure that frameworks and tools can exploit the hardware accelerators.

GCC and OpenBLAS optimizations

Starting with GCC 10.2 and later versions, the compiler includes built‑in MMA intrinsic functions and the -mcpu=power11 flag. This enables libraries such as OpenBLAS to compile MMA‑optimized matrix kernels automatically. These optimizations propagate to higher‑level packages, meaning Python libraries like Numpy, Scipy and Scikit-learn inherit MMA acceleration automatically.

Python runtime enhancements

Recent Python releases (3.13/3.14) have introduced Free‑threaded execution (PEP 703 / PEP 779), removing GIL-related bottlenecks and Experimental JIT compilation, reducing interpreter overhead. When combined with Power11's increased core density and MMA acceleration, these improvements significantly boost performance for Python‑based AI and data science workloads.

Summary

The Power11 processor architecture integrates AI acceleration directly into every core through Matrix Multiply Assist, VSX SIMD enhancements, improved memory bandwidth, and mixed‑precision support. Compiler and runtime optimizations extend these capabilities into the Python ecosystem, enabling efficient AI inference, analytics, and mathematical workloads on all Power11 systems—including the compact S1112. These built‑in accelerators allow organizations to deploy AI closer to their data, reduce reliance on GPUs, and support production‑grade inference with strong energy efficiency and enterprise‑class reliability.

References

• Implementing AI on Power11: Introducing the IBM Spyre Adapter
• IBM Power11 Scale-out servers: introduction and overview

No IBM Spyre Accelerator support for the Power S1112

The IBM Spyre AI Accelerator is not supported with the Power S1112 system. This does not mean the system cannot run AI workloads — it relies instead on the integrated CPU-based Matrix Math Acceleration (MMA) capabilities described in other topics. This makes it well-suited for local AI workloads, particularly in edge computing environments.

Why the Spyre Accelerator is not supported

The Power S1112 cannot accommodate the IBM Spyre AI Accelerator due to two fundamental incompatibilities: hardware and memory.

Hardware requirement

The Power S1112 supports only half of the ENZ0 I/O Expansion Drawer. Power11 Spyre offerings require a fully populated ENZ0 drawer, which the S1112 cannot provide.

Memory requirement

The IBM Spyre AI Accelerator requires a minimum of 1 TB of RAM. This exceeds the maximum memory capacity of the Power S1112.

Summary

RequirementIBM Spyre AI AcceleratorPower S1112ENZ0 Expansion DrawerFully populated ENZ0 requiredHalf ENZ0 onlyMinimum Memory1 TB RAMLess than 1 TB (max)

Because the Power S1112 supports only half of the ENZ0 drawer and cannot meet the minimum memory requirement, the IBM Spyre AI Accelerator cannot be used with this system. AI inference workloads on the Power S1112 are instead handled by the built-in Matrix Math Acceleration (MMA) engine in the Power11 processor.

CPU-based AI inference workloads on S1112

What this covers

This topic explains how the Power S1112 leverages its integrated CPU-based AI acceleration capabilities to run inference workloads efficiently without requiring external GPU hardware.

CPU-based AI inference architecture

The Power S1112 uses its built-in Matrix Multiply Assist (MMA) engines and Vector Scalar Extension (VSX) capabilities to perform AI inference directly on the CPU. This architecture enables organizations to deploy AI workloads at the edge, in branch offices, or alongside existing IBM i, AIX, and Linux applications without the complexity and cost of GPU infrastructure.

Suitable workload characteristics

CPU-based inference on the S1112 is well-suited for:

• Small to medium-sized language models
• Classification and regression tasks
• Lightweight natural language processing
• Document analysis and OCR
• Anomaly detection
• Time-series forecasting
• Edge analytics

Performance considerations

The S1112's CPU-based approach provides:

• Low latency for real-time inference
• Consistent performance across workloads
• Energy-efficient operation
• Simplified deployment without GPU drivers or frameworks
• Direct integration with existing applications

Key takeaways

• S1112 provides CPU-based AI inference through integrated MMA and VSX acceleration
• Suitable for small to medium AI models and edge deployment scenarios
• Eliminates need for external GPU hardware in many use cases
• Enables AI workloads to run alongside traditional enterprise applications

Matrix Math Accelerator (MMA) features

The IBM Power MMA is an on-chip AI Accelerator for high throughput AI inferencing, machine learning, and Matrix Math processing.

Beginning with IBM Power10 and enhanced with the Power11 platform, each core integrates four Matrix Math Accelerators (MMAs) that support a wide range of AI inference workloads directly on the CPU. This innovation enables customers to run AI models natively on IBM Power cores without requiring a discrete GPU. The architecture is suitable for traditional AI use cases such as fraud detection, text extraction, document analysis, domain adaptation through retrieval augmented generation (RAG), pattern recognition, forecasting, and image, video, or audio processing.

The Power11 processor further enhances these capabilities. Although it does not always match the raw throughput of high-end GPUs in certain generative AI scenarios, Power11 technology provides excellent performance for tasks such as entity extraction, translation, summarization, and classification, while offering reduced energy consumption and improved data protection. By enabling AI workloads to run closer to where the data is, Power11 processor-based servers support secure, efficient, and scalable deployment of AI without unnecessary data movement.

MMA architecture

Each Power11 core contains four MMA units that can perform matrix multiplication operations in parallel. These units are tightly integrated with the CPU pipeline, providing:

• Low-latency matrix operations
• High throughput for AI inference
• Support for multiple precision formats
• Efficient memory bandwidth utilization

Supported operations

The MMA units support various matrix operations essential for AI workloads:

• Matrix multiplication (GEMM operations)
• Dot product calculations
• Vector-matrix operations
• Mixed-precision arithmetic

Performance characteristics

MMA acceleration provides:

• Significantly faster inference compared to standard CPU operations
• Consistent performance across different model types
• Energy-efficient computation
• Scalable performance with core count

Key takeaways

• Four MMA units per Power11 core enable parallel AI acceleration
• Suitable for a wide range of AI inference workloads
• Provides CPU-based AI without requiring external GPUs
• Enhanced in Power11 with improved throughput and efficiency
• Supports mixed-precision operations for optimal performance

Which AI workloads are suitable for S1112?

This section presents clear, practical guidance on the types of AI workloads that perform well on the IBM Power S1112. Because the S1112 relies exclusively on its CPU‑based AI acceleration — through Matrix Multiply Assist (MMA), VSX SIMD execution, and high‑bandwidth DDR5 memory — it is best suited for lightweight and medium‑complexity inference workloads, AI‑enabled edge processing, and IBM i‑adjacent analytics running in Linux partitions. These capabilities align with the system's compact form factor, low power requirements, single‑socket architecture, and tight integration of IBM i, AIX, and Linux environments through PowerVM.

The examples in this section illustrate appropriate use cases for the S1112's on‑chip AI capabilities. They highlight how organizations can deploy meaningful AI inference without requiring external acceleration, while placing modern analytical and AI‑driven services close to the workloads and data they support.

Automated document processing in remote offices

Many distributed or branch office environments depend on manual document handling: processing invoices, forms, service tickets, claims, or intake documents. The S1112 can automate these tasks using lightweight OCR pipelines and text classification models running in a Linux LPAR. These models can extract content, classify document types, identify key fields, or trigger workflow actions.

A typical implementation includes:

• A Linux LPAR hosting Python‑based OCR and document‑classification tools
• Local inference for extracting structured data
• Secure handoff to IBM i applications for storage or downstream processing

Running these models locally reduces latency, maintains uptime even during network outages, and avoids bandwidth overhead associated with transmitting raw documents to centralized systems. This scenario is ideal for CPU‑based MMA acceleration, as OCR post‑processing and text classification rely heavily on vector math and mid‑sized embedding models.

Local demand insight for retail inventory systems

Retail environments can use the S1112 to run inference models that analyze point‑of‑sale data, foot‑traffic patterns, or inventory signals to generate short‑term demand projections. Lightweight forecasting models such as gradient‑boosting regressors, tree‑based ensembles, or compact neural networks can run efficiently on a Linux LPAR.

Key characteristics of this scenario include:

• Real‑time inference with predictable CPU latency
• Local processing for rapid decision making and store autonomy
• Tight integration with existing IBM i inventory and POS applications
• Minimal resource footprint suitable for back‑office deployment

The S1112's memory bandwidth helps accelerate time‑series analysis and numerical pipelines used in small retail prediction workloads, while Linux LPAR isolation ensures the store's IBM i processing remains unaffected.

Real‑time anomaly detection in manufacturing environments

Manufacturing floors produce continuous telemetry from equipment—vibration data, thermal readings, acoustic signatures, and cycle timing. The S1112 is well suited for hosting real‑time anomaly detection models that operate directly on this telemetry to detect early signs of failure or inefficiency.

An example deployment would include:

• A Linux LPAR running time‑series inference models
• Feature extraction pipelines using VSX acceleration
• Lightweight neural networks or statistical anomaly engines
• Integration with IBM i‑based MES or maintenance systems

Processing at the edge minimizes latency, ensures continued operation during WAN interruptions, and avoids sending sensitive operational data to offsite systems. Models of this type typically require frequent small tensor multiplications rather than large deep learning workloads, making them a good match for the S1112's CPU‑based MMA engines.

Entry‑level AI through the Wallaroo AI Starter Kit (WASK)

The Walaroo AI Starter Kit complements the S1112 by enabling turnkey AI adoption for IBM i customers with minimal infrastructure overhead. WASK runs in a dedicated Linux LPAR and includes:

• A pre‑packaged AI inference model
• A pre‑built AI API service
• A managed runtime optimized for Power11 CPUs
• Automatic updates without requiring Linux administration
• A fully contained environment that acts as a "black box" AI engine

WASK provides a low‑risk pathway for organizations beginning their AI journey, especially those without internal data science teams. The S1112's compact footprint and IBM i compatibility make WASK especially attractive for:

• Remote or branch office deployments
• Small and midsized IBM i environments
• Environments seeking minimal operational overhead

Common workloads supported through WASK include:

• Conversational digital assistants for HR, sales, and IT
• Internal question‑answering models using enterprise documents
• Product or catalog search
• Document‑summarization tasks
• Lightweight inference pipelines
• Python‑based analytics and rapid experimentation

By colocating a simple AI runtime alongside IBM i, organizations avoid complex data movement and benefit from local, low‑latency inference.

Lightweight analytics and CPU‑based ML pipelines

Beyond industry‑specific examples, the S1112 is a strong platform for a broad category of small to mid‑scale analytical models, including:

• Logistic and linear regression‑based scoring
• Small embedding models
• Clustering algorithms (K‑Means, DBSCAN)
• PCA or dimensionality reduction
• Signal processing with vectorized kernels
• Small transformer‑derived models optimized for inference

These workloads rely more on vector math and MMA‑accelerated matrix operations than on deep accelerator stacks. With support from optimized libraries (NumPy, SciPy, OpenBLAS using -mcpu=power11) and emerging Python runtime optimizations, the S1112 can deliver strong CPU inference performance for many business‑ready AI models.

Summary

The IBM Power S1112 is well matched to edge workloads, CPU‑based AI inference, IBM i‑adjacent intelligence, and lightweight analytical pipelines. It delivers meaningful AI acceleration through its built‑in CPU features (MMA, VSX, mixed precision, and high‑bandwidth DDR5 memory) without requiring external hardware.

Suitable workloads emphasize local inference, small-to-medium model sizes, and direct integration with existing IBM i, AIX, or Linux applications. The S1112's combination of compact design, strong CPU vector performance, and integrated AI‑acceleration features makes it an efficient and practical platform for deploying operational AI into distributed, branch, and departmental environments.

References

• Wallaroo AI Starter Kit

S1112 hardware limitations

Core

The S1112 is limited to either 4 cores or 10 cores depending on the configuration.

• 10 cores available in rack-mount configuration only.
• IBM i: maximum 4 cores and 64 GB memory total across all IBM i partitions.
• Licensing at 0.25 core granularity.

Memory

Maximum memory is 4 DDR5 DDIMMs, 512 GB total.

Internal storage

Maximum 4 NVMe drives, 12.8 TB total internal storage.

I/O drawer

1/2 ENZ0 I/O Expansion Drawer.

PCIe slots

4 PCIe slots: 2 PCIe Gen4 x16 direct, 2 PCIe Gen4 x8 direct.

RDX

No RDX support.

Comparison with S1012

The performance improvements seen on Power11-based systems, such as the S1112, arise from a series of architectural enhancements introduced with the generation transition from Power10 to Power11. While the exact uplift varies by workload, several design changes consistently contribute to improved AI/ML inferencing efficiency:

• Higher core density in the Power11 processor design, enabling more parallel MMA-accelerated operations within a similar power envelope
• Refinements to cache hierarchy and memory subsystem behavior, improving throughput for data-intensive and memory-bound inference workloads
• Increased effective memory bandwidth when moving from DDR4 (S1012) to DDR5 (S1112)
• Pipeline and instruction-scheduling improvements that enhance core-level efficiency, especially in workloads involving dense linear algebra and transformation kernels
• Continued use of four MMA engines per core, benefiting from a more efficient surrounding architecture

These changes collectively allow the S1112 to handle CPU‑driven AI/ML inference workloads more efficiently than the S1012, even though both rely on the same fundamental on‑core MMA acceleration architecture (Power ISA v3.1).

Workload validation performed on Power11‑based systems demonstrated reduced execution times in several representative Python AI/ML examples, aligning with the architectural gains expected when moving from a Power10‑based design (i.e. S1012) to Power11 (i.e. S1112). Although the test results were not collected on these specific models, they illustrate the broader performance scaling trends inherent to the Power11 architecture.

Comparison summary

The following table summarizes the key differences between Power S1012 and Power S1112 for AI/ML inferencing workloads:

AspectPower S1012Power S1112ComparisonProcessor architecturePower 10-basedPower 11-basedS1112 benefits from improvements on the processor architecture with the transition from Power10 to Power11.Core densityLower core densityUp to 25% higher core densityPower11 enables more parallel MMA-accelerated operations within a similar power envelopeCache and memory subsystemEarlier-generation cache and memory behaviorRefined cache hierarchy and memory subsystemS1112 improves throughput for data-intensive and memory-bound inference workloadsMemory technologyDDR4DDR5S1112 gains increased effective memory bandwidth over S1012Pipeline efficiencyPower10 pipeline behaviorImproved pipeline and instruction schedulingS1112 improves core-level efficiency, especially for dense linear algebra and transformation kernelsMMA engines per coreFourFourBoth systems retain four MMA engines per core; S1112 benefits from a more efficient surrounding architectureISA / MMA foundationPower ISA v3.1 with on-core MMA accelerationPower ISA v3.1 with on-core MMA accelerationFundamental MMA architecture is the same on both systemsAI/ML inference efficiencyBaselineHigherS1112 handles CPU-driven AI/ML inference workloads more efficientlyObserved workload behaviorOlder-generation baselineReduced execution times in several Python AI/ML examplesValidation on Power11-based systems aligns with expected gains, though not measured on these exact models

Note: The performance observations are based on Power11-based workload validation trends and not on direct side-by-side measurements of these exact server models. For more details on AI/ML performance testing on Power10 and Power11, see the Workload Validation section in Implementing AI on Power11: Introducing the Spyre Adapter.

References

• Implementing AI on Power11: Introducing the IBM Spyre Adapter
• IBM Power S1012 Technical Overview and Introduction

When to use larger Power11 systems

The IBM Power S1112 provides meaningful on‑chip AI acceleration through its integrated Matrix Multiply Assist (MMA) and VSX SIMD capabilities, making it well suited for small‑to‑medium inference workloads, edge analytics, and IBM i‑adjacent AI functions running from Linux partitions. However, some AI workloads require greater compute density, memory capacity, accelerator support, or cluster-scaled orchestration than a single‑socket S1112 can provide.

This section outlines clear decision criteria for when organizations should move to larger Power11 systems, such as the Scale-out Power L1122/L1124, S1122/S1124 or the midrange and enterprise Power E1150/E1180 platforms, to meet advanced AI requirements.

AI model size exceeding CPU-only capabilities

The S1112 performs well when running:

• small-to-medium transformer or encoder models
• classification and regression algorithms
• forecasting and signal‑processing models
• anomaly detection pipelines
• OCR and lightweight natural language tasks
• Small Language Models (SLMs) optimized for CPU inference

However, once models grow beyond this range, CPU-only execution becomes insufficient. It is advised to use larger Power11 systems when the models require:

• multi‑billion parameter model inference
• large transformer architectures
• batch inference at high concurrency
• multi-modal or high-throughput embedding generation
• RAG pipelines requiring vector DB throughput beyond the S1112's memory/compute limits

These workloads demand significantly more computational throughput than a single Power11 socket can provide. Larger Power11 systems deliver higher memory capacity, larger core counts, and critical access to off‑chip accelerators such as the IBM Spyre Accelerator.

Off‑chip AI acceleration (IBM Spyre Accelerator)

Spyre is a PCIe‑attached AI accelerator designed for high‑throughput, low‑latency processing of modern AI workloads, including multi-precision training and inference. It provides:

• 32 low‑power AI compute cores per module
• High throughput for vector/matrix operations
• Support for FP16/INT8/INT4 low‑precision numerics
• Energy‑efficient acceleration at just ~75W per card
• Linear scaling using multiple Spyre cards in an accelerator expansion drawer
• Support for Red Hat OpenShift AI and frameworks such as PyTorch and vLLM

The S1112 does not support Spyre. Therefore, AI workloads requiring Spyre must use Power11 systems such the Power L1122/L1124, S1122/S1124, E1150 or E1180 (supports multiple drawers; high-end AI). For example, workloads that require enterprise‑scale LLM inference or high‑throughput embedding generation for semantic search. It also applies to Retrieval‑Augmented Generation (RAG) pipelines that depend on large context windows and sustained vector processing.

Models that require low‑precision acceleration such as INT4, INT8, or FP16 also benefit from these systems. The same is true for environments that must serve many inference requests per second or host multiple models or model instances concurrently. Power11 servers equipped with the IBM Spyre Accelerator are specifically designed for these scenarios and deliver the performance needed for advanced modern AI workloads.

Training AI models

The S1112 is well suited for inference, but it is not designed for training modern deep learning models. Larger Power11 systems are required when workloads involve transformer training, deep neural network training, or any scenario that demands large, sustained vector and matrix throughput. Training tasks often require memory footprints that exceed the S1112's capacity, along with fast data movement across NUMA domains — capabilities available only in multi‑socket systems. Distributed training workflows also rely on multi‑node orchestration that the S1112 cannot provide. Power E1150 and Power E1180 systems address these needs with higher memory bandwidth, bigger system memory configurations, and cluster‑scale training capabilities suitable for advanced AI development.

AI workloads exceeding the memory capacity of the S1112

The S1112 supports up to 512 GB of Memory, which is sufficient for most small language models, classical machine learning workloads, and edge‑focused inference tasks. However, some AI workloads demand significantly larger memory footprints. Environments that involve memory‑intensive LLM serving, vector database indexing over large document collections, multi‑model hosting with substantial resident memory requirements, or analytic workloads that exceed 1 TB of Memory will require larger Power11 systems. The Power E1150 and Power E1180 platforms address these needs by offering multi‑terabyte memory configurations and higher system bandwidth, enabling them to support large‑scale, memory‑bound AI workloads effectively.

AI workloads demanding high concurrency or complex orchestration

Some AI environments require the ability to serve large numbers of small inference requests in parallel or coordinate multiple pipeline stages across several partitions or containers. In these situations, larger Power11 systems become necessary. High‑concurrency inference pipelines, multi‑partition or multi‑node OpenShift AI deployments, load‑balanced model serving, distributed feature‑extraction workflows, and scenarios where multiple AI services must run simultaneously, such as embeddings, LLM inference, and re‑ranking, quickly exceed the capacity of a single‑socket S1112. Systems such as the Power L1122/L1124, S1122/S1124, and the Power E1150/E1180 offer higher core density, multiple Shared Processor Pools, and larger overall system capacity, enabling them to sustain demanding, highly orchestrated AI workloads.

AI workloads requiring greater storage or I/O throughput

Some AI pipelines depend on fast, scalable I/O to move data efficiently between storage, memory, compute engines, and accelerators. These workflows often require multiple PCIe adapters, NVMe-heavy storage layouts, high‑throughput SAN connectivity for vector databases, or multi‑adapter SR‑IOV networking to support distributed inference. Because the S1112 is a compact single‑socket system with limited PCIe expansion, it cannot support the larger I/O footprints needed for these scenarios. Larger Power11 systems such as the L1122, L1124, S1122, S1124, E1150, or E1180 provide greater I/O riser capacity, higher PCIe bandwidth, and the ability to attach accelerator drawers such as the IBM Spyre enclosure. These platforms are better suited for AI workloads that rely on parallel I/O paths, extensive NVMe storage pools, or multiple network fabrics to achieve the required performance and scalability.

Enterprise-scale AI services

Enterprise‑class AI Services such as Granite model inference, embedding generation, RAG pipelines, re‑ranking, and entity extraction typically rely on transformer‑based architectures, large context windows, heavy vector math, and the ability to index and search large document stores while serving multiple models concurrently. These workloads exceed the compute, memory, and throughput capabilities of a single S1112 system. Deployments of this scale require either larger Power11 systems equipped with the IBM Spyre Accelerator or distributed clusters designed for high‑volume AI processing. Organizations should move to larger Power11 platforms when planning to implement IBM's enterprise AI Services stack.

Summary

The Power S1112 is well suited for CPU‑based inference, small‑to‑medium AI models, edge analytics, and Linux‑based AI services that operate alongside IBM i workloads. However, some AI requirements extend beyond the capabilities of a compact single‑socket system. Organizations should consider larger Power11 platforms such as the L1122, L1124, S1122, S1124, E1150, or E1180 when:

• model sizes outgrow CPU‑only execution
• Spyre acceleration is required
• workloads involve AI training rather than inference

Larger systems also become necessary when:

• memory demands exceed 512 GB
• high concurrency or complex orchestration is needed
• substantial PCIe or storage expansion is required
• deploying enterprise‑scale AI services such as LLMs, embeddings, or RAG pipelines.

Larger Power11 systems provide higher compute density, more memory, greater scalability, and access to off‑chip acceleration, enabling enterprise-class AI workloads that the S1112 is not intended to support.

References

• Implementing AI on Power11: Introducing the IBM Spyre Adapter
• IBM Power E1180: Introduction and Overview
• IBM Power11 E1150 Introduction and Technical Overview
• IBM Power11 Scale-Out Servers: Introduction and Overview