A4 Standard delivers higher per-core performance and greater memory bandwidth than prior Ampere-based OCI A1 and A2 instances. For a 16vcpu (8 OCPU; 1 OCPU=2 AmpereOne M cores) VM configuration, customers can expect sizeable performance improvements on A4 compared to the previous generation A2, as shown in the results published by Oracle. For reference, a 24% performance improvement was measured on SPECrate®2017_int_base (est.) on A4 compared to A2.

With industry-leading pricing at $0.0138 per OCPU-hour and $0.0027 per GB-hour, A4 Standard delivers exceptional price-performance value. Industry standard benchmarks such as SPECrate®2017_int_base (est.) demonstrate 16% higher performance and 27% better price–performance on OCI A4 compared to AMD EPYC Turin–based E6 shapes.

Fig 1: SPECrate®2017_int_base (est.) Performance and Price-Performance for OCI A4 vs. E6 VMs

As organizations accelerate AI adoption, CPU-based inference offers a balance of performance, cost, and operational simplicity, making it a compelling solution for deploying models at scale.

LLMs today demand more than just raw compute. They depend on high memory bandwidth, large token buffers, and fast data movement across cores. AmpereOne M was designed specifically to meet these needs.

With 192 cores, 12 channel DDR5 and a large, distributed cache architecture, it delivers highly scalable inference, predictable multi-tenant performance, and seamless integration into existing cloud environments. This architecture also allows for larger context windows and more complex model graph structures.

Results on STREAM Triad, a popular benchmark that measures memory bandwidth, show A4 Standard throughput at 143 GB/sec vs. E6 at 53 GB/s, a 3.8x performance advantage for A4, underscoring its suitability for LLM inference.

Fig 2: STREAM Triad Performance (GB/s) for OCI A4 vs. OCI E6 VMs

To further enhance LLM deployment and performance, AmpereOne M is supported by Ampere® AI Optimizer (AIO), a framework that helps developers transform, optimize and tune LLMs for Ampere’s architecture. It streamlines quantization, layer fusion, and format conversion to improve throughput and reduce latency, without requiring model rewrites. Ampere also maintains upstream contributions and ecosystem integrations with popular AI frameworks, ensuring that developers can bring models to production with minimal friction.

NVIDIA A10 is the only small AI inference solution in OCI. By comparing the number of simultaneous AI users that can be supported while still meeting a Service Level Agreement (SLA) of 10 tokens/second, a 90-core A4 Standard shape delivers a 2x price–performance advantage over NVIDIA’s A10 Bare Metal GPU instance on Llama 3.1 8B model inference.