AI is present at Advanced Micro Devices

Meta description: AMD uses GitHub Copilot to triage one million nightly simulations and an SAP AI assistant to shorten customer order inquiries and release engineering capacity.

Tags: AMD, GitHub Copilot, SAP S/4HANA, AI in engineering, supply chain AI

Featured image: Abstract semiconductor circuitry with nightly simulation logs and order-routing lines, no people, no logos, no text.

AMD GitHub Copilot and a generative AI assistant built on SAP S/4HANA are the two internal workflows highlighted by Advanced Micro Devices as its 2025 filing reports a larger AI budget. The case studies show why the work matters: pre-silicon verification produces about one million nightly simulations, while changed customer ship dates can force up to 14 manual verifications.

Background: AMD's AI Budget and 2025 Scale

Advanced Micro Devices designs processors and other semiconductor products for data centers, personal computers, gaming systems and embedded devices, and also supports AI computing. Its 2025 annual filing reported $34.6 billion revenue, including $16.6 billion from the Data Center segment, up 32% from 2024. Research and development spending rose 25% to $8.1 billion, mainly due to higher employee-related costs from headcount expansion for the AI strategy. AMD also reported acquiring teams with software compilers, machine learning, photonics and AI inference expertise, covering hardware and software needed for AI applications.

Development: GitHub Copilot Moves from Repositories to Overnight Triage

Before manufacturing, engineers run virtual tests to confirm intended chip behavior. A failure may indicate a design defect, a test problem, or a known issue. Harry Foster, chief scientist for verification at Siemens, explained that verification oversights can require multimillion-dollar manufacturing reruns to correct designs and disrupt product cycles. Combining multiple chiplets into integrated systems increases the workload. AMD corporate fellow Alex Starr said the complexity created roughly four verification engineers for each hardware designer, and sorting known failures consumes capacity needed for new design problems.

AMD consolidated hardware and software repositories on GitHub Enterprise before introducing Copilot. GitHub's customer account describes an agent checking simulation logs against historical issues and proposing fixes. AMD worked with GitHub to adapt the tool to hardware description languages such as Verilog, addressing context general-purpose AI models initially lacked. The GitHub video Powering next-generation engineering: AMD + GitHub accompanies the account. The agent triages overnight, giving engineers sorted results and proposed fixes, while they still investigate new failures and validate that changes resolve underlying design problems.

The account reports automated triage of more than 1 million nightly simulations and zero manual morning triage time. It provides no baseline hours, total engineering review time after deployment, triage-accuracy measure or chip-quality outcome, and broader productivity claims cannot be attributed solely to this workflow. More than 20,000 engineers use GitHub Enterprise; the number using the verification agent is undisclosed. A separate paper by AMD verification engineers describes human review of AI-generated assertions before formal testing, but that different application does not establish how engineers approve nightly fixes or escalate unresolved failures. For chip-development leaders, the reported reduction in sorting is a starting point; measuring unresolved morning failures, accepted fixes, established-test results and recurring defects would test whether capacity is released without harming verification quality.

Progress: SAP S/4HANA and an OpenAI GPT Assistant Shorten Order Questions

A second workflow followed operational consolidation. Beginning in January 2023, AMD moved its operations and acquired Xilinx business onto SAP S/4HANA. AMD's information technology team worked with SAP's Lighthouse generative AI program to build an assistant that helps employees investigate changed delivery commitments and answer customer order inquiries faster. It analyzes order allocation and supply availability to identify causes and recommend action. When a customer-facing ship date changes, supply chain specialists must explain why to planning and customer teams. An AMD account published by SAP AppHaus says the investigation could require up to 14 manual verifications; finding an allocation problem or supply shortage helps teams decide the response and customer message.

A specialist enters a sales order number through the interface, and SAP Business Technology Platform connects the assistant to the enterprise resource planning environment. It analyzes allocation timing, constrained quantities, available supply, scheduled delivery and material availability. The output supports investigation and corrective action; the published workflow does not show automatic changes to order commitments. AMD's architecture diagram, sourced from AMD's information technology document Transforming Order Scheduling with Generative AI, shows the specialist application connected to SAP S/4HANA and AI services. An architecture presentation identifies an OpenAI GPT deployment and anonymization component. AMD later expanded access to customer operations and sales teams for self-service and added controls restricting sensitive order information to authorized personnel. AMD's as-is and proposed AI flow diagrams compare manual order checks with the AI-assisted inquiry, using the same source.

The company estimates the assistant saves 3,000 staff hours annually and reports shorter inquiry response times. Its target was reducing inquiry response times from 20 to 30 minutes to under two minutes. A later partner article describes a 90% reduction and 3,120 annual hours saved as achieved, but provides no measurement period or calculation method. AMD does not publish measured current response time, active-user counts, answer-accuracy results or customer-service outcomes; the differing savings figures cannot be reconciled or treated as demonstrated financial returns. Architecture materials leave model version, anonymization rules and employee validation or escalation process unspecified. The evaluation should pair response time with answer accuracy and further-specialist frequency, sample answers against enterprise resource planning records, and track first-contact resolution to show whether customers receive sufficient information.

Current Stage: Useful Claims Await Verification Metrics

Together, the cases point to three tests. For verification, track triage time with unresolved failures, accepted fixes and defects found in later tests. For order answers, sample allocation and delivery responses against enterprise resource planning data and record inquiries needing further investigation. For capacity, use observed inquiry volume and time saved after review, then identify how employees use released hours. The 2025 filing shows AMD's AI investment is rising, but GitHub and SAP accounts stop short of full before-and-after measurement. Next evaluation will depend on current response times, agent accuracy, escalation outcomes and whether released hours shift to new chip or supply-chain work.