AI is not about one chip type; CPUs, GPUs, and interconnects must work together at scale: Manik Kapoor, AMD India

In this recent interaction with Express Computer, Manik Kapoor, Senior Manager – Sales, Data Centre and AI Group, AMD India, outlines how the company is expanding access to AI-ready infrastructure across India’s rapidly evolving enterprise landscape. Kapoor explains why AI infrastructure demands a full-spectrum, workload-specific approach, spanning CPUs, GPUs, interconnects, and open software, and how AMD’s open ecosystem is helping enterprises, regional cloud providers, and mid-sized organisations deploy scalable, power-efficient AI faster, without vendor lock-in. From sovereign cloud platforms to real-world AI deployments in healthcare, he shares how AMD is trying to position itself at the centre of India’s AI build-out.

How is AMD broadening access to AI-ready infrastructure across India’s enterprise landscape?

In AI, there is no one size fits all approach – powering AI at scale requires diverse, purpose-built compute and an open ecosystem that brings together the capabilities of the entire industry to accelerate innovation.

AI is not about one chip type. Besides the parallel computing capabilities of GPUs, CPUs remain indispensable in AI pipeline, handling pre-processing, orchestration, and data access, keeping GPUs optimised for heavy inference workloads. Equally critical is interconnectivity, the glue that enables true performance at scale. Not every AI workload needs data centre-class compute. AMD’s full-spectrum compute portfolio, from AI PCs and edge servers to hyperscale infrastructure – ensures organisations can deploy the right-sized solution for their unique needs. With industry-leading CPUs, GPUs, adaptive computing, and networking, AMD is well-positioned to deliver scalable, efficient AI systems across the full compute continuum.

The AMD ROCm software platform plays a pivotal role in broadening access to AI development. ROCm is an open software stack designed to enable high-performance AI and HPC development across a wide range of AMD GPUs. It supports popular frameworks like PyTorch and TensorFlow, and provides tools, libraries, and APIs that simplify memory management, distributed execution, and workload orchestration. ROCm also supports portability across edge, cloud, and on-prem environments, helping organisations build and scale AI workloads flexibly and efficiently without relying on proprietary ecosystems. Backed by a vibrant open-source community, ROCm empowers Indian enterprises with transparency, flexibility, and long-term support as they build next-generation AI solutions.

What role does AMD play in enabling regional cloud providers to offer AI-as-a-Service, and what integration and support features help them deploy AI models quickly?

Regional cloud providers play a key role in bringing AI closer to users and meeting data sovereignty needs. Our focus is to help them offer AI-as-a-service in a way that is fast, flexible, and cost-friendly.

We support this by giving them a complete AI stack with AMD EPYC processors, AMD Instinct GPUs, and an open software ecosystem through ROCm. ROCm works with the most popular AI frameworks and includes ready-to-use libraries, containers, and model tools. This removes much of the early engineering effort and helps the deployment of AI models into production faster.

On the integration side, we provide reference designs, optimised containers, support for Kubernetes and Slurm, and direct guidance from our engineering teams during deployment. This usually brings the time to deployment down from months to weeks and lets regional cloud partners launch AI services without complex rework.

In India, we work with next generation cloud providers such as NxtGen, Neysa, and Reliance Jio that use AMD EPYC processors and AMD Instinct accelerators for their sovereign cloud platforms. These services support AI, analytics, and real-time enterprise workloads while meeting data localisation needs. This is how we help regional cloud providers bring AI to market quickly while giving them an open and flexible platform for future growth.

Can you share examples of AMD-driven AI applications in sectors like BFSI, telecom, or healthcare?

AMD has developed an innovative tool which can act as an extension to e-Sanjeevani, India’s telemedicine platform. EyeCare AI – Your Virtual Ophthalmologist is an innovative healthcare solution to address doctor shortages and language barriers. It leverages advanced LLMs to provide AI-driven diagnosis and prescription assistance.

The software solution runs on AMD EPYC CPUs and AMD Instinct GPUs deployed on our OEM partner platforms. The tool supports over 22 official languages, and enables low-cost, high-resolution imaging, capturing eye images and patient feedback in regional languages, generating AI-assisted diagnoses and prescriptions on the spot.

What differentiates AMD’s latest GPU and CPU offerings for AI workloads compared to other market solutions, especially when serving mid-sized enterprises?

The latest AMD CPUs and GPUs stand out due to their leadership in performance, energy efficiency, and cost-effectiveness, which are crucial for mid-sized enterprises balancing workload demands and TCO.

The 5th Gen EPYC 9005 Series processors offer up to 192 high-performance cores with exceptional memory bandwidth, enabling up to 86% fewer racks compared to legacy hardware and significant reductions in power consumption and licensing costs.

The AMD Instinct MI350 Series GPUs provide scalable AI acceleration with up to 288 GB of HBM3E memory, support for massive models (up to 520 billion parameters), and flexible cooling options for easy integration.

Together with AMD ROCm, our open-source, developer-first AI software platform, these offerings enable seamless integration into existing environments. ROCm supports industry-standard frameworks like PyTorch and TensorFlow, and is optimised for containerised workflows such as vLLM and Hugging Face TGI. This provides mid-sized enterprises with the flexibility to deploy AI without vendor lock-in, accelerate time to value, and leverage a rapidly growing open ecosystem backed by community and industry collaboration.

Additionally, AMD leads in supporting open standards like Ultra Accelerator Link (UALink) and Ultra Ethernet Consortium (UEC), which simplify scaling AI workloads. UALink provides a high-speed, low-latency “scale-up” interconnect for dense GPU clusters, enabling accelerators to access shared memory efficiently as a single, large GPU. UEC focuses on scalable, robust “scale-out” networking, facilitating large, distributed clusters with high bandwidth and congestion management. These open standards ensure multi-vendor interoperability, reduce deployment complexity, and future-proof enterprise AI infrastructure.

Together, these innovations make AMD’s CPU and GPU platforms easier to deploy, manage, and scale for mid-sized enterprises pursuing AI initiatives.

How does AMD address the power efficiency and scalability requirements of AI inferencing on its platforms, and what specific innovations are driving this?

AMD enables efficient, scalable AI inferencing through a combination of high-performance silicon, advanced interconnects, networking, and open software. EPYC processors offer power-optimised SKUs like the 9575F for inference orchestration, while Instinct MI350 Series GPUs deliver up to 288 GB of HBM3E memory for large model support and high throughput at lower energy cost.

To scale inference across nodes, AMD supports open standards like UALink and UEC, alongside Infinity Fabric, to ensure low-latency, high-bandwidth performance. Our ROCm software stack further optimises resource utilisation with native support for PyTorch, TensorFlow, and inference-ready containers like vLLM.