I was reading this interesting challenge on Reddit:

https://www.reddit.com/r/CIO/comments/1rxtfcy/our_cio_just_asked_for_our_ai_adoption_number_and/

And I started thinking about it. How would you monitor the progress of implementation of AI in your organization. Can we modify the Capability Maturity Model to build an AI Maturity Model? One that shows how an organization moves from ad hoc experiments to repeatable, governed, and continuously improving AI capability. CMMI’s progression from Initial to Optimizing provides a strong backbone for this kind of model, and AI-specific governance can be anchored in structured risk practices such as Govern, Map, Measure, and Manage.?

Why a maturity model matters

AI programs often start as isolated pilots, but scale only when strategy, data, governance, delivery, and risk controls become consistent across teams. A Gartner report says 86% of all pilots failed in 2025. There is another report that says 99% failed. Depending on which report you read and what its data sources and timeline considerations are the numbers change. But one thing is consistent. They are all pessimistic and bleak.

A maturity model gives leaders a common language to assess where they are today, identify gaps, and prioritize investments that turn AI from a series of experiments into an enterprise capability.

A simple AI maturity model

You can track maturity across seven dimensions: Strategy & Governance, Data Management, AI Model Development, AI Deployment & Operations, Skills & People, Risk & Compliance, Performance Measurement & Value, and Agentic AI & Scaling. Each dimension should be rated from Level 1 to Level 5, where Level 1 is mostly manual and reactive, and Level 5 is optimized, automated, and continuously improving.

How to use it

The easiest way to apply this model is to score each dimension independently, then look for the lowest-scoring areas that could block scale. For example, an organization may have strong model development but weak governance, which usually means pilots will keep growing faster than controls. A monthly or quarterly review works well because maturity changes more slowly than project delivery, but still needs visible executive ownership.

What good looks like

At Level 3, the organization has standard processes and shared ways of working. At Level 4, it can measure AI performance, risk, and value with consistent metrics. At Level 5, AI becomes a managed enterprise capability, with continuous improvement, automated controls, and governance built into delivery rather than added afterward.

This is just the beginning of my thoughts. Let me know what you think. How is your organization tracking the progress of AI?

#AIMaturity #EnterpriseAI #MLOps #DigitalTransformation #Leadership

Most boards obsess over AI strategy while quietly starving the very thing that makes it possible: a resilient, scalable infrastructure backbone.

I have often noticed that Infrastructure investments are not taken as seriously as other IT and Technology investments are. I am not saying that infrastructure teams are left completely high and dry, but often it so happens that projects are planned without keeping in mind the infrastructure investments. And when it comes to cutting costs, infrastructure teams are given targets that are unreasonable without compromising on the expectations that are there from them.

Considering AI is booming and there are quite a few companies aspiring to go AI first, this becomes even more important. I see organizations putting way too much focus on AI models whether it is foundational models like GPT, Claude, Mistral, etc.; Or domain specific models like BioGPT or BloombergGPT or fine tuned enterprise models like Customer support copilots or document summarization models, and such.

But Data and Infrastructure are the foundation of any organization going AI first. Is your data quality good? Do you have good data governance? What’s the architecture?

And then topic of our conversation here. Infrastructure and HPC. Things such as HPC Clusters, GPU/TPU infrastructure, Cloud Landing Zones, Storage for ML, Networking for high throughput workloads, MLOps compute orchestration. These are very important topics that need to be considered before we start thinking about AI Products and Models.

Of course there is also culture and ways of working but I am not going to talk about that right now. What I am trying to say is if organizations have an AI strategy, infrastructure plays a very important role. While the executives don’t have to get into every nook and corner understanding every technical jargon, it is important that they take technology and infrastructure seriously and include them in the board meeting agenda.

AI strategy without a corresponding infrastructure strategy is a liability in disguise. Boards that treat infrastructure as a core part of business risk, resilience, and growth — not just as IT plumbing — will be the ones whose AI ambitions actually make it into production and stay there.

Enterprise Architecture is a field in which people who have spent years in technology usually enter. That’s been the trend. And counter intuitive as it may be, the skills needed to be an enterprise architecture are not what you would expect from someone who has spent an entire career doing technology. Let me explain. Usually people who come to EA are introverts, head down and do your work, geeks. This is because, for some reason, Enterprise Architecture is seen as an extension to Technology Architecture (as defined in the industry). Which is an extension to being a Senior Engineer. And people who excel at Engineering (programming, infrastructure, technology) can excel only if they spend tremendous amounts of time studying. Learning new languages, new paradigms, new technologies. And usually, people who are introverts can do this more easily considering they don’t have as much of a social life as an extrovert. I am an introvert and I read 50 books in a year.

But EA is a leadership function. It has more to do with people than it has with technology. Stakeholder Management, Strategy, Organizational Priorities, Leadership Vision. These are the things they eventually end up doing. Sure initially they will have to go through the grind of collecting information, building the repository, building reports and dashboards. But with AI coming in so fast, I am thinking very soon the grind work will be out of Enterprise Architecture. Which leaves the sexy part of the job to the EAs.

EA (Enterprise Architecture) is not very different from EA (Executive Assistant) in one particular context. An EA (in both cases) works with the leader, gets them the necessary information so that they can do their job better. Enterprise Architecture has to do with Processes, Org Structure, Systems and Technology and Executive Assistants work with Emails, Calendars, Facilitation and Administrative tasks. But primarily it’s the same thing. All decisions are taken by the stakeholder. EAs Facilitate.

And that’s why I don’t care much about the EA job as a designation. I do, however, care about the knowledge of Enterprise Architecture. I believe understanding the organization better helps you become a better, stronger leader. You take better decisions.

And so if you want to become an Enterprise Architect to escape leadership responsibilities. If you are an introvert and think EA is a good step up from your current individual contributor role, think again. Sure in an advisory role you will be able to gather information and help leaders make decisions. But you will have little control over the decisions. And if your leader doesn’t care much about Enterprise Architecture, which is usually the case in the industry today, over and over you will see leaders making decisions you don’t agree with and there is little you will be able to do.

So I say learn EA and then strive for power. Become a leader. Hold the decision making authority in your hands. Because you care about the organization, about its long term future, about the right strategy, you are a better leader. And being a better potential leader, it becomes your responsibility to take the reigns. Or someone else will. And that won’t be good for anyone.

Here are some of the Artificial Intelligence related acronyms and their brief explanations. Now this is a bit of a starter for you. If you find any of these fascinating, you should do some further research. Let this not be the end.

AI (Artificial Intelligence): The simulation of human intelligence in machines, allowing them to perform tasks that typically require human intellect, such as learning, problem-solving, and decision-making.

ML (Machine Learning): A subset of AI where algorithms analyze data, learn patterns, and make predictions or decisions with minimal human intervention.

DL (Deep Learning): A type of machine learning that uses multi-layered neural networks to analyze and learn from large amounts of data, often used for complex tasks like image and speech recognition.

NLP (Natural Language Processing): No it is not Neurolinguistic Programming :-). This is the branch of AI that enables machines to understand, interpret, and respond to human language in a meaningful way.

CV (Computer Vision): AI that allows machines to interpret and make decisions based on visual inputs, such as images and videos.

RL (Reinforcement Learning): A type of machine learning where an agent learns to make decisions by receiving rewards or penalties based on its actions in a given environment.

ANN (Artificial Neural Network): Computing systems inspired by the structure and function of the human brain’s neural networks, used for pattern recognition and decision-making.

GAN (Generative Adversarial Network): A type of neural network architecture where two networks compete to generate realistic data, often used for creating images and other media.

RNN (Recurrent Neural Network): Neural networks designed for processing sequential data, such as time series or text, by maintaining a memory of previous inputs.

CNN (Convolutional Neural Network): Neural networks specifically designed for analyzing visual data, such as images, by using layers that mimic the human visual cortex.

AGI (Artificial General Intelligence): The hypothetical concept of AI that possesses the ability to understand, learn, and apply knowledge across a wide range of tasks, similar to human intelligence.

ASR (Automatic Speech Recognition): Technology that converts spoken language into text, enabling voice control and transcription services.

HCI (Human-Computer Interaction): The study of how people interact with computers and designing technology that facilitates this interaction in an efficient and user-friendly manner.

IoT (Internet of Things): The network of interconnected devices that communicate and exchange data with each other, enabling smart systems and automation.

LSTM (Long Short-Term Memory): A type of recurrent neural network architecture designed to remember information over long periods, useful for tasks like language modeling and time series prediction.

BERT (Bidirectional Encoder Representations from Transformers): A pre-trained NLP model that understands the context of words in a sentence by considering both the left and right surroundings.

GPT (Generative Pre-trained Transformer): A language model that generates human-like text by predicting the next word in a sentence based on the context provided.

RPA (Robotic Process Automation): Technology that uses software robots or “bots” to automate repetitive and rule-based tasks, improving efficiency and accuracy in various business processes.

HITL (Human-in-the-Loop): AI systems that involve human interaction to provide feedback, validate results, or guide the AI’s learning process for improved accuracy and reliability.

MITL (Machine-in-the-Loop): Systems where machines assist humans in decision-making by providing data analysis, predictions, and recommendations to enhance human capabilities.

I hope you have enjoyed this, and I really hope I am a little more creative next week. :-)

I am sure you must have heard about Deepseek. If you haven’t, here’s a quick rundown.

What is Deepseek?
Deepseek is a Chinese AI Startup founded in 2023.

What’s the ruckus about Deepseek?

Deepseek is in the news in the AI world lately because of its innovative and extremely cost-effective models such as Deepseek Coder and Deepseek LLM. They are popular because of their extreme performance and their extremely low costs. They are comparable with OpenAI’s GPT-4 in performance and yet at a fraction of the cost. In fact, it has rattled the entire western world and has brought the tech stocks crashing down. People are asking if AGI can be developed at a fraction of the cost what are the billions of dollars being used for by the leaders of OpenAI? There are also concerns about Geopolitical instability as more and more western developers turn towards Deepseek. In fact, some of the people in the pentagon in US were found to be using Deepseek until it was banned.

And it's not just that Deepseek has made quite a few of its models open source. Deepseek-V3 is a MoE language model with 671 billion parameters trained with 14.8 trillion tokens and is available on Github and Hugging Face. To give you more context, each parameter within the 671 billion parameters is like a brain cell or a neuron in our brain. Each one capable of making decisions, like a small brain in itself. So the more the parameters, the more nuanced and complex the decisions the AI can make. GPT-4 has 1.8 trillion parameters.

Well so if GPT-4 has more than twice as many parameters, what’s the big ruckus about? Apples to apples, deepseek performs better than GPT-4 on several benchmarks. And that’s because of its Mixture of Experts (MoE) Architecture. This means it activates only a subset of its parameters at a time, to accomplish a task, which makes it a lot more efficient.

Deepseek R1 is also open sourced and is known for its advanced reasoning capabilities. It rivals OpenAI’s Model o1 and is available on the Web App and API. And they continue to contribute to the Open-Source community by providing its powerful models for public use.

How did they make it so cost effective?

Hardware: Deepseek used Nvidia’s H800 GPUs which are older and less powerful. They optimized their hardware usage to achieve higher performance.

Algorithm Efficiency: Deepseek focused on Algorithmic efficiency rather than using brute force hardware power. They used mixed precision training and advanced pipeline algorithms to make the most out of their outdated hardware.

Innovative Techniques: they used PTX Programming, which is very much like assembly language, which allowed them a much more efficient use of the GPU resources. It is very challenging to squeeze out the very last resource out of the GPU resources, but this demonstrates the skills of the engineers on the job.

Resource Constraints: Classic example where resource constraints actually compel people to become more innovative and come up with more efficient with available resources. The western world had put export controls and hardware limitations which actually worked in their favor.

What does this mean for you?

Innovation doesn’t require large resources. On the contrary, lack of resources often becomes the trigger for innovation. Look up Richard Stallman, Denis Ritche, Kenneth Thompson. These were some of the original stalwarts of the Open Source Community. And I think throwing money at a problem is the easiest of the solutions, but not always the best. Perhaps it is time we look at the open source community to leverage it, to contribute to it.

Thoughts?

Once upon a time in a galaxy far far away, there was a bandwagon. It was called the Cloud Bandwagon. Everyone tried to climb on it. Then the bill came and then everyone tried to jump ship. So what is it? Is cloud good or bad?

One thing I am sure everyone has come to realize by now is that climbing a bandwagon because everyone else is climbing it is never a good idea. Critical thinking before taking a major steps is always a good idea. So let’s explore.

Moving to cloud because everyone else is moving to the cloud and you don’t want to be left behind, is a bad reason to move to cloud. Moving to cloud because you want to save money reduce capital expenditure and optimize on operational expenditure is also not the best reason to move to cloud. I say, “not the best” because cost savings are possible. Cloud infrastructure is definitely a lot more optimized than a standard on-premises data center because of magnitude and scale.

But there is no apples-to-apples comparison of cost. Think about it. We have to calculate the total cost of managing the on-prem data centers, divide it by compute, storage, network, power, cooling, licenses, managed services cost on operations, rent for the space, etc. So there is no standard formula. And then Compute, storage and network cost will depend on the brand and quality of the hardware. Managed services will depend on the contract negotiation that you have with the service provider. And then there is the matter of hardware depreciation. Generally we know that compute is about 45–50% of your total cost and storage is about 10%, power and cooling 1%, rent 1% and so on. But these are educated guesses. We can get ballpark figures but can’t be ever sure.

So if you are counting exact savings from a move to cloud, it is tricky. Best I would say is if you are smart about the migrations and don’t lift and shift everything as is, if you re-platform and refactor and modernize, the cost can be less than on-prem data centers because of economies of scale. But then again, that’s not the best reason for moving to the cloud.

I think the best reason is modernization. AI is a big thing and we already know it. One of the key benefits of moving to cloud is it preps your landscape to become AI Enabled. How? Scalability, Data & AI Services and Innovation and Updates.

We all know Cloud is scalable. Increase or decrease resources as required on the fly. But with AI and HPC this takes it to the next level. So I am not going to talk a lot about this here.

Let’s talk about Data Management and AI Services. On AWS, S3, Glacier and EFS were always there, but they are not very different from our traditional data centers. But AWS provides DynamoDB and Aurora. And that’s just the beginning. We have Sagemaker, DataZone, BedRock that allow us to play with our data. Train models, catalog data, perform governance and what not. And then we have other ancillary services like Textract, Rekognition that can be used for various small tasks in coordination with existing models.

And then let’s talk about Innovation and Updates. Iterations are quicker, turnaround times are better and deployments are faster on Cloud. Forget everything else, just the fact that we don’t have to buy hardware itself makes a huge dent on time to market. Prototyping becomes a breeze. And the speed at which the technology is evolving, we cannot afford to not be agile.

Having our applications on AWS modernizes our landscape and prepares it for leveraging AI. But only if done right. The question is, are we re-platforming and modernizing? Are we using the appropriate native services? Are we truly Cloud Smart?

Last week I had a very interesting discussion with someone about the future of infrastructure teams and how the entire landscape is evolving.

Now we have been talking about High Performance Computing for the last couple of articles and it is definitely a very interesting to talk about. I mean we definitely need to understand the business and do everything that we can to support them and one of the big demands that’s coming is around AI and HPC considering the AI Enabled vision that we have now.

But let’s look inwards a little bit. How are we evolving as infrastructure professionals?

So let’s start by asking another question. How much can we automate? And I am not talking about Terraform for provisioning automation or Ansible or RPA. Think AI Ops, think Agentic AI. AI Ops can automate and optimize performance monitoring, workload scheduling, data backups and incident response. We can use AI to peek into various sources of information like systems logs, performance metrics, network data and proactively resolve issues, reduce operational costs and improve overall efficiency.

Using AI we can autonomously perform complex tasks without human intervention. That’s what agentic AI is. We can have AI agents that do not require any human interaction to perform tasks. A good example would be an agent managing infrastructure by automatically adjusting resources based on real-time demand, ensuring optimal performance and cost.

There are tools available already. Take Splunk for example. It provides a host of capabilities that are a part of AI Ops. Like Anomaly detection, Event correlation (takes data from various sources to pinpoint the root cause of incidents), predictive analytics for issues along with proactive measures, automated incident response, performance monitoring, real-time threat detection, resource optimization and it also allows custom ML Model integration.

Now let’s talk about Agentic AI. There are some tools already, that provide capabilities that could, for example, autonomously manage infrastructure, allocating and deallocating resources, based on real-time demand. And then there is a whole ecosystem that could allow us to build and program such tools for specific purposes. And then there are service providers today that provide such specific services around building the ecosystem that could automate infrastructure management.

So from a tools and technologies standpoint, we are not far from achieving that level of automation. But does that mean we transform overnight? Even if we assume, for a moment, that these tools and technologies can be implemented by 5.00 pm tomorrow. Can we transform overnight? There are a lot of things that will have to change. Things such as processes, business integrations, organizational structures and what not. This is not just a technology transformation; it is an organizational transformation. Now I don’t want to get into all of that. But I can talk about the one aspect that we can control.

Which brings us to back to the original question that I started with. How are we evolving as infrastructure professionals?

The role of infrastructure professionals, in my opinion, should stop focusing on the mundane and move towards “more interesting” stuff as AI slowly comes and takes over the simple tasks. Few things come to mind.

• Strategic planning, long term goals, ensuring automated systems stay updated with the latest tech, but most importantly they continue to meet the business objectives.
• Monitoring, Oversight and continuous improvement ensuring everything continues to work. Here we are talking about ensuring there are no biases creeping in, and the results are as expected. Perhaps an engineering team continues to fine tune the algorithms where required, improving data quality, etc. This team can also focus on innovation, coming up with newer ways to use tech to meet business objectives.
• There will always be complex situations that autonomous agents cannot handle. Human intervention will be required there, so a small team, could continue to work on those areas.
• Then we have the job of ensuring that all automated process are compliant. There is also the case of responsible, secure and ethical AI that will require human intervention, at least for the foreseeable future. Of course once we become a part of the Matrix… Or are we aready…? :-)

The Infrastructure Professional role, like everything else, is evolving and I think its bringing along a lot of fun stuff. Fun if we are thought workers. Fun if we are malleable to change. Fun if we consider it fun

We have all heard about the Moore’s law. It predicts that the number of transistors on a microchip doubles approximately every two years, leading to increased performance and decreased costs. And I think we have broken the Moore’s law. Now interestingly, I know many articles that talk about Moore’s Law being broken in both directions. On one side I read that Moore’s Law is broken because our microchips are getting more than twice as fast every two years. On the other hand I also read that we can’t add twice as many transistors to a microchip because we have reached the physical limitations. But that’s not what this article is about. There are many other laws and metrics related to high performance computing and I thought this would be an interesting read.

Law of Diminishing Returns

Dates back to early 18th Century. French economist Anne Rober Jackues Turgot discussed diminishing returns in the context of agriculture. It states that in a production process, if one factor of production (such as labor) is increased with other factors (such as capital and land) are held constant, the resulting increase in output will eventually become smaller and smaller. Simply put, if you are thirsty, drink a glass of water. You will enjoy it. The second glass will provide a lesser satisfaction. The third glass will give you no satisfaction and the fourth glass may become punishment. That’s law of diminishing returns. Amdahl’s Law, Gustafson’s Law and Karp-Flatt Metric are all in a way related to the Law of Diminishing Returns.

Amdahl’s Law

A principle that helps us understand the potential speedup of a task when using multiple processors. Imagine there is some work and you have a friend to help you. Some steps, like mixing ingredients can be done together and will go faster with two people. However, there is other steps like baking the cake in the oven, that can only be done by one oven at a time. So this part cannot be sped up by adding more friends. Amdahl’s Law tells us that there are limites to how much faster a task can be completed by adding more helpers (processors). This highlights that the overall time to finish a job is constrained by the parts of the task that cannot be parallelized.

Gustafson’s Law

Imagine you have a big project and you have a group of friends who can help you finish faster. Gustafson’s Law helps you understand how much faster you can get the project done by using more friends. So instead of focusing on how much time you save by adding more people, like Amdahl’s Law, Gustafson’s Law focuses on increasing the size of the task you can complete with the extra help, without making it longer. In essence it shows us that by adding more resources (processors) you can tackle larger problems more efficiently, even if some parts of the projects can’t be further divided.

Karp-Flatt Metric

Proposed by Alan H. Karp and Horace P. Flatt in 1990, it is a measure of parallelization in parallel processor systems, designed to complement Amdahl’s Law and Gustafson’s Law. This is done by calculating the serial fraction of a parallel computation. Imagine you have a task that you want to split among several friends to get it done faster. This metric helps you understand how much of your work can’t be sped up by adding more people (or processors) because some parts of the task must be done one at a time. It also takes into account the extra effort needed to manage and coordinate everyone’s work. Sounds like Amdahl’s Law? Well the difference is Amdahl’s Law gives you a theoretical limit to speedup, while Karp-Flatt Metric provides a more nuanced, practical view of parallel performance, considering the real-world overheads.

Little’s Law

A fundamental theorem in queuing theory, which is a branch of operations research that deals with the study of waiting lines. Formally, Little’s Law is expressed as:

L=λ⋅WL

Where:

• L is the average number of items in the system (also known as the average “length” of the queue).
• λ is the average arrival rate of items into the system.
• W is the average waiting time an item spends in the system.

Little’s Law was first published in 1961 by John D. C. Little and yet it has its applications in high performance computing in 2025. Its beauty is in its simplicity and generality, allowing it to be applied to any system where items or people wait in line. We can use this for Memory Access Patterns, Buffer Management, Workload Distribution and measuring System Throughput and Response Times in High Performance Computing

Metcalfe Law

Formulated by Robert Metcalfe, the inventor of ethernet technology and a pioneer in computer networking. It helps us understand the value of networks, such as Internet or social media platforms. It states that the value of a network grows exponentially with the no. of its nodes. So in HPC, the more notes in the cluster, the more the cluster’s performance. Also as the interconnected devices on the network grows, the network’s value increases because the data can be transmitted more efficiently. In summary, it provides valuable perspective on the importance of networking in HPC

And finally, the last but not the least… The most feared law in the history of laws. The ghastly, the horrific. THE MURPHY’S LAW…!!!

The Murphy’s Law

Named after Edward A. Murphy Jr. an American aerospace engineer. The law originated in the context of aerospace engineering in the late 1940s. Simply put when you drop something, it will fall and role to the most remotest, most unreachable corner under the desk. When you are early to office, your boss is late and when you are late, the boss is early. When you put the USB cable in the USB Port, it will invariably and always be the wrong side up, unless you look and check first. Basically if things can go wrong they will. And in HPC, whether it is picking the wrong hardware, installation, network configuration, software installation, cluster management, job scheduling or data management and storage. There are a 100 things that can go wrong. And if we are not careful, like checking the USB before we stick it into the port, it surely will.

I talk about these laws because these laws and principles give us a unique perspective into High Performance Computing. I hope these pieces that I write serve as a beginning of the readers’ learning process. A place where they can start their research. Because that’s what it is for me.

This is the last and final part of this series. And I will try to connect all the dots in this one. In the first three parts, we talked about HPC and options that we have. We chose 5 specifically because they do different kind of activities.

1. Quantum Processors
2. Intel Data Center GPU Max Series
3. Amazon EC2 P5
4. NVIDIA BlueField DPU Series
5. Google Cloud A3 Ultra VMs

And then we picked 10 use cases from the pharmaceutical industry

1. Drug Discovery and Development
2. Molecular Modeling and Simulation
3. Clinical Trials and Optimization
4. Personalized Medicine
5. Genomic Data Analysis
6. Bioinformatics
7. Drug Repurposing
8. Predictive Analytics
9. Manufacturing Process Optimization
10. Market Analysis

So now connecting the dots.

First Quantum Processors. Google demonstrated that their quantum computer, Sycamore, could solve a specific problem in 200 seconds. This same problem would take the world’s fastest supercomputer, Frontier, approximately 47 years to solve! So we may assume that since Quantum Processors are so many magnitudes faster than other HPC hardware, they could be used anywhere. Well not so fast. While quantum processors have the potential to revolutionize computing, they are still not the silver bullet. They excel at certain tasks that involve complex simulations, optimization problems and certain types of cryptography due to its ability to perform parallel computations and exploit quantum phenomena like superposition and entanglement. However there are several challenges and limitations to consider.

• Error Rates: Quantum processors currently have high error rates and require sophisticated error correction techniques
• Coherence Time: Quantum states are fragiel and maintaining coherence for long periods is challenging
• Specific Use Cases: Quantum processors are best suited for specific types of problems and may not be efficient for all tasks

So while Quantum Computing has great potential in the future, it is more a hassle at the moment than an opportunity. Sure we can dabble and experiment, but for now for real use cases, I am not considering Quantum Processors.

So let’s look at some of these use cases and see what our best options could be.

Accelerated Drug Discovery and Development: Given the nature of drug discovery and development, which involves a mix of data-intensive and parallel processing workloads, the Intel Data Center GPU Max Series and Amazon EC2 P5 instances are likely to be the most suitable options. However, adding the NVIDIA BlueField DPU Series can enhance data and networking loads if needed. Having said that, if Google DeepMind and AlphaFold is being used for Molecular Modelling then Google Cloud A3 Ultra VMs have a seamless integration with these services enhancing overall efficiency and performance and hence have an advantage over the other two.

Clinical Trials and Optimization: NVIDIA H100 Tensor Core GPUs are specifically designed for deep learning and high-performance computing applications, making them ideal for AI-Driven Analysis and predictive modelling tasks. EC2 P5 also offer computational power, which is essential for handling data-intensive tasks like decrypting and analyzing OCT Images. Also, the P5 instances allows seamles sintegration with other AWS Services, enhancing data management, storage and security. So Intel Data Center GPU Max Series and Google Cloud A3 Ultra VMs are strong contenders, but Amazon EC2 P5 instances are optimized for specific AI and machine learning workloads involved in clinical trials.

Personalized Medicine: Under this category, I am going to take just 2 Cell and Gene Therapy and RadioLigand Therapy (RLT). For Cell and Gene we need Genomic Data Analysis processing and analyzing large genomic dtaasets to identify genetic mutations and targets for therapy. Considering this is heavy data-intensive workload, the NVIDIA BlueField DPU Series can be considered. Then there is Gene Editing, Cell Culturing and Expansion, Protein Modeling and Clinical Trial Simulation. All of these are parallel processing intensive activities and Either Google Cloud A3 Ultra VMs or AWS EC2 P5 instances can be used. These are both powered by H100s and specifically designed for Machine Learning and Deep Learning. As far as Radio Ligand Therapy is concerned, activities involved are molecular modeling, dosage optimization, radiation dosimetry (calculating the distribution and absorption of radiation within the body to ensure effective targeting of cancer cells) imaging analysis (processing and analyzing imaging data to monitor the distribution and efficacy of radioligands) and clinical trial simulations. All these involved high computational tasks and again the AWS EC2 P5 and Google Cloud A3 Ultra VMs are both good choices.

Bioinformatics: This is Advanced Quantitative Sciences. This requires high computational power, AI and Machine Learning Optimization, Scalability and Versatility. For all these cases the Intel Data Center GPU Max Series is the best option. Of course the H100 powered machines (both AWS and Google) are also very powerful and designed for AI / ML but there are subtle differences, with different optimizations and strengths. Intel Data Center GPU Max Series is specifically designed for versatility (broad range of high performance computing tasks), Scalability (can scale efficiently to handle large computational workloads) and Integration (may be better integrated with certain HPC environments that rely on Intel Architecture). On the other hand the H100s are optimized for AI and deep learning, known for high computational throughput and efficiency in parallel processing and include libraries and tools optimized for AI and HPC tasks. Here, I am going with Intel Data Center GPU Max series because Advanced Quantitative Sciences require Versatility, Scalability and Integration. And it is comparable with the H100 when it comes to computational performance as well

Predictive Analytics: Predictive Analytics requires AI / ML Optimizations, High Computational Power, Scalability and Integration with Cloud Services. So while Intel Data Center GPU Max Series is also powerful, the H100 in AWS EC2 P5 and Google Cloud A3 Ultra VM are more specialized for AI / ML tasks, making them the better choice for predictive analytics.

Manufacturing Process Optimization: This involves the Data Center Edge Optimization. We are talking about improving data processing speeds and reducing latency for manufacturing workloads and can be done using AI powered engineering solutions. NVIDIA BlueField DPU Series is the best option here. DPUs are specifically designed to enhance data center infrastructure, including networking, storage and security. They are also well suited for edge computing environments, where processing data closer to the source can significantly reduce latency and improve performance and DPUs can offload certain tasks from the main CPU, improving overall system efficiency and resource utilization.

That’s the end of the HPC series. Obviously designing HPC infrastructure for specific business use cases is not so simple. There is a lot more that must be considered, and this should only be the beginning of the learning process. Which, by the way, raises an interesting question as well. How much medical science do I need to know to be able to design an HPC solution for a business use case in the pharmaceutical industry. I mean I can’t be a Double PHD scientist now. I mean don’t get me wrong, I am just 48 years old, so technically still young ;-). But this is not the field of specialization that I have chosen.

I think I don’t need to be a double PHD. I do, however, need to have a good understanding of the requirements. What is it that they are trying to do with the HPC hardware. Is it data intensive or computation intensive or parallel processing intensive? Are they processing graphics? Is it latency sensitive? These are some things that need to be understood well. Of course understanding the science at a high level definitely helps there is no denying that.

So while there are ever changing innovations happening in the hardware world and the AI / ML / DL world is also ever evolving with new innovations happening every day, we must also evolve the integrations between the business and the technology world. The ways of working must change too. The word Agile takes on a very different meaning. Fail Fast is no longer a guideline that we may or may not follow. It becomes quintessential for success.

In the last episode we talked about the various areas that HPC and AI can impact a pharmaceutical industry. There were 10 topics I was going to talk about and we talked about the first five.

1. Drug Discovery and Development
2. Molecular Modeling and Simulation
3. Clinical Trials Optimization
4. Personalized Medicine
5. Genomic Data Analysis
6. BioInformatics
7. Drug Repurposing
8. Predictive Analytics
9. Manufacturing Process Optimization
10. Market Analysis

Let’s talk about the rest of the five today.

BioInformatics:

Ok, so Bioinformatics is a sort of a mix and match of everything. It is interdisciplinary which means it has within it Biology, Computer science, Information Engineering, Mathematics and statistics to analyze and interpret biological data.

• Advanced Quantitative Sciences: Quantitative scientists focus on biostatistics, statistical programming, pharmacometrics, data sciences and machine learning. They provide quantitative input to drug development strategies, from early development through life-cycle management
• Bioinformatics Communities: Pharmaceutical organizations often form Bioinformatics Communities. They work on sequence and structure analysis to guide engineering efforts, selection or optimization of candidates and data capture and structure to facilitate access and sharing within this communities

Drug Repurposing:

Drug Repurposing involves finding new therapeutic uses for existing drugs. This approach significantly reduce the time and cost associated with drug development since the safety profiles of these drugs are already well established. A good example is Hydroxychloroquine. Originally developed as an anti-malarial, it is commonly used for autoimmune diseases, was explored during COVID-19 as a potential treatment and recent research also suggest it may have potential in cancer treatment by disrupting cancer cells’ ability to recycle resources.

Now any accelerated drug discovery program that looks at potential candidates will not just look at new potential candidates but also look at existing drugs and their repositioning. We talked about DEL in my previous article. DEL has the potential to identify existing drugs as repositioned candidates. MELLODDY, the project we talked about last time also has the potential because it is also in the accelerated drug discovery domain

Predictive Analytics:

Predictive Analytics is core to pharmaceutical industry and is involved everywhere. It is there in drug development and Bioinformatics to analyze biological data such as genomic sequences and metabolic pathways to identify potential targets for new therapies. It is used in clinical trial optimization enabling exploratory data analysis, extracting knowledge and insights on disease compounds and patients. It is used in manufacturing process optimization to ensure higher efficiency and quality control.

But it is not just core sciences. We must not forget that a pharma organization is not just medicines. It is also finance and accounting and legal and compliance and marketing and mergers and what not. In Finance AI can be used for Financial Planning and Forecasting where predictive analytics plays a key role. There could also potentially be a a decision engine for the next best action designed to enhance customer engagement by leveraging advanced AI / ML techniques. It could analyze HCP demographics, sales data and survey data to create a unified view of each HCPs profile. And based on that it could recommend the most effective engagement strategies tailored to each HCP.

Manufacturing Process Optimization

Data Center Edge Optimization refers to the process of enhancing the performance, efficiency and reliability of data centers located at the “edge” of the network, closer to the data source. This involves improving data processing speeds and reducing latency for manufacturing workloads and can be done using AI powered engineering solutions. It also could leverage an augmented delivery model centred on AI infrastructure. AI can also be used for data processing and latency reduction, predictive maintenance and automation and optimization.

Market Analysis

I have already covered Next Best Action as a part of Predictive Analysis, but it is a part of Marketing and engagement with the HCPs. There could be solutions that use AI and HPC to enhance its revenue management solutions, providing advanced analytics and insights to support decision making process. There is predictive analysis, automation, insights and reporting, high performance data processing, the works. And finally AI Assistant Functionality aiming at enhancing the capabilities of AI assistants to provide more accurate and actionable insights. Using Predictive Analytics of course, but Natural Language Processing to understand and respond to user queries more effectively, improving the overall experience. And HPC providing the computational power to scale up data analysis processes, allowing the AI assistant to handle complex and large-scale datasets.

So there is a tremendous amount of work going on in in the Pharmaceutical Industry where AI and HPC is already being used. Not just in the Accelerated Drug Development space but also in other areas such as Finance, Manufacturing, Marketing, etc.

Now, in Part 4 of this series, I will talk about different types of HPC models that we covered in Part 1 and talk about what can be leveraged for doing what kind of business use case. Because while it is important to understand the business, we are not all scientists and can’t get into the nitty gritty of the core sciences. But a high-level understanding is required to be able to do our jobs well. So, I will try to build that bridge of understanding needed so we can support the business the best.

But I have a question for all of you. How much science do we (Technology and Infrastructure people) need to know to support the pharma business? Do we even need to know the science? Or do we need to just know what they are going to do with our systems? Like for example, is it data intensive workloads or parallel processing workloads? Is it an interactive workload or batch processing workload? Is it memory intensive, or high throughput (large number of small independent tasks that need to be processed concurrently)?

What are your thoughts? Let me know and let’s talk.