← Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
How AI and Accelerated Computing Reshape Manufacturing
00:42:53
Explore how Dell and NVIDIA drive manufacturing efficiency with AI, GPUs, and physical AI. Discover real-world strategies for future-ready factories.
Loading summary
Transcript33 lines
- [00:00]
A
Foreign.
- [00:05]
B
Welcome to Reshaping Workflows with Dell Pro Precision and Nvidia, where innovation meets real world impact in high performance computing.
- [00:20]
A
Welcome back to another exciting episode of Reshaping Workflows with Dell Pro Max with Nvidia RTX Gaming GPUs. I'm your host, Logan Mahler. Today I'm very excited for this episode. If you've been listening and following along for a while, we've done everything kind of from AI to ACO to media, entertainment, and believe it or not, we've even done oil and gas. But one topic we haven't dove into too much, it's kind of Dell Pro Max Nvidia RTX Pro GPUs role in accelerating, you know, manufacturing workflows. So for that conversation has been requested, I gotta thank no one better than Himachu Iyer from Nvidia who's with me today, who's my esteemed guest. And we'll be able to go through kind of all things, you know, manufacturing industry related and from the Nvidia side, kind of helping us tie it all together. So with that, Himachu, thank you for coming on today. You know, take about 30 seconds, introduce yourself, you know, your role at Nvidia, a little bit on your background and we'll jump right into it.
- [01:16]
B
Thanks, Logan. My name is Himanshu Iyer. I lead industry marketing and strategy for the manufacturing industry at Nvidia. I have been in this industry, manufacturing industry for over 20 years. My background is mechanical engineering and I have worked in this field in product design, engineering, simulations, additive manufacturing, conventional manufacturing over the last 20 years. So I have fairly good experience both on the software side and the hardware side of all of these different workflows in the manufacturing industry. And really looking forward to the conversation today and provide some suggestions, ideas, thoughts, inputs on how AI is reshaping workflows in the manufacturing industry.
- [02:02]
A
I love that. And it wouldn't be reshaping workflows if I didn't bring on a guest that can talk about how AI is. And just in general, the workflows are changing. So we're going to start with the first piece I think this became. And by no means am I a manufacturing expert. Like out of all the industry that Dell Pro Max supports, I probably know the least about manufacturing, even though I'm trying to learn from Amach himself. But you know, in the conversations we've had, I've learned that, you know, that the manufacturing process is kind of generally two things, right? Is pretty complex, right? And then it's also the more efficiency we can drive into that complex kind of, the easier and more streamlined things can be. So with that, let's start with kind of an easy question. Like let's talk about some of the, that multi step kind of process in manufacturing, like from R and D to design to production to documentation. You know, let's talk a little bit about that and how Nvidia and Dell Pro Max play a role in that kind of in the current state.
- [03:00]
B
Yeah, certainly. So, so like you said, right, Manufacturing is incredibly complex. It is a multi step process, right? It starts from requirements gathering, R and D, you know, conceptual design, detail design. That's where engineering, simulation, CAE plays a role. Of course, then the design needs to go to production. How is it going to be manufactured, assembled together and finally shipped to the customer. And all of the documentation, the technical documents that are necessary for maintenance, all of that need to be available at those right stages. But you know, the manufacturing is one of the oldest industries, right? We have been manufacturing products for a very, very long time. So over this time, you know, these designers, engineers, manufacturing experts, they have come up with processes that improves the process, makes it more efficient, it's agile just in time processes in order to, you know, make the whole process streamlined and eke out as many efficiencies as possible. But there are still plenty of opportunities to harness the power of accelerated computing, to harness the power of AI, to revolutionize, to reshape this, this industry, how products are designed, how products are manufactured, how products are maintained. And Dell, Promax, Nvidia, we can play a significant role in this reshaping of workflows by providing the right hardware, by providing the right software at each of these different stages in the manufacturing process to make it more efficient, improve the product quality and keep the costs down.
- [04:48]
A
I mean that it makes a lot of sense and you've touched on AI. But we're going to hold that thought for a quick second, right? And you're right. I mean there's a lot of, you know, I mean really, I think anything in life can always be improved upon, right? Like, and nothing is ever perfect. And I want you to touch on something that you talked about. Kind of the accelerated compute of, you know, Nvidia and kind of the underneath kind of cuda layer of that. And I'm going to remember a conversation that we had and I hopefully I'm getting it right or we're going to have to stop this recording and rerecord this part. But I remember you were telling me, I think Sean, your, your boss, Sean Young was there as well, and you were kind of talking about kind of the simulation piece was all very CPU based in kind of the past. Right. And you know, and then now it's moving to more of kind of a gp or not everyone, but it's starting to move kind of more towards GPU compute. So can we one kind of talk about that for a second before we get to AI? And then also, and like, I want to call that out because it's not just the efficiency within AI that is bringing even just traditional processes are being accelerated by Nvidia and Delpro Max and Nvidia RTX GPUs. But can you talk a little bit about that change from CPU to GPU processing for simulation?
- [05:58]
B
Absolutely. So you, you know, going back to the previous question, we discussed this process, right? How products are developed. Products are designed before they go to manufacturing, right. So it goes through this step by step process of requirements, gathering, R and D, conceptual design, detailed design, and then, you know, once the, once the design goes through reviews and approvals, it will go to manufacturing. But during this phase of conceptual design, to detail design, there is a very important step of simulations, engineering simulations. And what it does is the product is going to experience in its operating environment. It's going to experience some forces, pressures, temperatures, lot of these different physical processes. Right. So we need to make sure that the process or the product is able to withstand all of these external factors that I mentioned. Temperatures, pressures, vibrations, all of that. Right. And perform the way it is supposed to perform without any failure. Right. So that is what engineers simulate digitally on a computer before the product is manufactured. Right. They want to capture the performance, the behavior early on in the design cycle. Now, this process has traditionally been done on CPUs. And it takes a long time to do these simulations. It can take hours, it can take, in some cases, it can take weeks. But what has happened over the last five to ten years is a lot of these ISVs or independent software vendors who develop these applications where you can do these simulations, they have ported these solvers to GPUs. So the solvers that were entirely running on CPUs previously are now doing these computations on GPUs. And the main difference between GPU calculation and CPU calculation is on GPUs. There are thousands and thousands of cores that are doing these simulations in parallel. So the GPU solvers are much, much faster as compared to CPU solvers. So previously the simulation that would have taken hours or days is 10x20x or even 50-100x faster, depending on the type of application and what is being simulated. So in the industry we see this transition going from traditional CPU based computation to accelerated computation on GPUs where the engineers are able to run these simulations much, much faster. And that enables to run them more simulations so they can come up with an optimized design, they can run more iterations. So the final product that is designed is, you know, fully optimized on lot of different parameters. It has, you know, minimize the weight, maximize the performance, even costs and manufacturing processes can be all simulated to, to, to make sure that you're getting the best possible product.
- [09:05]
A
Makes total sense. And yeah, I'm glad my memory served because I would have felt like a crazy person if I got that story wrong. But I remember you saying that. And I think, you know, the thing to realize is that, you know, whether it be delpromax or you know, Nvidia, we're not just necessarily looking at, I mean AI is important, which we're going to get to in the next question. But even those, you know, moving, you know, off for simulation, for isv is moving from CPU to GPU compute, doing it faster, all of those things are still improving. Right. But let's talk about AI, right?
- [09:33]
B
So let me add one more point to that, right? One more change or transition that is happening in the industry going from traditional compute to accelerated computer is that a lot of times these simulations needed very high end or server or data center type of an infrastructure to run. Because these were such large simulations previously they could not run on workstations like the Dell Promax systems. Right. But now with many of these simulations running on GPUs and the GPUs, the memory available in GPUs growing with every generation. A lot of these simulations that were previously either cloud based or data center server based can now be run on workstations like the Dell Promax Towers, which can incorporate multiple Nvidia GPUs. You can run these simulations on these workstations which the engineers can have at their desks rather than having to wait for expensive cloud computer or data center type of systems to be able to run those simulations.
- [10:45]
A
Exactly. I mean, I think that that's the thing that I've noticed is that, you know, I haven't been in this role for a long time. But you know, workstation GPU compute, I mean the, you know, the RTX Pro Blackwell, you know, Workstation Edition 96 gigs, like it's not an H100, right? Because H100 is more, you know, tensor based, AI based where more Cuda cores and the RTX Pro 6000. But in terms of just VRAM getting pretty close, right? Like you know, in terms of that. So I mean you hit the nail on the head is that it's not only from the location that the computation is happening, but from you know, CPU versus GPU offload, but also the physical occasion of the system that is being used. And I think that kind of ties into the next point with, with AI, right. So I obviously taking things by storm, we don't need to rehash that because we have, but kind of have a question for you is that, you know, I wouldn't say more so but manufacturing firms, when you know you're doing these simulations, you have tons and tons of data, right? And you have a team that. And hypothetically if you're building a car that is testing the braking function versus hey, the side crash impact, you know, all testing this. But much like with any company, many times, you know, different teams, different business units, all kind of work, you know, unfortunately somewhat siloed way. And the fuel that really powers, you know, AI or artificial intelligence or machine learning is really data. So maybe talk a little bit about one kind of what is the benefit of from a manufacturing standpoint, consolidating those data sets through AI and then how can we use agents to transform like the imprecise or unstructured processes into something actionable from, you know, the consolidation of that data?
- [12:29]
B
Great question. I mean the very foundation of AI is data, right? And the many of the manufacturing companies, you know, have data going back three, four decades, maybe even more. Right? But all of these data coming from different systems. It could be a PLM system, it could be an ERP system, it could be a manufacturing execution system, right? Manufacturing companies have it ot and these type of data that's in these systems may not necessarily be talking to each other. It may not be compatible, it may be in different formats, right? For manufacturing companies to be able to, you know, benefit from all of this data, there, there needs to be some data governance policy. There need to be some way of linking all of these data together so that they can then build on top of that, right? They can derive intelligence out of all of that data rather than having to you work with each and every system individually, right. One thing that is becoming very common in the manufacturing industry, right. We have heard about digital transformation, but underlying the digital transformation is a digital thread which is connecting all of these process that we talked about previously, right? And this digital thread can really be transformed because of AI being able to connect these different teams, these different silos.
- [13:54]
A
Right.
- [13:54]
B
And that is what is going to empower these manufacturing companies to benefit the most from AI. And we have seen this journey within AI. Right. We started with perceptive AI systems, then we went to generative AI. Now we are somewhere, you know, agentic AIs, where teams can build AI agents that are autonomous to a large extent. They are able to string together a workflow with minimal inputs from the designer or the engineer. And that is where we see benefits in different phases of the production or the manufacturing cycle that we talked about.
- [14:36]
A
Right.
- [14:36]
B
It could be an agent that helps the design engineers in coming up with many different design iterations. It could be an agent that helps a simulation engineer check many different types of behaviors for these different physics that need to be simulated. So these agentic AIs are assisting these designers and engineers at the stage that they are working on for this manufacturing process.
- [15:08]
A
Yeah, I mean, that makes a lot of sense. I mean, we've talked about agentic AI on the show, right? Like, and that's kind of, I think we're. I wouldn't say we're at the beginning of it. We're not at the end of that phase. Right. I think a lot of the generative phase is kind of gone. Like, that was kind of very early, but like with the agentic phase. Right. I mean, that's literally just for those that haven't listened to an episode before. It is think of a workflow kind of from A to B where you're putting kind of bounds on it at the end of the day, where AI within the certain confines of what is structured can go out and do and complete tasks. Right. Or that we talked about the simulation piece. Right. A hypothetical that would be, hey, you know, in Agentica, it's going to test this. And then because it's let's say out of bounds because, you know, okay, let's say the part failed, then it would adjust and then you can kind of automate this workflow and kind of automate the process. But what you're kind of talking about is very experimentation, I think that is, especially when it comes to Pro Max and RTX Pro gpos, is that that's where it really thrives. Right. Like the experimentation, having the resources at your desk or at your fingertips. You're not waiting on that. You're not paying a cloud bill, you're not waiting for a data center resource. And you know, you're allowed to iterate and type or, you know, prototype very quickly. But let's talk about like practical deployment, right? As people in companies are starting to do a growing investment in AI, they're really starting to look at more of, hey, this is not an experiment anymore. This is not something that we're thinking about doing. Like, this is something we're going to deploy across our company. So kind of two questions for you is what kind of is the key? Like for example, when you're trying to establish a firm foundation for the future with AI infrastructure, what should you think about? How do you do that and what are companies, whenever you're out talking to engineering companies, which you do all the time, what are the big points that they're concerned about when it comes to that scalable deployment?
- [17:03]
B
Manufacturing industry is adopting AI, no doubt about it, right? Every industry is adopting it at a different scale, different speed. But manufacturing industry no doubt is adopting AI in many of its workflows, use cases. We have seen POCs taking place for a long time now, but we are at a time in the industry where a lot of these PoCs are scaling up to become an integral part of workflows. You know, one of the key areas where we have seen this taking a significant foothold, becoming almost, you know, daily kind of a process is in quality and inspection, right? IoT has been in the industry for a long time now. IoT technology has matured. So we are getting lot of data through cameras, through different sensors and such, right? So, you know, going back to our previous point about agentic AI, right? There are, you know, very good examples of these agentic AIs monitoring all of these different streams, all of the Data coming from IoT sensors and giving real time inputs to the technicians, to the folks on the factory floor. If there are any quality issues, there is a real time alert that is being sent about whatever the problem is. So predicting anomalies and triggering resolutions before the defect actually makes it further, right? So these are the types of agentic AIs that are taking hold in the manufacturing process. But going back to your original initial question about how, you know, what type of difficulties are there in scaling these processes, right? So it's twofold, right? First of all, there needs to be a team that comes together to be able to implement these AI workflows in their work, in their use cases, right? There needs to be management backing, there needs to be some key KPIs defined that will decide outcome of what is successful implementation, what is not. And then once this basic framework is put together, the architecture or the platform that is going to support these workflows from a hardware and software point of view is equally important. A lot of these AI workflows really need a lot of compute capabilities. And that is where having the right compute capabilities in terms of the systems, in terms of GPUs, becomes really important. Right. All these projects are going to start as a POC on a smaller scale. So you quickly need to be able to iterate on this, need to come up with that actual workflow using the hardware and the software. And once you are able to prove its efficiency, its success, then you are going to scale it right to your broader organization. So making sure that you have the right resources, the training that will be needed to implement this on a broader scale, building the right skills in your team, in your group, all of these become really important to make sure that you are able to scale AI. Right. And finally, I would say make sure that you have that backing from the management because this can take some time to really show the success from the effort. So, you know, kind of multiple factors that are needed to be able to start from a PoC and scale your
- [20:29]
A
AI projects makes sense. And you, you made a really interesting comment and I would love to get your perspective on this because I know that you're out at events in the field, you know your space really well. You kind of mentioned almost like a, you know, in in tandem with the management backing for any AI initiative, whether that be financial or for, you know, support of launching this initiatives or buying or whatever. You kind of mentioned like a center of excellence. Right. And I would love to hear your perspective because I've, I've always hear different mixed opinions on this and I'd love to hear what you say related to kind of manufacturing. But I see kind of companies do one of two things. Kind of a center of excellence that you kind of described, right? Where you have the right stakeholders from the right groups that are making the decision on, hey, what is the tech, what is the software layer, what, you know, all of that kind of standardizing it and like rolling it out or it being kind of the other end, where it's completely decentralized, where it's, hey, marketing's going to use this and you know, another group's going to use this. So like in your opinion, I guess what are kind of the pros and cons to both of those approaches? And for someone listening that, you know, is in manufacturing or considering it, which one would you recommend?
- [21:40]
B
I think experimentation is going to happen, right. The world of AI right now is, is so dynamic, right? We are getting these new models, LLMs, new type of methodologies coming out almost every month, right. We have, you know, every two weeks or so there is some, some major new model or an update to a model that's coming. So people are excited about the technology and they are always going to, you know, experiment to some extent on their own. And which is fine. Which gives the team exposure to these different methodologies, different technologies, right? But then if that keeps happening, if that becomes the primary nature of how you're adopting AI or an organization is adopting AI, I think it can lead to the same problems that were there previously, right? A siloed approach. Every team is using different tools, different framework for trying to do what they want to do, right? And I think this is where forming a core team, a center of excellence, let's say, will make sure that this siloed approach doesn't take hold. Right. To benefit from AI, it needs to be. Adopting a certain LLM by itself is not going to give you the results that you are after, Right. Training the LLM on your own data, right. Like we discussed, manufacturing companies have data going 30, 40 or more years, right? So being able to train these LLMs on your data, providing those guardrails and then, you know, making sure that the whole team contributes to this. And there is no siloed approach. There is, you know, going back to a previous point, there is that digital thread that is connecting all of these different, you know, approaches, different models that the organization might be using. I think that is going to provide the most benefit. Rather than, you know, every team experimenting on its own and doing things on their own, having that centralized approach from a software point of view, from a hardware point of view, I think that is going to give the most benefits to the organization.
- [23:52]
A
I agree wholeheartedly. I mean, and I would also kind of add this, is that with a center of excellence, you know, centralized approach, it will probably be a tad bit slower, right. Just from the fact that you have multiple personalities, points of view, people from different organizations, right. So it'll come a little bit slower to come to a time and alignment on what some of these things, right. Would be a decentralized approach. Yeah. You can operate very quickly, right. Like, you know, it's very quick, you can make decisions fast, kind of autonomous. But then the trade off of that is that downstream, kind of like digital thread that you talked about is then you're, you're back to kind of square one, right? You have your, your silo data, you know, you have been from a cost perspective. You know, I've seen situations where two different parts of the business had competing tool sets and like, just crazy, right? But yeah, no, I agree. The COE approaches is 100% the way to go. And kind of the next topic is, I think what I'm most excited about. I remember I was not lucky enough to make Jensen's keynote. I was at GTC this year, but I did not make it to the venue. The line was too long and I ended up have to set up demos at the booth. But I remember very, very vividly there was a slide, right, that he kind of showed of like where we're at on this continuum of AI, right? And you know, we were kind of in. I don't remember. It was like more the. The perception. Then there was like the generative, right? Think of comfy ui, stable diffusion, all that. Then it was a move into the generative and then to the far right was physical AI. So I think where we start first, before we jump into physical AI with you is let's take a step back and say like physical AI, can you def. First off, can you kind of define it and talk about how does kind of core pieces of manufacturing, you know, like robotics or AI or digital twins, how does that kind of fit into the concept of physical AI?
- [25:47]
B
Yeah, what you talked about, right? The journey, the various milestones in AI, right. We started with perceptive AI, generative AI, where, you know, these LLMs model can generate text or images or even videos based on the prompts that you enter. Right? And then to agentic AI, where a lot of the workflows become almost autonomous because these agents, these AI agents are able to string together different pieces of that workflow. And now moving on to physical AI, which we see as the next step. So where we are right now, all of these LLMs and such, they don't necessarily fully understand the laws of physics. They are mainly based on texts and videos and images. That's what they have been trained on. And it's very hard to communicate or integrate the laws of physics into the responses that we will get. That is where we see the journey taking us. Right? These new models will be physically aware. They will know the laws of physics, you know, temperature and friction and mass and velocity. The model will be aware of these. Before answering the question that, that you have asked these models, Nvidia announced one such World foundation model, Cosmos, which has these laws of physics built into it. So when, let's say a robot is going to be working on an assembly line or on a welding line, right? These robots need to go through hundreds or thousands of hours of training to be able to do that step accurately on the factory floor. Now, lot of this training needs to happen in such a way that it is quite expensive to train those robots, right? There isn't enough data available to train those robots. So where we see physical AI playing a role is lot of this data that is needed for training these robots will be generated synthetically in these World foundation models, right? It could be training a robot, it could be training an autonomous vehicle, it could be training a pick and place type of a device, right? So all of these devices, all of these physical agents that will be working in these spaces will be trained, trained digitally in these World foundation models which are aware of physics, which, which take into account the laws of physics. And that will make all of these processes, all of these systems a lot more easier to simulate and train before they are deployed in the, in the physical world.
- [28:33]
A
That makes a lot of sense. That was a great explanation because I've heard physical AI in and I've heard people talk about it. I really like how you broke it down. I mean it is the physical world, but it's really physics AI, right? You know, and I, I love the example about train rainbow lots. And I played with Cosmos model as well. Like by no means is it great. Some of the stuff that you know, I've done, I haven't like trained a robot for synthetic data. But you know, going back to the car driving example, right, like for example, you know what, whoever it is, Uber Waymo, whoever, whoever's doing the self driving cars is, you're right, you have to have a ton of data to be able to say, hey, that's a stop sign, hey, that's a trash can, not a dog. You know what I'm saying? You have to have all of that. I guess the question two, two part question of this is in like let's say, you know, the foundation model, Nvidia Cosmos, how are you actually imparting the physics into it? I guess like how are you teaching the model physics? And then the next part of that is how can you test the physics accuracy before training, you know, the computer vision model that you're going to put on a self driving car.
- [29:42]
B
So this goes back to, you know, one of our previous points of discussion. We talked about simulations, right? Engineering simulations, right? So in these physics accurate models, it's essentially doing these simulations in the background. Let's say you have a tabletop and an object is rolling on the Tabletop. We all know, you know, naturally, inherently, that if it comes close to the edge because of gravity, it's going to, you know, fall down and come in contact with whatever it's a floor or whatever it is, right? But an LLM doesn't necessarily know that. It needs to be taught what happens if the object falls on a hard surface, Right? It might bounce a couple of times before coming to rest. If it bounces on soft surface, it might not bounce, right? Now imagine what happens in a warehouse, right? A robot is picking up something from a conveyor belt and dropping it into a container which will essentially be shipped out, right? The robot needs to be trained on how to drop the object in that container, right? Is it a hard container? Is it a soft container that will decide the motion of the robot? Right. And the robot will only know that based on the training data or the training material that has been provided to it. So that is what we do beforehand. So in that physics AI model, we teach it about gravity, we teach it about, you know, contact forces and friction and all of that. So. So there is a physics engine, essentially, that is powering all of the physical interactions that are taking place. Let's say that in a manufacturing facility, you have an existing machine that is going to be replaced by a new machine. But surrounding this machine, there are going to be, you know, hundreds of other components and piping, wiring and all of that equipment already present, right? So now when you bring in the new machine, there, the first thing that you need to make sure that there is, you know, there is no collision. There is no. With the existing objects that there are. So again, what is collision? It is, you know, two objects colliding with each other or they are trying to occupy the same space, Right. If you create a simple CAD model, right. It does not understand this. You can model two components right over each other and the software will not complain. But physically, it cannot happen, right? That physics model, that AI model needs to be smart enough to tell you that you cannot place these two objects at the same location because there will be a collision. And when you try to go into the real manufacturing facility to, to install that, you're going to have problems. So you need to be able to simulate all of that before you do anything physically. And that is what these physics simulations inside the models will enable us to do very soon.
- [32:50]
A
I love that explanation. It is almost, and this is a bad analogy, but within something like Nvidia Cosmos, it's almost like a physics engine powering that for it to know exactly like you said when you were giving the example about the ball and rolling and say, hey, well, you know, if it's at this angle, it's going to roll faster or slower. If it's, you know, something that's like a glass ball, it could break, but if it lands on carpet, it might not. Which I think that's so interesting and fascinating. I mean, I can see that it probably makes a lot of sense because you can run a lot of simulations. But outside of, you know, preparing some really good, you know, training data that you can use internally to train, you know, and fine tune or distill other models, what are some of the other, you know, advantage. Well, not say advantages, but the benefits of physical AI that you're starting to see manufacturing companies start to realize as they move towards it.
- [33:41]
B
We discussed it already. But I think robotics is one of the key areas, right. In the manufacturing industry we know some of the challenges, right? A lot of the workforce is retiring and new talent, engineers, designers, they are not coming in at the same rate. So there is a lot of experience that is being lost because of this retiring workforce. Right. But at the same time we see that there is a lot of re industrialization that is happening, right? Many new manufacturing facilities are being set up. So it is going to need some kind of, you know, technology that assists the workers that are going to be there. And that is where I think robotics is going to play a big role over the next few years. And these robots, like we discussed, they need to be trained to be able to do these different operations or different steps. So training them with synthetic data is really going to be important. Training them in different environments is going to be really important. Right. If we go take an example of autonomous vehicles, right? They need millions of hours of training, an autonomous vehicle, millions of hours of training. And doing that physically is impossible. So generating all the different lighting scenarios during the day, in different weathers, in summer, in winter, when there is snow, when it's raining, doing all of this synthetically, doing all of this virtually in these physics AI model is the only way we can move forward. There is no way that we can do all of this testing or simulations physically.
- [35:23]
A
I mean, you're right, like it would be to collect the data, it would actually take longer until the, the data was outdated. Right. Like, I mean, it makes perfect sense. And I could see how that is becoming specifically in manufacturing. Right. And I know Machu, we've been going for a while and we're going to start kind of winding down the episode. But probably the last Question I'm going to ask is, you know, we've kind of talked through, you know, an arc in this episode. Episode a little bit, right? We've talked about, you know, the complexity of the manufacturing. We've talked about how AI can unify and how to bring down some of the silos so we can consolidate that data to give us real opportunities with AI, how to scale AI. Talking about how accelerated computer through Nvidia, RTX Pro and Cuda and other kind of computational layers is helping AI advance. We got all into physical AI, but the last that I want to leave people with because I think it's kind of the most important, right? And there's folks that are going to be listening to this podcast and watching this, you know, on YouTube that may be very advanced and there's going to be some people that are, you know, novices, right? And I want you to take a second and kind of talk to the novices. You know, we had a good talk before this call about kind of, there was, I think, five points that you had said is like, hey, when it comes to preparing your organization for AI, there's kind of five points that you've seen. Manufacturing companies who have either done these and part and parcel have seen more success. So could you go through those five points and kind of give a quick summary on, you know, what people really need to think about when bringing AI into a manufacturing process?
- [36:54]
B
It starts with data, right? Make sure that the data you have is good quality data. It is a data, you know, that has been kind of scrutinized closely because it goes back, you know, garbage in, garbage out. Everybody knows that, right? So if your AI models are trained with bad knowledge, bad information, bad data, the outcomes are not going to be accurate. Second thing I would say is LLMs is something that, you know, we all know, we have all heard about over the last three or four years, right? But now where we are in the AI journey, there are a lot of small language models that are also becoming quite useful, right? And these small language models are much more easier to fine tune with your own data and these can be tailored more specifically to your workflows rather than having to train a very massive model which can be very, very resource intensive, compute intensive. Take a look at these domain specific small language models which are easier to work with, easier to fine tune and easier to deploy for your particular use cases, right? And then look at the infrastructure layer that you have, because like we discussed, a lot of these AI models are compute intensive. So make sure that you have the Right. Infrastructure to start small and then be able to scale as your projects grow in size. Right. So you can start small with simple workstations, desktops, laptops that have the right GPUs. Another product Dell Nvidia has is called the GB10, which is an AI development and inferencing type of device that can sit on your desktop and help you with many of these workflows. And then you can seamlessly scale to, to more and more powerful hardware as your projects are growing. And one, finally, what I would say is make sure that you're using the right combination of hardware and software for your AI projects. And all of this goes back to our discussion about center of Excellence. Right. There is a core team that is strategizing and formulating the AI workflows. AI transition in your organization with Powered by AI.
- [39:31]
A
I love that. And yeah, you can't have a conversation without talking about GV10, which I'm really, I mean, I'm really excited about. I've been hearing about it for six months now, and we're, you know, on the precipice, at least at the time of recording of this launch. So, you know, very excited about what that can do. So how much I really appreciate the time, do me a favor, let's go ahead and wrap the episode. Tell everyone where they can kind of find you on LinkedIn and then, you know, give the URL. We'll also connect to it in the description. But anywhere in Nvidia you'd like the guests to check out, whether it's, you know, the developer portal or whatever you'd like, let everyone know where to find you and what they should be checking out at Nvidia.
- [40:08]
B
Feel free to connect on LinkedIn with me. Search by my name. I don't think there are too many profiles by my name.
- [40:14]
A
Probably not, probably not.
- [40:17]
B
So, so that helps. And you know, some of the discussions that we have had, right. In terms of hardware, in terms of software, in terms of workflows, there are many free tools that Nvidia makes available. Go to build.Nvidia.com you will see many, you know, free tools that you can download there to play with. Dell Pro Max has a very good collection of resources. You know, use cases, case studies, solution overviews. I think we'll be able to provide those links in this podcast. Right? So make sure access those links and feel free to reach out to Logan and I if you have, you know, specific questions about how AI can help in your manufacturing workflows.
- [41:03]
A
I love that. Well, you really appreciate you coming on. You are a very giving and person with your time and oh, it's been a great resource. So if you're listening, I would definitely reach out to Machu. He's a wealth of knowledge. He's probably forgotten more about engineering than most of us ever hope to learn. So, and I don't say that physician. I mean it's true. He does know everything and I've, I've learned a lot from him. So, you know, in closing, you know, this, like I said, engineering, obviously a lot has changed from, you know, that over the last several years, right from the traditional compute of moving, you know, ISV simulation software from CPU to GPU to more, you know, to be able to do that at the desk side because the amount of CUDA cores that are now available in GPUs, right to the advent of AI, whether that be kind of the generative we talked about agentic all the way to physical AI and being able to talk about the physics of models and the example of, you know, the ball rolling off the table. What happens is it break? Does it bounce? Does it continue to roll? So, you know, this is, I absolutely love this episode. And with that I'll quit talking and we'll wrap it up. So this is Logan with reshaping workflows with Dell Pro Max and Nvidia RTX GPUs. Until the next time, make sure you keep your engineering workflows, whether those be simulations or AI agentic physical AI running locally on your Del Pro Max and your video RTX gpu. And we'll see you on the next one. Do what you want. Do what you want. This podcast was produced in partnership with Amaze Media Labs.