Taming the AI Hardware Cambrian Explosion — Abdul Dakkak, Modular

0:01 [music] Uh so let's get started. Uh hi, my name is Abdul. Uh uh I'm from modular. Uh today I'm going to be talking about what modular is and

0:23 what value it brings to you. So as you probably know today there is a huge crunch uh for compute and there's like no alleviation in sight. A lot of these things are booked for the next few

0:37 years. But there's also a kind of as a result you have an explosion of hardware. We call it like a Cambrian explosion. You have Nvidia providing you know a few

0:47 generations of GPUs, the Intels, the ARMs, the Apples, everything uh is is present. They all have some niches uh that they cover. But the problem is that you have to ask is how do you program

1:01 all of these uh systems. Now if you look at the current software as it is today um it's kind of a mess as well and that's because a lot of these hardware choices bubble up from the you

1:17 know from down under uh into upwards into the stack. So this is a diagram showing how uh VLM supports both Nvidia and AMD. If you add extra hardware the the the

1:32 diagram becomes even more difficult. Now today folks actually uh deploy not only using VLM but also SG lang TRTLM uh Llama CPP etc. So at Modra we thought well this is kind

1:47 of a huge mess. Let's rearchitect the uh the stack from the ground up and let's build a unifying layer uh that unifies all the hardware so everything runs on our stack.

2:00 To realize this, you have to actually go back to first principles and build a complete software stack with portability, performance, and programmability in mind. So at the

2:13 bottommost level, we've developed Mojo, a heterogeneous programming language that's targeted towards AI as well as general purpose programming uh as well. It's kind of pythonic in style. So folks

2:26 that actually know Python today pretty much know 90% of Mojo. Uh so we've built that language and that targets both CPUs, GPUs as well as all our accelerators. On top of that, we

2:39 leverage Mojo to build max our inference engine. Uh so that provides you with uh AI inference for both for a bunch of different modalities like image, video, speech as well as text. We've built all

2:54 our kernels underneath the hood that are built on top of Mojo. Finally, today's AI inference is distributed. So you actually need a cloud solution as well. uh you need to

3:06 pick which hardware to deploy your model on, how to route uh different user requests to different uh machines, how to do KV cache offloading and and and distributed inference and disagreed

3:19 inference uh to be able to serve the model at the speed and the cost that that folks expect. So putting it all together, we built a modular platform that's fully

3:30 programmable. It's hardware abstracted. It's it's soda perf. Uh and then it actually can be deployed both on prem as well as on our cloud as well. So it can be if you have a cloud uh uh setup, you

3:45 can actually download the modular cloud and and deployed there. So let me kind of dig into each one at a time, each part of the stack. So the first is Mojo.

3:56 Uh so Mojo as I mentioned is a Pythonic system programming language. It's not unlike C++, Rust or CUDA C++. The thing is what it enables is uh programmability across different accelerators. It's

4:11 blazing fast. It's again it actually competes with CUDA in terms of performance. uh and you can think of it as the best way to program CPUs and GPUs within a Python-like programming

4:24 language. In terms of how we think about Mojo, uh you know, you can think about there's three different quadrants, usability, performance, and portability. Rockcom

4:37 and KOD and CUDA gives you massive performance, but they sacrifice on both usability and programmability. A lot of folks use agents to actually program these system nowadays.

4:48 Open CL back in you know a few years ago was popular. It gives you great portability but is sacrificed on performance and usability. Triton is has a good sweet spot that gives you 70% of

5:00 the Perf uh but it gives you portability and usability and Mojo is is trying to kind of fill the niche of the sweet spot where it's the best of both worlds. It gives you

5:12 the abstraction to allow you to write portable kernels and and uh and usable kernels while giving you also the low-level access uh to get the maximum performance.

5:23 We've actually open sourced uh practically all of our kernels so you can actually test this stuff. You can benchmark it on your own machine. Uh we have state-of-the-art kernels on Nvidia,

5:34 AMD and other hardware. So to tell you what state-of-the-art is, it means that our matrix multiplication is faster than the coupe losses, the hip losses, everything out there. And it's not by 1%

5:47 in some cases by 2%. We have the fastest attention uh kernel out there, faster than flash attention for uh as well as like the AI it and things like that. Uh one thing that we've announced uh a

6:01 month ago or so is Mojo 1.0, know the stabilization of our APIs. Uh, and we're going to be releasing Mojo 1.0 in the next few months. And we're also open sourcing Mojo uh, in the next few months

6:14 as well. Sitting on top of Mojo is Max. So, Max is trying to kind of solve a few of the uh, assumptions that were made uh, back in the day for PyTorch. PyTorch

6:29 is about 10 years old. It was not really designed for the genai inference of today. As a result, you have like VLM and SG lang that stepped into the void to give you things like KV cache, page

6:41 detention and and things like that. And but all of that kind of kind of complicated the entire PyTorch ecosystem. A lot of folks don't really uh use PyTorch as is. They're resorting

6:55 to these layers on top. So think of a new approach and we call that the max framework. So again it's built from the ground up uh with the best technologies with the best

7:07 state-of-the-art graph compilers uh the eager mode execution. It's portable across devices. It understands your mojo programs to do novel fusion uh and it gives you the maximum performance and

7:20 deployability that you're otherwise not able to do in in in regular PyTorch. So compared to other stacks uh this is a single stack that allows you to serve develop the model as well as optimize

7:34 everything in one cohesive uh system. it within Max we support thousands and thousands of models in fact thousands of architectures uh everything from Gemma 4 to M uh Miniax 3 to GLM uh so any model

7:55 that you think about is is practically supported in uh in in uh in Max and it has the same interface that you'd expect uh from what is standard today. So we also provide a open AI endpoints that

8:09 you can use uh within the stack. If you're familiar with like PyTorch programming uh the figure on the right would or the left will not be like massive departure of your mental model.

8:25 Uh you can build a model as you expect. You know you can import modules. The uh programming language for developing models is Python. So as familiar to folks uh and you can compose uh things

8:38 as as you as you as you um as you build them. We also support like the eager execution to enable debugging as well. Now the other thing that one has to realize is modern LLMs uh are run on

8:56 across different GPUs and in many cases multiple nodes. So also max the max APIs provide you with a way to uh program multiGPU system AC as well as multi-node systems from within the API so that the

9:12 model is natively supported across uh different GPUs and different nodes as well. And this can also be heterogeneous. So there's nothing to stop you from saying the prefill is on

9:24 hardware X and the decode is on hardware Y. Finally, layered on top of all of this is the modular cloud. So, as I mentioned, uh today's AI inference is

9:39 distributed. It runs on many many different machines. Uh it's heterogeneous. So we provide the module cloud which is a performance native you know we think

9:50 of it as a glass box because you can actually tinker with all the variables uh that are available for you in both max mojo and everything in between uh within your endpoints. So it provides

10:03 provides endpoint deployment as well as bring your own cloud deployments as well. We also have a module cloud lab that allows you to experiment with all these different parameters and pick the

10:14 best point in the perto frontier that you have for that model or based on your workload as well. Uh so in terms of like the UI uh if you're familiar with like any uh cloud

10:28 provider I mean it's it should be pretty familiar. You just go to the website, you click on which model to you want to run, you get an AP an API key and then you plug that in uh to your HTTP request

10:42 and uh you can run it. You can have dedicated endpoints as well as we can serve your own custom model. Uh you know in many cases folks have fine-tuned models or models with their own custom

10:54 speculative decoders. Uh we could do all of that uh for you as well. Now, why do you care about all of this, right? It's a lot of tech. Uh why do folks care? Well, it's because you can

11:07 actually uh accomplish more things uh through the system, right? So, you can actually decrease cost. You can get greater flexibility because you can actually use different hardware. You're

11:21 not beholden to one hardware provider. Uh you get portability. You can use multiple different models. Uh etc. So how does Perf look like? Well, we benchmarked a lot of these uh and this

11:34 is like against uh state-of-the-art, you know, the best combination of VLM or SG lang [snorts] in a bunch of cases. These are these benchmarks are independently verified as well. So in terms of Kimmy

11:46 Kimmy K2.5, we're 5.5 faster on TTFT uh compared to VLM. We can also get 1.5x throughput improvements. So the token per minutes

12:00 on per GPU is 1.5x faster uh on B200 than it is on VLM. In terms of Flux 2, and I encourage folks to actually visit the booth behind this curtain, uh there's a nice demo of

12:19 Flux 2. uh we're about 7x faster uh compared to ters that compile with diffusers. So again, Max supports different modalities within the same framework. So you don't have to use an

12:31 omni variant of a of a piece of software or anything like that. We support both video generation, audio generation, image generation as well as text generation.

12:42 Uh we also support you know things on AMD where you can get a TCO benefit as well where we're 4x faster than the equivalent uh on um on on the MI355 systems. Actually you know some of this

12:58 is not a fair comparison because you know some folks might say oh 7x is much larger than 4x but this is you cannot do that comparison very easily. uh that you can you have to factor in the cost uh

13:13 per uh GPU and AMD is usually 30% cheaper than uh the Nvidia equivalent. Uh artificial analysis actually we've been partnering with them uh and we've been and they've independently been

13:28 verifying our results. Uh this is our results like as of uh today. Uh we're about 3x faster than the next competitor uh on the flux 2 dev model as well. And this is not playing tricks of

13:44 distillation and you know quantization or anything like that. This is just raw speed because we can leverage the hardware from the ground up and make sure that uh we're getting the max uh

13:55 the max PF benefit from the model and you can actually go to artificial analysis today and see and see these results. Now what does that mean to folks again?

14:08 It means you can open up applications that are otherwise kind of not possible because of slow inference. So this is actually uh a kind of a live demo. It's just somebody's pasting in the prompts.

14:22 Uh but the generation is is accurate. So you can generate things in subsec generate images in subsecond uh per in many cases these are.3 seconds. Again you can go to the booth. There's a nice

14:36 demo where uh you can give a prompt and then it reads the video feed and then renders it as well. Uh the other one data point is Gemma 4. So Gemma 4 was a a recent model that was

14:53 just introduced about a month ago, a month and a half ago. Uh again we're quite faster than other any other competition out there. Uh so uh you know many cases actually some of these are uh

15:09 hardware companies. you know, Samanova is a hardware accelerator. Uh we're, you know, let's say 6x faster than that. And in general, we're much much faster than others. And by the way, we're for all

15:22 these cases, we're just starting. There's more optimizations that we're landing uh as as as time goes on. So, we're never never stopping to improving these things.

15:34 So, we have a bunch of uh places where you can get involved. You can join our open source community. So, a lot of our code is open source. We have over 700,000 lines of kernel code that we've

15:45 open sourced. We've over open open sourced all of our model definitions, all of our max serve. Um, we have AI skills. So, if you're not familiar with Mojo or Max, you can actually use the AI

16:00 skills to port uh from languages that you're familiar with into Mojo or Max. You can also join us. We're hiring. So, there's pamphlets on the booth where we're hiring for a variety of different

16:13 positions. So, as I you probably know, we we hire from low-level kernel engineers to high level folks that work on Kubernetes and everything in between. You can actually try all of this. So,

16:26 this is not uh demoare or slideware. You can pip install modular right now if you want to and try all of this stuff. And I welcome everybody to validate these results.

16:38 So, with that, I kind of that's it. And uh but I'm open to questions if folks have it. And I'll also be on the booth uh for the next few hours as well. [music]