Discussion: From Cuda to tile-based programming: Nvidia's Stephen Jones on shaping the future of AI

As AI models grow in complexity and hardware components to meet demand, the software layer that connects the two must also adapt. We recently sat down with Stephen Jones, a respected engineer at Nvidia and one of the original designers of the Cuda.

Jones, who was behind the mechanical origins of Aerospace Mechanics, gave a deep insight into NVIDIA's Software Testrations, including the transformation of tile-based systems, the introduction of “how AI rewrites the rules of code development.

Here are some key takeaways from our discussion.

Replacement of tile-based output

For years, the Cuda system has been largely responsible for managing grids, blocks, and threads. With the latest updates, NVIDIA presents a higher level of release: Cuda tile.

According to Jones, this new approach allows developers to directly edit arrows and wishes rather than managing individual threads. “It carries over the existing cuda,” Jones explained. “What we have done is to add a way to talk to the program directly, the strings, the data arrows … To allow the language and the compiler to see what works in every area of ​​the new construction”.

This modification is a response to the rapid evolution of hardware. As tensor cores become larger and reduce the resistance to more's law reduction, the mapping of the code in silico is becoming more and more advanced.

• Future proofs: Jones noted that by expressing programs as vector functions (eg
• Strength: This ensures that the design of the system remains stable even in the change of basic GPU architectures from ampere to hopper to blackwell.

Python first, but not only python

Realizing that Python has become the lingua franca of artificial intelligence, Nvidia introduced cuda tile support in Python first. “Python is an AI language,” Jones said, adding that structured representations are “very natural for Python programmers” who are used to profit.

However, functional cleaners need not worry. C++ support is coming next year, in keeping with Nvidia's philosophy that developers should be able to speed up their code regardless of their language of choice.

“Green conditions” and reduce latency

Developers using large-scale linguistic models (LLMS) in production, latency and jitter are deeply concerned. Jones highlighted a new feature called Green conditionswhich allows GPU-specific partitioning.

“Green instances allow you to separate the GPU … into different components,” Jones said. This allows developers to allocate certain parts of the GPU to different tasks, such as running pre-completion and cleanup simultaneously without competing for resources. This is a small level technology within the GPU in one view of the GPU Difference seen in the scale of the data.

No black boxes: The importance of tool discovery

One of the most common fears regarding high-level writing is loss of control. Jones, drawing on his experience as a Cuda user in the aerospace industry, emphasized that NVIDIA tools will never be black boxes.

“I really believe the most important part of the Cuda is the engineer's tools,” Jones assured. He assured developers that even if they were using tile-based abstracts, tools like nisight accoute would allow inspection of the instructions and registers of each machine. “You have to be able to synchronize and damage and position well … it can't be a black box,” he added.

Accelerating Time-to-Result

Ultimately, the purpose of these reviews is productivity. Jones described the objective as a 'left-shifting' performance curve, enabling developers to reach 80% of potential performance in half the time.

“If you can get to the market [with] 80% of the work in a week instead of a month … then you spend all your time,” explained Jones

Lasting

As AI algorithms and scientific computing converge, nvidia positions cuda not as a low-end tool for hardware professionals, but as a flexible platform that adapts to the needs of Python developers and HPC researchers alike. With support from Ampere's Blackwell Blackwell and Rubin architectures, this update promises to improve the development of the GPU Ecosystem as a whole.

For complete technical details on Cuda tile and green modes, visit the NVIDIA Developer Portal.