Despite this, programmers say that moving the code from the miracle of Nvidia to ROCM is not an productive process, which means that they usually focus on building only one system supplier.
“Rocm is amazing, it is open source, but it works on the equipment of one supplier,” Lattner He told the crowd at AI AI promotion in June. Then he made Modular software more portable and makes GPU much faster.
Lattner’s conversation in AMD is representative of the type of dance that Lattner and Davis must do when they spread the modular gospel. Today, Nvidia and AMD are the company’s key partners. In the future universe they are also direct competitors. Part of the proposal of the Modura value is that it can send the software for GPU optimization even faster than NVIDIA, because there may be a difference between the months between this when NVIDIA sends a fresh GPU, and when it releases the “bridge of attention”-the cluster part of the GPU software.
“At the moment modular is free for AMD and Nvidia, but over time you could see how both companies feel threatened by ROCM or miracles that are not the best software that lies on their fries,” says Monichiello. He is also worried that potential cloud customers can be examined, the need to pay for an additional layer of software, such as Modular.
Writing software for the GPU is also “dark art”, says Waleed Atallah, co -founder and general director of MAKO, a GPU optimization company. “Maping of the algorithm on the GPU is incredibly difficult to do. There are one hundred million software programs, 10,000 who write GPU nuclei, or maybe a hundred who can do it well.”
Mako builds AI agents to optimize coding for GPU. Some programmers believe that this is the future of the industry, instead of building a universal compiler or a fresh programming language, such as a modular. Mako has just collected $ 8.5 million for seeds for seeds from Flybridge Capital and the startup NEO accelerator.
“We try to apply an iterative approach to coding and automate it with AI,” says Atallah. “By making it easier to write code, you exponently develop the number of people who can do it. Making another compiler is a more fixed solution.”
Lattner notes that the modular also uses AI coding tools. But the company intends to solve the entire coding stack, not just the testicles.
There are about 250 million reasons why investors believe that this approach is profitable. Lattner is something like a luminarian in the world of coding, earlier building the Open Source LLVM compiler infrastructure project, as well as the SWIFT Apple programming language. He and Davis are convinced that this is a software problem that should be solved outside a immense technological environment, in which most companies focus on building software for their own technological stack.
“When I left Google, I was a bit depressed because I really wanted to solve it,” says Lattner. “We realized that it is not about intelligent people, it is not about money, it is not about ability. This is a structural problem.”
Monichiello shared a universal mantra in the world of technological investments: he says that he is put on the founders themselves, like their product. “There is high opinions and impatient, as well as a lot of time,” said Monichiello about Lattner. “Steve Jobs was also like that – he did not make a decision -based decisions, but he was often right.”