OpenAI Builds First Chip With Broadcom and TSMC, Scales Back Foundry Ambition (reuters.com) 12
OpenAI is partnering with Broadcom and TSMC to design its first in-house AI chip while supplementing its infrastructure with AMD chips, aiming to diversify its reliance on Nvidia GPUs. "The company has dropped the ambitious foundry plans for now due to the costs and time needed to build a network, and plans instead to focus on in-house chip design effort," adds Reuters. From the report: OpenAI has been working for months with Broadcom to build its first AI chip focusing on inference, according to sources. Demand right now is greater for training chips, but analysts have predicted the need for inference chips could surpass them as more AI applications are deployed. Broadcom helps companies including Alphabet unit Google fine-tune chip designs for manufacturing and also supplies parts of the design that help move information on and off the chips quickly. This is important in AI systems where tens of thousands of chips are strung together to work in tandem. OpenAI is still determining whether to develop or acquire other elements for its chip design, and may engage additional partners, said two of the sources.
The company has assembled a chip team of about 20 people, led by top engineers who have previously built Tensor Processing Units (TPUs) at Google, including Thomas Norrie and Richard Ho. Sources said that through Broadcom, OpenAI has secured manufacturing capacity with Taiwan Semiconductor Manufacturing Company to make its first custom-designed chip in 2026. They said the timeline could change. Currently, Nvidia's GPUs hold over 80% market share. But shortages and rising costs have led major customers like Microsoft, Meta, and now OpenAI, to explore in-house or external alternatives.
The company has assembled a chip team of about 20 people, led by top engineers who have previously built Tensor Processing Units (TPUs) at Google, including Thomas Norrie and Richard Ho. Sources said that through Broadcom, OpenAI has secured manufacturing capacity with Taiwan Semiconductor Manufacturing Company to make its first custom-designed chip in 2026. They said the timeline could change. Currently, Nvidia's GPUs hold over 80% market share. But shortages and rising costs have led major customers like Microsoft, Meta, and now OpenAI, to explore in-house or external alternatives.
Another vapour ware they (Score:2)
Wut? (Score:3)
Did they ever read a primer on semiconductor physics? Tell me they know a transistor is more than a switch? Oh right, they were going to let AI take care of all that.
This is beyond arrogance. The idea is just madness. Did they ever wonder why Apple did not start its own foundry yet?
I can go on like this for hours. Mind boggling.
Re: (Score:3)
If they didn't want bleeding edge nodes, they could have licensed something from Samsung (for example). But even then they'd be competing for the equipment and talent to build and operate their own fab. It's probably for the best that they stuck with TSMC.
Re: Wut? (Score:2)
Re: (Score:2)
I was triggered here by IT guys thinking they are smart and can do anything. I have a few arrogant ones in the far family that think they are on top of the world because they can program a computer. One of them once explained to me how dcdc converters work. You know with magnets and a square wave. (I was involved in designing a dcdc then at work. The dude was serious.) He changed the subject when he could not keep up when I told him I was struggling with pointers in C and memory leaks for a hobby project. He said to me that that was the old way of programming. They used Python these days. Even on microcontrollers. There is this specific group amongst programmers that really needs a reality check. I switched to teaching when I no longer could keep up with the fresh young ones. Guess what. I started together with a programmer ("He designed Natos network infrastructure", part time). On his first day, he explained to the teachers (old and experienced) what was wrong with the education system. In his speech he made rookie mistakes that are well explained in any decent book about education. He left after a year through the back door. Students threw stuff at him, that is very exceptional. I can go one (again). Oh well...
I have a, let's say relative, though it's through marriage, that walks into new jobs and lectures the management on how fucked up their entire business is until they get fired, then they complain to everyone that if the stupid bastards would just listen to her they'd be making money hand over fist. Literally on the first day, even when switching business sectors. Some people just have arrogance in their DNA. It's not just programmers. She couldn't program her way out of Hello World in PHP.
Re: (Score:2)
Cocaine is a hell of a drug!! (Score:2)
Re: (Score:2)
Translation (Score:3)
The company has dropped the ambitious foundry plans for now due to the costs and time needed to build a network, and plans instead to focus on in-house chip design effort,
Translation: OpenAI has fucked around and now they have found out.
Silicon is an expensive medium to revise and it takes a different approach than software projects to get it right the first time.
Re: (Score:3)
Diversification (Score:2)
Seeing as how the hardware involved with AI is the lion's share of the cost and it is a physical resource that everyone seems to be ready to dump unlimited money on, it makes very good sense for OpenAI to try to create their own. They hired top guns, they "secured manufacturing capacity with Taiwan Semiconductor Manufacturing Company", and maybe they will have something valuable out in 2026. Seems like a reasonable plan.
Would they produce different results? (Score:2)
It would be interesting to find out if given a complex AI task that didn't have a binary answer, would their chip produce different results than the nVidia chip. Something like "design an efficient electric motor". If the two chips produced the same result, ok. If they produced different results, you could build both and test them out.