news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

there’s new news about openai’s first in-house ai chip.

according to taiwan economic daily news,tsmc will develop an a16 angstrom-level process chip customized for the openai sora video model, aiming to improve sora's video generation capabilities.

for a brief introduction, 1 angstrom is equivalent to one tenth of 1 nanometer. after the current semiconductor process has broken through the 2-nanometer process, angstrom has become a new battlefield for the world's leading chip giants.

the name "a16" has little to do with apple's a16 chip. the name a16 represents the process of 16 angstroms, or 1.6 nanometers.

this is the most advanced process node tsmc has disclosed so far, and it is also tsmc's first foray into the angstrom process.mass production is expected to begin in the second half of 2026.

it is reported that a16 will adopt the next generation of nanosheet transistor technology and super power rail technology (spr). spr is an original and industry-leading backside power supply solution and is the first technology in the industry.

super rail technology can transfer the power supply network to the back of the wafer, freeing up more space on the front of the wafer, thereby significantly improving logic density and performance, making the a16 suitable for high-performance computing (hpc) products that require complex signal routing and dense power supply networks.

compared with the n2p process, the chip density of a16 is increased by 1.10 times, and the speed is increased by 8-10% at the same operating voltage; at the same speed, the power consumption is reduced by 15-20%.

although tsmc's a16 process has not yet entered mass production, the first batch of customers who placed orders have already been exposed.

it is reported that apple has reserved the first batch of production capacity from tsmc, and openai has also reserved the production capacity of a16 due to the manufacturing needs of its self-developed ai chips.

in fact, with the continued demand for generative ai, ai chips have become the next battlefield for giants to compete for, and openai has never given up on developing its own ai chips.

taiwanese media reported that openai had originally actively negotiated with tsmc to cooperate in building a dedicated wafer fab, but after carefully evaluating various factors including profitability, the plan was ultimately canceled.

turning back the clock to july, according to the information, citing three people familiar with the matter, openai has recently been actively recruiting former employees who were involved in the production of google's ai chip tpu and seeking to develop ai server chips.

there have been previous reports that openai will provide these google senior engineers with equity worth millions of dollars each year. it is reported that openai's chip research and development team will be led by richard ho, who once participated in the tpu project at google.

in addition, openai is also considering working with an american company to assist in the development of this new chip.

for example, as one of google's tpu manufacturing partners, broadcom has discussed chip manufacturing with openai's chip research and development team.

however, the openai team does not appear to have started designing the chip yet, and it will not go into production until 2026 at the earliest. the team is evaluating various chip packages and memory components in an effort to maximize chip performance.

earlier this year, the wall street journal reported that openai ceo sam altman was in talks with investors to raise money for a chip joint venture, noting that it could require raising up to $7 trillion.

altman later debunked the rumor, saying the figure represented the total investment that participants in such a venture would need to make over a few years, including everything from real estate and power for data centers to chip manufacturing.

regarding openai's possible manufacture of ai chips, the attitudes of all parties are somewhat ambiguous.

on the one hand, nvidia is openai's main chip supplier, and nvidia ceo jensen huang even personally delivered the first dgx h200 chip to openai. therefore, openai's choice to develop its own ai chip may cause dissatisfaction with nvidia.

on the other hand, although developing a new server chip that can compete with nvidia is unlikely and will take years to achieve, developing its own ai chips may provide openai with potential bargaining chips in future pricing negotiations with nvidia.

altman once said that without major breakthroughs in energy technology, artificial intelligence will not be able to reach the next stage.

therefore, many people believe that instead of spending time on the manufacture of ai chips, openai should pay more attention to the energy crisis facing ai development.

as of press time, tsmc declined to comment on the above report, and openai has not yet responded to the matter.