Write down this name: Blackwell. Because it can make a difference at a delicate moment for the transition of artificial intelligence. Two years after the launch of ChatGPT and after the main players in the AI craze have invested close to $200 billion in the sector, many companies; starting with OpenAI, they have to deal with income that still does not justify the enormous expenses incurred and with balance sheets in the red.
If there is a company that has not had too much difficulty taking advantage of the explosion of generative AI, it is Nvidia. The latest bimonthly figures once again demonstrate the unstoppable growth of the colossus founded by Jensen Huang, which with its graphic processing units (GPU) has made its own artificial intelligence revolution possible. In the third quarter of 2024, Nvidia announced revenues of $35.1 billion, exceeding expectations of $33.2 billion, and marking a 94% increase over the same period last year. Although this is an inevitable slowdown compared to previous quarters; when the growth compared to 2023 had even exceeded 260%, it shows that the line of companies waiting to get their hands on Nvidia chips is still very long, and above all, that it is not going to end soon.
The Blackwell era
The biggest speculation surrounding Nvidia is its chips will continue to impress investors and continue to break records. Your new product could be the key: Blackwell is the successor to the previous Hopper architecture, the basis of the H100 chips used to train almost all recent neural networks.; is specifically designed to ensure the best performance in the field of AI.
It was first announced in March during the GTC 2024 conference, where Huang said Blackwell would “play a critical role in advancing the latest large language models (LLM).” A product highly anticipated by all the giants of Silicon Valley, given the series of disappointing results obtained by the latest versions of the systems developed by OpenAI, Google, Anthropic and others.
According to early reports, Blackwell’s GPU, which consists of 208 billion transistors and is made up of two silicon units, each comparable in size to the previous Hopper chip, combined into a single component, appears to be delivering on its promises. Blackwell was put to the test during the first half of November in the MLPerf Training test, which measures “how quickly systems can train models to achieve a given quality metric.” According to Nvidia data, Blackwell managed to multiply its performance by 2.2 compared to its previous version, Hopper.
But there was no shortage of difficulties: first, overheating problems, which reportedly plagued Blackwell for months when the chips were connected to Nvidia’s custom rack servers. The silicons were withdrawn from the market to change their design, but serious delays affected their commercialization. Patrick Moorhead, an industry analyst, explains that “it is not unusual for there to be tensions between the divisions responsible for design, engineering and manufacturing when a new chip is launched.” Now, however, the first thousands have already been delivered to Nvidia’s main partners, news that Microsoft confirmed in X.
Delays in chip production and rumors about overheating problems have not stopped Nvidia, which has once again reported a quarter of great profits and claims that the Blackwells are now in the hands of Microsoft and OpenAI.
A complex production
As Nvidia’s own CEO explained, Blackwell’s production chain involves a practically endless supply chain: Asian semiconductor companies such as Taiwan Semiconductor Manufacturing Company (TSMC), SK Hynix and SPIL, Taiwanese electronics giant Foxconn, American fiber optics manufacturer Amphenol, and data centers Wiwynn and Vertiv, among others.
Additionally, Nvidia’s profits could be tested by Blackwell’s production, which is much more expensive than its predecessors. Although, according to analysts, this is not a “major obstacle” for the company since the AI market will continue to grow in the coming years. At the moment Nvidia has practically no rivals, on the contrary, it has a monopoly in the GPU sector, with an 88% market shareand is pursued by a couple of adversaries, such as Advanced Micro Devices (AMD), who are after startups that are currently unable to match their production volumes.
As demonstrated by its rise to first place among the companies with the highest market value in the world, a position currently disputed with Apple, the great artificial intelligence race has been won by Nvidia. That from the top she can allow herself some stumbles, without fear of being overtaken.
Article originally published in WIRED Italy. Adapted by Alondra Flores.
#Blackwell #Nvidia #chip
