Bit-Rauschen: speculation about Intel's future fuels the share price

Spring fever at Intel as early as February: the share price of the beleaguered chip manufacturer shot up by a quarter. The background to this is speculation that the US government wants to retain Intel's manufacturing capacities in the USA. There have been rumors about the spin-off of the manufacturing division for some time, perhaps it will go to a consortium of several companies and investors. Allegedly, the Taiwanese market leader TSMC is also being urged to participate and is hoping for licenses for certain manufacturing processes. It is difficult to say whether this would be technically feasible in a reasonable amount of time and whether it is a good idea economically. Intel is also very confident that its in-house production technology 18A will be at the forefront and will actually go into series production in 2025.

Broadcom is increasingly being mentioned as a potential buyer for Intel's product division – processors, network adapters, graphics chips and AI accelerators –. Broadcom CEO Hock Tan has an impressive track record. Born more than 70 years ago in what is now Malaysia, the US American was once CFO of Commodore in the 1990s and briefly headed Integrated Device Technology (IDT), which is now part of Renesas, in 2005. From there, he moved to Avago, which had emerged from the former semiconductor division of HP called Agilent. In 2015, Tan very cleverly merged Avago with Broadcom and trimmed the company for profitability. Today, Broadcom – with the stock market ticker symbol AVGO – has almost exactly ten times the market capitalization of Intel. As already reported, Broadcom, unlike Intel, is benefiting enormously from the AI hype, partly because the company is a contract developer for Google's TPU, for example. Broadcom may be able to employ the many thousands of Intel developers more lucratively than Intel itself can.

ARM CPU

Meanwhile, there are increasing indications that ARM is developing its own physical server processor; the first buyer is said to be Facebook or Meta. It will be interesting to see how the existing ARM licensees deal with the fact that ARM is becoming a chip competitor. But it will also be interesting to see whether an AI computing chip with ARM technology will soon follow.

The powerful AI model DeepSeek-R1 from China is certainly raising hopes that it will be possible without super-expensive Nvidia hardware. However, there are only a few signs of this so far. For example, our colleague Jan-Keno Janssen from the YouTube channel c't 3003 managed to get DeepSeek-R1 running at full speed on a server with two AMD Epyc. But the responses came at only around four tokens per second. The developers of the SGLang serving framework for large language models (LLMs) recommend an Nvidia DGX H200 for DeepSeek-R1, which costs around 374,000 euros and also contains two server processors and 2 TB of RAM. But it also has eight H200 modules with a total of 1.1 terabytes of GPU memory. DeepSeek runs much faster on this. What's more, Nvidia makes it particularly easy for admins, at least if they have a subscription to the AI Enterprise Suite. Since the end of January, they have been able to download the pre-release version of an Nvidia Inference Microservice (NIM) for DeepSeek-R1.

We would have expected AMD to pull out both legs to show as quickly as possible that DeepSeek-R1 also runs on its AI chips and, above all, how it does so. However, up to the editorial deadline for this issue, we were unable to find any concrete measurement values from AMD. But at least we were able to find hints and tips on how to get DeepSeek running with SGLang on a server with eight Instinct MI300Xs.

Trump tariffs dampen PC market

The first gaming benchmarks of AMD's "super APU" have emerged: The Ryzen AI Max+ 395 aka Strix Halo certainly comes close to mobile versions of the Nvidia GeForce 4060.

US President Donald Trump could spit in AMD's soup with his tariff policy. Acer therefore wants to increase the prices for notebooks in the USA by around 10 percent.

There is also a regular podcast on bit noise. (ciw)