This article was originally published in German and has been automatically translated.
Unfortunately sold out! This is reported by two of the world's three manufacturers of High Bandwidth Memory (HBM), the type of RAM required by many of the most powerful AI computing accelerators. The US company Micron and the Korean company SK Hynix have sold their planned HBM production well into 2025. So prices are going up - this news should please investors. Everyone who can somehow wants to benefit from the AI boom. According to the Taiwanese market research company Trendforce, the expensive HBM chips already account for around 30 percent of total DRAM sales.
At the world's largest chip contract manufacturer TSMC, on the other hand, the capacity of the high-tech packaging plants is said to be fully utilized. These are the production lines that pack the AI chips developed by Nvidia and AMD, among others, and manufactured by TSMC together with the HBM memory stacks in the same housing. These bottlenecks make it clear that newcomers to the AI chip market have to get in line. Those who have not yet reserved production capacity will probably only be able to deliver larger quantities from 2026. Nvidia's quasi-monopoly on AI chips seems to be set in stone.
Heating up the race
The supercomputers for classic high-performance computing (HPC) look like pocket calculators next to the insanely expensive AI giant machines. What is the 500 million US dollar purchase price and 40 megawatts (MW) power consumption of the exaflops system Aurora compared to 3.2 billion euros and several hundred MW for a Microsoft data center in the Rhenish lignite mining area? At the ISC 2024 supercomputer conference in Hamburg, there was talk of cooling systems that shovel 100 kilowatts out of a single server rack. Let's hope that the power is enough for AI.
The market for traditional servers for companies (enterprise) has been shrinking for years, but recently also that for "normal" cloud servers. This is because the dominant hyperscale cloud companies are investing their money in AI machines. The classic CPU providers are looking for arguments. Ampere Computing - whose 192-core Ampere One was announced a year ago but is still rarely seen or rented - believes that AI inferencing also runs well and efficiently on many "small" ARM cores. AMD (Eypc 9754 "Bergamo" with 128 cores) and soon Intel (Xeon 6700E "Sierra Forest" with 144 E cores, 6900E with 288 cores) will certainly be telling potential customers the same thing.
Intel's eagerly awaited "Granite Rapids" with P-cores is also coming as a Xeon 6, i.e. as a 6700P or 6900P. What was not so clear until now is that there are two performance classes for different mainboards. The "SP" platform with the LGA4710 version connects eight RAM channels, with LGA7529 (AP) there are twelve per CPU. While the SP types can consume up to 350 watts, the AP can consume up to 500 watts; this is the only way to achieve the aforementioned 288 E-cores or up to 128 P-cores. This would finally bring Intel back into line with the Epycs; the "Turin" (9005) is also expected to have 128 Zen 5 cores or up to 192 of the more compact Zen 5c type. Allegedly, even 550 watts TDP are required for this.
For small servers, AMD also intends to launch an Epyc 4004P series for AM5 mainboards, which actually contains Ryzen technology. Intel's Xeon E sends its regards.
There are also rumors that Qualcomm wants to make a new attempt at ARM server chips, namely with the purchased "Nuvia" cores, which compute as Oryons in the Snapdragon X Elite/Plus. However, the market potential seems to be evaporating: Following Amazon (Graviton), Microsoft (Cobalt) and Google have also announced their own ARM server chips. The only major customers left for Ampere's ARM server chips are Oracle Cloud and a few smaller or regional providers, such as Hetzner in Europe.
Hatched
A 311-page internal PDF document from Dell, which has surfaced on the online platform Scribd, reveals just how complex the development of a new notebook is and what components cost. It is about an XPS 13 with Qualcomm Snapdragon X Plus. The ARM CPU with ten cores costs 145 US dollars, 48 percent less than Dell pays for Intel's Core i7-1360P. However, the mainboard for the ARM CPU is more expensive and with a thermal design power (TDP) of 25.5 watts, it also requires a fan. Dell receives considerable discounts from Microsoft and Qualcomm for the Windows license and on the chip prices - as is also known from Intel's "advertising subsidies". The competition between ARM and x86 is entering a new round.
There is also a regular podcast on Bit-Rauschen.
Alle 14 Tage präsentiert Ihnen Deutschlands größte IT-Redaktion aktuelle Tipps, kritische Berichte, aufwendige Tests und tiefgehende Reportagen zu IT-Sicherheit & Datenschutz, Hardware, Software- und App-Entwicklungen, Smart Home und vielem mehr. Unabhängiger Journalismus ist bei c't das A und O.
- c’t-Ausgaben online lesen
- c’t im heise-Shop kaufen
(ciw)
