La division GPU AI d’Intel connaît des difficultés, notamment avec le chip Gaudi 3, dont les ventes ne devraient pas atteindre les 500 millions de dollars visés pour 2024. En revanche, AMD a réalisé 1,5 milliard de dollars de ventes de ses puces MI300 en septembre. Pat Gelsinger a également signalé que la mémoire intégrée au chip Lunar Lake ne sera pas reproduite dans les futurs modèles. Intel réorientera sa production de silicium TSMC et reste concentré sur ses CPU x86, sans nouvelles concernant les cartes graphiques Battlemage.
Recently, we noted that AMD’s AI GPU sector has surged from nothing to match its overall CPU sales in under a year. In contrast, the news from Intel is less favorable, as the Gaudi 3 AI chip appears to be a disappointment.
During the latest earnings call with analysts and investors, CEO Pat Gelsinger expressed that Intel will not meet its target of $500 million in sales of Gaudi 3 GPUs for 2024.
For comparison, AMD announced that it achieved over $1.5 billion in sales of its MI300 AI chips in September, while Nvidia is expected to surpass Intel’s Gaudi 3 sales target by more than a hundredfold next year. Clearly, the Gaudi 3 is falling short.
Gelsinger also highlighted that the fully integrated memory used in the innovative Lunar Lake laptop chip will be a unique occurrence.
‘It’s truly a one-off with Lunar Lake,’ Gelsinger stated regarding the chip’s integrated memory. ‘This approach won’t carry over to Panther Lake, Nova Lake, and their successors. We will revert to a conventional design with off-package memory while keeping CPU, GPU, NPU, and I/O capabilities intact in the package. Volume memory will be off-package moving forward.’
The primary motivation for this shift appears to be cost-related, as integrated memory yields lower profit margins for Intel compared to separate memory. Gelsinger also revealed that Lunar Lake wasn’t intended to be a major contender in the mobile chip lineup or a true successor to Meteor Lake.
‘Lunar Lake was initially conceived as a niche product aimed at maximizing performance and battery life, but with the rise of AI PCs, it transitioned from a niche offering to a high-volume product,’ he explained.
In other developments, Gelsinger confirmed Intel’s intention to decrease its reliance on TSMC silicon for its CPUs in the forthcoming product cycles. He mentioned that ‘most of the Panther Lake wafer capacity is returning to Intel, while some Nova Lake SKUs will still rely on external sources, but a significant portion will also be managed in-house.’
Curiously, Intel’s Battlemage gaming graphics were not addressed. This technology is supposed to succeed the initial Arc gaming GPUs and has been frequently referenced by Intel, raising expectations for an earlier release this year. However, Gelsinger’s omission casts doubt on any new graphics cards utilizing Battlemage tech being available this year.
This is unfortunate, as Battlemage will now contend with Nvidia’s upcoming RTX 50 GPUs and AMD’s RDNA 4 architecture, both projected for early next year, rather than competing against previous generations. This scenario will likely intensify the challenges faced by Intel.
In fact, rather than discussing gaming graphics, Gelsinger reaffirmed Intel’s dedication to its x86 CPUs. ‘Our focus remains on our x86 franchise and the ecosystems we’ve built over more than 40 years,’ he stated. ‘We are actively pursuing ways to enhance and maximize the potential of our x86 franchise.’