Track Hyper | Intel's first GPU commercialization encounters setbacks

After "ousting" former CEO Pat Gelsinger, Intel's progress in technology and products remains challenging.

On January 30, interim co-CEO Michelle Johnston Holthaus stated during Intel's earnings call, "Based on industry feedback, we have decided to utilize Falcon Shores as an internal testing chip. Without bringing it to market, we will support efforts to develop rack-level system solutions in collaboration with Jaguar Shores to address AI data center issues more broadly."

Falcon Shores is Intel's first AI GPU, manufactured using TSMC's advanced 3nm and 5nm processes, and employs TSMC's CoWoS-R packaging technology for HBM integration.

Falcon Shores features HBM3 (288GB) with a total bandwidth of 9.8TB/s, supporting 8-bit floating-point operations, meeting the high demands for memory and data transfer in large-scale data processing and complex AI tasks; like Gaudi 3, Falcon Shores also utilizes Intel's Xe-HPC architecture (the first GPU from Intel to adopt this architecture), integrating fully programmable architecture and the system performance of the Gaudi 3 AI accelerator, allowing for seamless switching between the two in system development.

However, the consequences of Falcon Shores' numerous issues are more severe than the software problems faced by Gaudi 3, leading Intel to refrain from bringing it to market: Intel has canceled the market release of the Falcon Shores GPU aimed at AI and high-performance computing applications, opting instead to use it for internal software testing in collaboration with the upcoming Jaguar Shores product to build rack-level AI solutions.

Jaguar Shores is also a GPU, serving as the successor to Falcon Shores and a follow-up product to Gaudi 3, utilizing Intel's latest 18A manufacturing process and introducing new technologies such as RibbonFET and backside power delivery.

Compared to mainstream GPU design approaches, Jaguar Shores integrates GPU and ASIC design philosophies, offering performance advantages over general-purpose GPUs when handling specific AI tasks; it possesses multimodal processing capabilities and excellent software compatibility.

Falcon Shores was originally seen as a crucial part of Intel's transformation into the AI market, intended to compensate for the low market acceptance of Gaudi 3 due to software issues, but it now amounts to abandoning Intel's only "finished" GPU (Holthaus considers it a "semi-finished" product).

Unlike NVIDIA's AI accelerator card design, which is primarily used for IDC cluster training, Falcon Shores' design philosophy is aimed at developing a hybrid processor hardware and software ecosystem, which also means that during its internal development process, it could provide an important technical foundation for Intel's next-generation AI accelerator cards, including the Jaguar Shores iterative GPU Because Falcon Shores has a richer ecological color, Intel plans to share samples with partners (that is, provide them to partners rather than just using them internally) despite using it as an internal testing chip. This is to ensure that external developers and independent software vendors can efficiently adapt to the new GPU architecture, paving the way for compatibility for Intel's future AI card development.

These partners include independent hardware vendors (IHV) and independent software vendors (ISV).

This approach actually introduces a new idea to the industry: enhancing the completeness of ecosystem construction when developing new products may be more important than launching a single technology.

From this perspective, Intel's abandonment of the commercialization of Falcon Shores and its shift to using it as a tool for improving the AI accelerator card co-development ecosystem may seem like a commercial setback in the short term, but it may not be in the long term.

If Intel's change in the use of Falcon Shores is driven by necessity, the current situation is even more challenging, as Intel's AI chip now only has the flawed Gaudi 3 left, especially since this is considered a failed product.

Holthaus candidly stated at Barclays' annual global technology conference, "Gaudi cannot help me reach the masses; it is not a GPU that can be easily deployed across global systems."

Currently, Intel is in a predicament. Initially hoping that the Gaudi 3 processor and the subsequent Falcon Shores (GPU) would secure an entry ticket to AI, the market can only turn its attention back to Jaguar Shores with the commercial failure of Falcon Shores.

However, how much patience does the market have to wait for Intel's slow transformation?

Fortunately, Intel has made significant progress in the traditional CPU field, especially in server CPUs. For example, Granite Rapids has directly closed the performance gap that AMD's Genoa series chips had previously surpassed.

At the same time, Intel has also resorted to price war measures, reducing the price of the flagship version of the Granite Rapids series chips from $5,340 to $12,460 without official announcement, positioning itself below AMD's comparable chips to resonate with the market.

This has won valuable time for Intel's long transformation.

Intel's fourth-quarter financial report for 2024, released after the U.S. stock market closed on January 30, showed that revenue for the quarter decreased by 7.4% year-on-year to $14.26 billion, better than market expectations.

Excluding certain one-time items, the loss per share was $0.03, with a quarterly loss of $126 million, far less than the $16.6 billion loss in the third quarter of 2024, partly due to Asian customers' concerns about tariffs, which led to early orders, as well as obtaining subsidies from the CHIPS Act.

At the same time, management announced that Intel would not abandon its IF business (foundry) and plans to break even on losses by 2027. The overall cost of new factories and equipment in 2025 is estimated to be about $20 billion, lower than the previous estimate of up to $23 billion Nevertheless, the abandonment of projects like Falcon Shores and the underperformance of Gaudi 3 (which did not meet the expected sales target of $500 million) have negatively impacted Intel's performance.

Intel expects revenue for the first quarter of 2025 to be approximately $11.7 billion to $12.7 billion, below analysts' estimate of $12.85 billion, with earnings per share estimated to break even, falling short of the market expectation of $0.08 per share, due to weakening demand and declining market share