[News] Intel Renews AI Chip Push with Crescent Island, 2H26 Sampling; Reportedly Targeting Annual GPUs

After unveiling its first major in-house 18A product, Panther Lake, last week, Intel is renewing its push into the AI chip market following the limited success of its earlier Gaudi series. As announced at the OCP 2025 Global Summit and reported by Tom’s Hardware and Reuters, the company’s next-generation data center GPU, Crescent Island, is scheduled to begin customer sampling in the second half of 2026.

With Crescent Island, Intel is reportedly targeting a lean, inference-focused GPU to regain ground in the fast-growing AI data center market. Notably, CTO Sachin Katti told Silicon Angle that Intel plans to ditch irregular release cycles and launch new GPUs annually, keeping pace with NVIDIA, AMD, and cloud players like AWS and Google, who design their own AI chips.

As per the report, Katti also stressed that Intel will focus on AI inference, the hardware that drives models in production rather than training.

Specs Focus: Memory

As per Intel, Crescent Island is being designed to be power and cost-optimized for air-cooled enterprise servers and to incorporate large amounts of memory capacity and bandwidth, optimized for inference workflows. Tom’s Hardware notes the chip features one or possibly two Xe3P-based GPUs — an enhanced version of the Xe3 architecture powering the Core Ultra 300 “Panther Lake” lineup for notebooks and small desktops.

As Wccftech points out, the chip is equipped with an impressive 160GB of LPDDR5X memory—a notable shift from the industry’s current reliance on HBM. The report adds that while rivals like NVIDIA and AMD are pushing ahead with HBM3e and even preparing for HBM4 in their next-generation Rubin and MI400 accelerators, Intel’s choice of LPDDR5X reflects a strategic trade-off.

With HBM supply tightening and costs rising, Intel appears to be betting on more accessible and scalable memory solutions to stay competitive in the rapidly expanding AI inference market, Wccftech suggests. According to the report, Intel’s use of LPDDR5X memory could give it a notable cost-performance edge. The new architecture is also expected to support a broad range of data types, making it ideal for “Tokens-as-a-Service” and AI inference workloads.

Tom’s Hardware further reveals that Team Blue already has early prototypes of Crescent Island, with sampling planned for the second half of 2026.

Software Updates

On the other hand, as noted by Intel, the company is already testing its open, unified software stack for heterogeneous AI systems using its Arc Pro B-series GPUs—meaning future models will benefit from these optimizations early on.

According to TechNews, citing Intel’s remarks at OCP, the company is developing a unified software stack and orchestration infrastructure designed to enable applications to “just work” across PyTorch, Hugging Face, and LangChain—without requiring developers to modify their code. The system will reportedly profile agentic AI workloads, assign components to appropriate hardware, and orchestrate end-to-end service level agreements (SLAs). Intel expects to roll out this infrastructure in Q4 2025, as per TechNews.

Read more

• [News] What to Expect from Intel Panther Lake, Featuring 18A and Reportedly TSMC N3E/N6 Processes

(Photo credit: TrendForce at 2025 OCP Global Summit, Intel)