Intel officially names Crescent Island GPU D-Series at OCP 2026, revealing 1.3 PFLOPS FP8 performance and 1.5 TB/s memory bandwidth

At the Open Compute Project (OCP) summit today, semiconductor giant Intel provided a significant update regarding its highly anticipated AI inference accelerators, formally giving the codename "Crescent Island" a production-ready brand: the Intel GPU D-Series. Alongside this official naming announcement, the company shared the first-ever theoretical peak performance targets for the hardware, pointing to a staggering 1.3 PFLOPS of FP8 computing power. Furthermore, Intel disclosed the final memory bandwidth specifications users can expect from these specialized accelerators, projecting it at 1.5 TB/s. This comprehensive batch of disclosures builds upon the steady stream of information Intel has released about Crescent Island throughout 2026, gradually lifting the veil on its strategy for the competitive enterprise artificial intelligence landscape.

The latest revelations from the OCP event clarify several key performance metrics that industry analysts and enterprise data center operators have eagerly anticipated. While Intel has previously touched upon the broad strokes of its hardware roadmap, today’s confirmation of a 1.3 PFLOPS FP8 theoretical peak performance metric gives a much clearer picture of where the GPU D-Series will position itself in terms of raw throughput for low-precision workloads. Low-precision data formats like FP8 have rapidly become the industry standard for accelerating large language model inference tasks, striking a critical balance between maintaining model accuracy and dramatically increasing processing speed and efficiency. By highlighting this specific metric, Intel is signaling its capability to compete directly in high-density enterprise environments where inference throughput and energy efficiency dictate data center economics.

Complementing this compute capability is the newly confirmed memory bandwidth specification of 1.5 TB/s. In modern artificial intelligence inference accelerators, memory bandwidth is often the primary bottleneck constraining performance, particularly when dealing with massive models that require rapid weight fetching from memory subsystems. A robust memory bandwidth of 1.5 TB/s ensures that the powerful underlying processing cores can be fed data at a rate sufficient to maintain high utilization rates, preventing the computational units from idling while waiting for data transfers. This balanced approach to compute and memory underscores the architectural thought process behind the Intel GPU D-Series, positioning it as a well-rounded solution for demanding inference workloads running at scale.

These newly announced figures add substantial context to the slow drip of disclosures Intel has made regarding Crescent Island throughout the course of 2026. As the semiconductor industry shifts its primary focus toward specialized artificial intelligence infrastructure, companies like Intel have had to adapt their product communication strategies, offering incremental architectural previews at major industry events before delivering full product launches. The OCP summit, known as a premier gathering for hyperscalers, data center architects, and hardware developers, proved to be the ideal venue for Intel to consolidate its previous updates and present a cohesive vision for the GPU D-Series.

Industry observers already knew from earlier disclosures that Crescent Island stands as a milestone product for Intel, serving as the very first chip to be built upon the company’s advanced Xe3P graphics architecture. However, a closer examination of the architecture reveals that labeling it purely as a graphics processor is a misnomer. In reality, Crescent Island functions as a massively parallel computing architecture tailored explicitly for dense computational workloads. The chip notably lacks most, if not all, of the graphics-specific functional units—such as traditional rasterization pipelines, display engines, and specialized rendering hardware—that would typically characterize a true graphics processing unit. Instead, the silicon real estate normally dedicated to visual rendering has been entirely repurposed for parallel processing units optimized for matrix multiplication and tensor operations vital to modern neural networks.

To maximize its utility within modern server racks and enterprise data center deployments, Intel has engineered the GPU D-Series to conform to standard industry form factors and thermal envelopes. The company confirmed today that it will release these specialized accelerators in a traditional air-cooled PCIe form factor. This design decision is particularly significant for data center operators because it means the GPU D-Series can be integrated directly into existing server infrastructure without requiring costly liquid-cooling retrofits or proprietary chassis designs. By adopting an air-cooled approach with a tightly managed thermal design power (TDP) of 350W, Intel is ensuring broad deployment flexibility, allowing enterprises to slide these accelerators into standard PCIe slots alongside existing host processors and networking cards.

Another striking physical and architectural characteristic of the Intel GPU D-Series is its massive capacity for local memory. Intel reiterated today that the accelerators will support up to 480GB of LPDDR5X memory. This represents a distinct architectural choice compared to competitors who rely exclusively on high-cost, high-bandwidth memory variants like HBM3e. By leveraging LPDDR5X, Intel can provide massive local capacity—up to 480GB—which is essential for holding large language models directly within the accelerator’s local memory pool, thereby reducing the latency penalty associated with fetching weights from system memory. This combination of a 350W air-cooled PCIe footprint, massive LPDDR5X capacity reaching 480GB, a solid 1.5 TB/s memory bandwidth, and a formidable 1.3 PFLOPS of FP8 theoretical peak performance paints a clear picture of how Intel intends to carve out a distinct and competitive niche in the burgeoning AI inference acceleration market as the year progresses.