Intel Details Crescent Island Xe3P Inference GPU at Computex, Scaling to 480GB of LPDDR5X Without HBM
Intel's data-center inference GPU scales to 480GB of LPDDR5X in a 350W air-cooled PCIe card, sidestepping the HBM shortage that constrains Nvidia and AMD accelerators.
Overview
Intel used Computex 2026 in Taipei to flesh out the specifications of Crescent Island, its forthcoming data-center GPU for AI inference, confirming that the design scales to 480GB of LPDDR5X memory in a single air-cooled card. According to Tom’s Hardware, Intel detailed the Xe3P inference accelerator at the show, framing its unusually large memory pool as a way to combat the memory shortages squeezing the AI hardware market.
The central design choice is what Crescent Island leaves out. Where Nvidia and AMD accelerators rely on high-bandwidth memory (HBM), Intel built its inference card around LPDDR5X — the commodity, lower-cost memory found in laptops and phones — trading peak bandwidth for far greater capacity and supply resilience. As The Register put it, “Intel has opted for LPDDR5x memory — the same kind used in high-end notebooks and smartphones — and quite a bit of it too.”
What We Know
Crescent Island is built on Intel’s Xe3P architecture. Tech Times reported that the card “is built on Intel’s Xe3P architecture” and is positioned as a part “built for agentic AI.”
On memory, the reference design starts at 160GB and scales upward. Per Tech Times, the part “supports up to 480 gigabytes of LPDDR5X memory” and “starts at 160GB.” The same report describes a “640-bit memory interface running LPDDR5X at 10.7 Gbps,” yielding “684 GB/s of memory bandwidth.” Tom’s Hardware likewise reports that the “reference design will include 160GB of LPDDR5X, but that the chip is designed to allow partners the flexibility to build accelerators with up to 480GB of memory,” and that “with 10.7 Gbps LPDDR5X, Crescent Island would offer 684 GB/s of memory bandwidth.”
The card targets mainstream enterprise infrastructure rather than liquid-cooled supercomputers. Tom’s Hardware reports that “Crescent Island will be a PCI Express add-in card with a 350W power target,” and Tech Times notes a “350W power envelope that fits standard air-cooled servers,” adding that “the card operates as a standard PCIe device.” The Register confirms it “will ship as a 350 watt Air-cooled PCIe card.”
On precision, Crescent Island casts a wide net. Tech Times reports the architecture “spans a wide range of data types, from FP4…all the way up to FP64,” with FP4 aimed at high-performance inference and FP64 at scientific computing.
The memory pool also dwarfs the competition’s. According to The Register, “Crescent Island will be offered with up to 480 GB of memory, significantly more than you’ll find on Nvidia’s flagship GPUs, which currently top out at 288 GB.”
The motivation is the memory market. Tech Times noted that the “HBM supply chain that powers Nvidia’s accelerators is sold out through at least 2027,” while “LPDDR5X is a commodity memory technology…not subject to the same concentrated supply chain.” That commodity-memory bet mirrors a strategy The Machine Herald has covered elsewhere: startup Positron AI built its inference accelerator around LPDDR5X rather than HBM for the same capacity-over-bandwidth reasons, as previously reported.
Background
Crescent Island is not new. Intel first revealed the GPU at the 2025 OCP Global Summit on October 14, 2025, according to Intel’s newsroom, describing it then as built on the “Xe3P microarchitecture with optimized performance-per-watt” with “160GB of LPDDR5X memory” and “power and cost-optimized for air-cooled enterprise servers.” At that launch, Sachin Katti, CTO of Intel, framed the inference shift directly: “AI is shifting from static training to real-time, everywhere inference—driven by agentic AI,” according to Intel. The Computex disclosure extends that picture, confirming the up-to-480GB ceiling and the 350W air-cooled card details.
The LPDDR5X approach lands against a backdrop of acute memory pressure. The Machine Herald has previously reported on the GDDR7 and LPDDR5X shortage rippling through consumer markets even as AI chip revenue booms — a structural squeeze on how the industry allocates memory capacity.
What We Don’t Know
Intel has not disclosed raw compute throughput for Crescent Island, leaving the part’s FLOPS, chip configuration, and real-world inference performance unquantified. The company also has not detailed pricing or named launch customers.
The timeline remains a preview rather than a ship date. Tech Times describes Intel as “targeting H2 2026 sampling,” and Intel said at the original reveal that “customer sampling” was “expected in the second half of 2026.” Whether Crescent Island’s capacity advantage translates into competitive cost-per-token against entrenched HBM-based accelerators will not be clear until independent benchmarks arrive.
Analysis
Crescent Island reframes the inference contest around capacity and supply rather than peak bandwidth. Large language and agentic models are frequently constrained by how much of a model fits in memory, not by how fast data moves through the compute pipeline — the bottleneck LPDDR5X addresses well. By stacking up to 480GB on a single air-cooled PCIe card, Intel is targeting buyers who want to run large models locally without provisioning liquid cooling or competing for HBM allocation that, by Intel’s own framing via Tech Times, is “sold out through at least 2027.”
The tradeoff is bandwidth: 684 GB/s is a fraction of what HBM4-class accelerators deliver, which makes Crescent Island a poor fit for training and a deliberate bet that the inference market is large enough, and capacity-hungry enough, to reward a different memory architecture. With compute figures still undisclosed and sampling not due until the second half of 2026, the strategy is clear well before the silicon is provable.