Deck: Intel’s inference-focused data center GPU pairs 32 Xe3P cores and 256 XMX engines with unusually high LPDDR5X capacity.[1]
The design is promising, but it remains a pre-release product without public pricing or independent performance results.[1][3][5]
Reporting cutoff: August 25, 2026, 8:38 a.m. Eastern (2026-08-25T12:38:22Z)
Hero disclosure: Original editorial concept; not an official Intel product render.
Intel has disclosed the clearest specifications yet for Crescent Island, its next-generation data center GPU for AI inference.[1]
At Hot Chips 2026, Intel described a 350-watt, air-cooled PCIe card with 32 Xe cores, 256 XMX matrix engines and up to 480GB of LPDDR5X memory.[1]
Hot Chips 2026 runs August 23 through August 25 at Stanford University.[4]
That combination makes Crescent Island noteworthy for a practical reason: it is designed to put a large amount of model data and working context on a conventional add-in card without requiring a liquid-cooled rack.[1][2]
Intel is aiming the accelerator at real-time inference—running trained models to answer requests—not at the far more compute-intensive job of training the largest frontier models.[1][2]
The boundaries matter.
Crescent Island is not a shipping retail product, a gaming card or a benchmarked performance champion.[1][3][5]
It is Intel’s code name for an upcoming inference-focused data center GPU and PCIe accelerator.[1][3]
Intel announced the project at the 2025 Open Compute Project Global Summit, added technical details at Computex 2026 and presented the fuller architecture at Hot Chips 2026.[1][2][3]
Crescent Island specifications Intel has confirmed
Intel’s August 24 Hot Chips summary lists three headline specifications:[1]
- Architecture: Xe3P
- Compute: 32 Xe cores and 256 XMX engines
- Memory: up to 480GB of LPDDR5X
- Power and cooling: 350 watts, air cooled
- Form factor: PCIe add-in card
- Primary target: data center AI inference, especially agentic and long-context workloads
XMX engines are Intel’s matrix-math accelerators.[1][5]
They are the portions of the GPU designed to accelerate the dense multiply-and-accumulate operations used heavily by neural networks.[5]
Intel positions the 32-core, 256-XMX design around sustained token throughput and performance per watt rather than conventional graphics output.[1]
Intel also says Crescent Island supports data types and microscaling formats ranging from native FP4 and MXFP4 through FP64.[2]
Lower-precision formats such as FP4 can reduce the memory and compute needed for inference when a model and software stack are prepared to use them.[2][5]
FP64 support broadens the arithmetic range, but Intel has not published application benchmarks establishing Crescent Island’s performance in scientific computing.[2][5]
Why 480GB of memory is the headline
Large language models do not consume memory only for their weights.[1][5]
They also need memory for working state, including the key-value cache used to preserve conversation context and serve concurrent requests.[5]
Longer context windows, larger models and more simultaneous agents can therefore increase capacity pressure even when raw arithmetic is not the only bottleneck.[1][5]
Intel says the card’s high memory capacity is intended to support larger models, longer contexts and more concurrent AI agents inside existing air-cooled data center footprints.[1]
At Computex, the company described up to 480GB for “large, token-intensive workloads.”[2]
The memory story has evolved.[1][3]
Intel’s October 2025 announcement specified 160GB of LPDDR5X.[3]
Its 2026 disclosures now advertise up to 480GB.[1][2]
Reporting from the Hot Chips presentation says Intel’s own branded card is expected with 160GB while original-design-manufacturer configurations may scale to 480GB; Intel’s public newsroom summaries do not explicitly document that branded-card-versus-partner split.[5]
This is a capacity-first design choice.[1][2][6]
LPDDR5X is widely used for efficient memory systems, while the leading high-end AI accelerators typically use high-bandwidth memory.[6]
Data Center Dynamics notes that LPDDR5X offers less bandwidth than HBM.[6]
Intel has not disclosed a final Crescent Island memory-bandwidth figure in its public newsroom material, so the 480GB capacity number should not be interpreted as proof that the card will outrun HBM-equipped competitors.[1][5][6]
A conventional deployment target
Crescent Island’s 350-watt PCIe design is meant to fit air-cooled enterprise servers rather than require a proprietary liquid-cooled rack architecture.[1][2]
That could lower adoption friction for organizations with suitable PCIe servers, available power and cooling headroom.[1][2]
“Easy to deploy,” however, is Intel’s positioning—not an independent installation finding.[1]
Final compatibility will depend on server slot dimensions, power delivery, firmware, supported operating systems, drivers and validated OEM configurations.[1][5]
Intel has not published a complete compatibility list, exact card dimensions or general-availability product page at this reporting cutoff.[1][3][5]
Detailed reporting from Intel’s Hot Chips presentation adds a PCIe Gen5 x16 interface, a 32MB unified L2 cache, four media decoders and four encoders.[5]
The same report describes error-detection and reliability features including ECC or parity on key memories, page offlining and PCIe Advanced Error Reporting.[5]
Those details were shown in Intel presentation material as reported from the event, but they are not all repeated in Intel’s short public newsroom specifications.[1][5]
What Xe3P changes
Intel identifies Xe3P as the architecture underlying Crescent Island.[1][2]
ServeTheHome’s account of the Hot Chips slides describes a 16-deep XMX systolic array, 1MB general register file and 512KB of L1 or shared local memory per Xe core, along with the 32MB unified L2 cache.[5]
Video-oriented fixed-function blocks remain in the reported design, while another account of Intel’s slides says Crescent Island removes traditional 3D and ray-tracing hardware to devote more silicon to compute.[5]
Because Intel’s public newsroom summary does not state that exclusion directly, it should be treated as event-slide reporting rather than a complete official product specification.[1][5]
The result is best understood as a purpose-built compute GPU.[1][2][5]
Intel calls it a GPU because it derives from the Xe architecture and contains Xe cores, but the product target is accelerator-class AI inference in servers—not display output or gaming.[1][2][5]
The software stack may be as important as the silicon
AI accelerators succeed only when models and frameworks can use them reliably.[2][5]
Intel says its open, programmable AI software stack follows an upstream-first approach and is intended to provide out-of-the-box model support.[2]
The company also positions Arc Pro GPUs as a development platform where software can be built and optimized before deployment to Crescent Island, with forward and backward compatibility across the Xe foundation.[2]
ServeTheHome’s report from the Hot Chips session lists vLLM, SGLang, llm-d and NVIDIA Dynamo among the orchestration or serving frameworks discussed.[5]
It also identifies Triton, SYCL-TLA, oneCCL, oneDNN, SYCL, Level Zero and an OpenCL compute runtime in the software stack.[5]
Those names show Intel’s intended direction, not proof that every framework feature is production-ready on unreleased Crescent Island hardware.[2][5]
Buyers will need validated versions, model coverage, operational tooling and measured reliability—not only a framework logo on a presentation slide.[2][5]
When will Crescent Island be available?
Intel’s original October 2025 announcement said customer sampling was expected in the second half of 2026.[3]
The August 2026 Hot Chips update provides architecture and card specifications but does not announce general availability, broad shipments, pricing or a final commercial product name.[1]
That leaves several distinct milestones unresolved:
- whether customer samples are already in partners’ hands;
- when production hardware will ship;
- which OEMs will offer 160GB or higher-capacity configurations;
- whether 480GB is available at launch or later;
- pricing, warranty and support terms;
- final driver and framework availability.
Until Intel confirms those milestones, Crescent Island should be described as upcoming, disclosed and targeted for sampling—not generally available.[1][3]
What Intel has not shown yet
Intel’s public disclosures do not provide enough evidence for a performance verdict.[1][5][6]
Missing or unresolved items include:
- measured tokens per second and tokens per watt;
- latency at defined batch sizes and context lengths;
- final memory bandwidth;
- clock speeds and peak arithmetic throughput by data type;
- manufacturing process and die size;
- multi-card scaling performance;
- pricing and total system cost;
- independent benchmarks against current NVIDIA, AMD or specialized inference hardware.
Data Center Dynamics reported after Computex that Intel had not yet supplied performance figures.[6]
ServeTheHome’s Hot Chips analysis likewise identified memory bandwidth and real tokens-per-watt results as open questions.[5]
Intel’s claims about better inference economics, utilization and return on investment are therefore vendor projections.[1]
They may be reasonable design goals, but they are not independently verified outcomes.[1][5][6]
Why Crescent Island matters
The design tests a different tradeoff from the highest-power accelerators.[1][2][6]
Instead of maximizing every dimension with HBM and dense rack-scale cooling, Intel is emphasizing large memory capacity, a 350-watt air-cooled envelope, standard PCIe deployment and an open software path.[1][2]
If the software matures and measured performance is competitive for the intended workloads, Crescent Island could give enterprises another way to host large or long-context models without rebuilding an entire data center around liquid-cooled systems.[1][2][5]
The large capacity could also reduce model partitioning across cards for some workloads, though that benefit depends on model format, precision, KV-cache needs and actual memory bandwidth.[1][5]
The most defensible conclusion today is narrower: Intel has now defined Crescent Island as a high-capacity, air-cooled Xe3P inference accelerator, but its commercial and measured-performance story is still unfinished.[1][3][5]
What IT leaders should watch next
Before putting Crescent Island on a procurement shortlist, ask for:
- Independent benchmarks using the same model, precision, context, batch size and quality target.
- Sustained—not peak—tokens per second, latency and power measurements.
- Final memory capacity and bandwidth for the exact card being quoted.
- A validated server, driver, firmware and operating-system compatibility matrix.
- Framework and model support demonstrated on shipping hardware.
- Pricing, support terms, availability and multi-card scaling details.
Crescent Island’s 480GB headline is technically interesting.[1][2]
The buying decision will depend on the numbers Intel has not published yet.[1][5][6]
Sources
[1] Intel Outlines Architectures for Agentic AI at Hot Chips 2026
32 Xe cores and 256 XMX engines based on Xe3P
Up to 480GB LPDDR5X memory
350-watt air-cooled PCIe card
[2] Intel Puts Agentic AI to Work with Xeon 6+, Networking, and AI Systems
Equipped with LPDDR5x memory, Intel’s Crescent Island delivers up to 480 GB capacity to efficiently handle large, token-intensive workloads while reducing total cost of ownership.
Leveraging a multi-generational Xe install base, Intel’s Crescent Island is designed for next generation AI workloads with support for a wide range of datatypes and microscaling formats, from native FP4/MXFP4 to FP64, including expanded support for advanced AI operations and improved memory and scalability.
[3] Intel to Expand AI Accelerator Portfolio with New GPU
160GB of LPDDR5X memory
Customer sampling of the new data center GPU code-named Crescent Island is expected in the second half of 2026.
Hot Chips 2026 will be held Sunday, August 23 – Tuesday, August 25, 2026 at Memorial Auditorium, Stanford, Palo Alto, CA.
[5] Intel Crescent Island 160GB to 480GB LPDDR5X AI GPU at Hot Chips 2026
Crescent Island is Xe3p, and the generation table shows how far the design has come from Battlemage. Where Xe2 delivered 20 Xe cores and a 4-deep systolic XMX, Crescent Island brings 32 Xe cores, a 16-deep systolic array, a 1MB GRF and 512KB L1 per Xe core, a 32MB unified L2, and up to 480GB of LPDDR5x.
Still, how much memory bandwidth does this have? I wish Intel went into more detail.
[6] Intel teases Crescent Island AI data center GPU
LPDDR5X has lower bandwidth than HBM.
Intel has yet to compute performance figures. Customer sampling of the Crescent Island GPU is targeted for the second half of 2026.
Specifications and availability were checked through August 25, 2026. Crescent Island was not generally available at the reporting cutoff; vendor projections are not independent benchmarks.