Intel has put an official name on the AI chip it teased a year ago. On October 13, 2026, the company confirmed that its “Crescent Island” accelerator now ships as the Intel GPU D-Series, a PCIe-based inference card built on the Xe3P architecture. According to Tom’s Hardware, the chip delivers up to 1.3 peak petaflops of FP8 compute and 1.5 TB/s of memory bandwidth, backed by as much as 480GB of LPDDR5X memory, all inside a 350-watt, air-cooled card. The announcement landed at the 2026 OCP Global Summit, where Intel laid out a broader pitch for what it calls open infrastructure for the “agentic AI era.”
For a company that has spent three years watching Nvidia and AMD split the data-center AI market between them, the GPU D-Series is less a moonshot than a calculated bet on a different problem: inference, not training. Intel is wagering that enterprises running large language models in production care more about memory capacity and deployment economics than raw peak compute, and the spec sheet reflects that logic almost entirely.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Intel Actually Announced
Crescent Island was first disclosed as a code name roughly a year earlier, and Intel used the OCP Global Summit to confirm it as a shipping product line rather than a research project. The official product name is Intel GPU D-Series, built on the Xe3P architecture, according to Intel’s own newsroom post. The card is designed as an air-cooled PCIe product, which matters because it means data centers can deploy it in standard server chassis without the liquid-cooling infrastructure that higher-wattage accelerators typically require.
Intel’s stated target workloads are AI inference tasks, particularly memory-intensive inference involving large language models, long-context processing, and what the company describes as agentic AI serving, where multiple models or tool calls run concurrently against a shared pool of memory. That emphasis on inference over training is deliberate. Intel is not claiming the GPU D-Series competes with the largest training clusters built around Nvidia’s or AMD’s top-tier accelerators. Instead, it is positioning the card as the economical option for running already-trained models at scale.
Sampling for customers was expected in the second half of 2026, based on Intel’s original disclosure a year prior, though Thursday’s announcement confirms the product has now moved from code name to named SKU. Pricing has not been disclosed, and Intel has not confirmed a specific model number for the lead product, despite some secondary reports referencing a “GPU D270” designation.
The Spec Sheet: 1.3 PFLOPS, 1.5 TB/s, 480GB
The headline numbers are straightforward. Peak FP8 compute tops out at 1.3 petaflops, a theoretical ceiling rather than a benchmarked result, according to StorageReview. Memory bandwidth is rated at 1.5 TB/s, and the card can carry up to 480GB of LPDDR5X memory, a capacity figure that dwarfs most HBM-based accelerators on the market today. All of this runs within a 350-watt power envelope, which is modest by data-center GPU standards and is explicitly meant to allow deployment in air-cooled racks rather than the liquid-cooled infrastructure Nvidia’s and AMD’s flagship accelerators increasingly require.
The choice of LPDDR5X over High Bandwidth Memory is the single most distinctive engineering decision in the announcement. LPDDR5X is cheaper to produce, more available, and draws less power per gigabyte than HBM3 or HBM4, but it also delivers substantially lower bandwidth per pin. Intel appears to be betting that for inference workloads bound by memory capacity rather than raw throughput, that tradeoff works in its favor.
| Specification | Intel GPU D-Series (Crescent Island) |
|---|---|
| Architecture | Xe3P |
| Peak FP8 compute | Up to 1.3 PFLOPS |
| Memory bandwidth | 1.5 TB/s |
| Maximum memory capacity | Up to 480GB LPDDR5X |
| Power envelope | 350 watts |
| Form factor | Air-cooled PCIe |
| Target workload | Memory-intensive AI inference |
| Customer sampling | Second half of 2026 |
| Pricing | Not disclosed |
Why Intel Is Betting on Memory Capacity, Not Just Speed
The logic behind the GPU D-Series only makes sense once you separate training from inference as distinct computational problems. Training a frontier model demands enormous, sustained compute throughput across thousands of chips working in lockstep, which is why Nvidia’s and AMD’s top accelerators pair extreme compute density with HBM memory that can feed data to cores fast enough to keep them busy. Inference is a different problem. Once a model is trained, serving it to users mostly means holding its weights, along with a growing context window of conversation history or document data, in memory that is fast enough to respond without excessive latency.
According to Intel Vice President Anil Nanduri, “Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost,” a comment made as Intel unveiled the GPU D-series at the OCP Global Summit, according to Pulse2. Nanduri, who leads AI Products and Go-to-Market for Intel’s Data Center Group, also said, “With Intel GPU D-series, we’re focused on making high-capacity inference practical to deploy and economical to scale,” underscoring that the product’s value proposition rests on deployment cost rather than peak throughput.
That framing matters because it is an implicit admission that Intel is not trying to out-compute Nvidia or AMD on their own turf. It is instead trying to carve out a category where a different set of tradeoffs, cheaper memory, lower power, air cooling, wins on total cost of ownership even if it loses on raw FLOPS. Whether that argument holds up depends entirely on how inference workloads actually behave in production, something that varies enormously by model size, context length, and traffic pattern.
Scaling From 300 Billion to Over a Trillion Parameters
Intel’s pitch includes a specific scaling story: a single GPU D-Series card is described as capable of handling models up to roughly 300 billion parameters, while an eight-GPU rack-scale configuration is aimed at models exceeding one trillion parameters, according to the Perplexity-sourced research citing Intel’s OCP presentation. That rack-scale configuration is central to Intel’s “open, heterogeneous infrastructure” argument, where GPU D-Series cards work alongside Intel’s Xeon processors rather than forming a closed, proprietary stack.
Kevork Kechichian, Executive Vice President and General Manager of Intel’s Data Center Group, framed the strategy in broader terms: “Agentic AI is changing what customers need from their data centers,” he said, adding that “success will depend not on any single component, but on open, heterogeneous infrastructure that brings together Xeon, purpose-built acceleration and interoperable software,” according to Pulse2’s coverage of the OCP Global Summit announcement. That statement is effectively Intel’s answer to Nvidia’s tightly integrated CUDA ecosystem: rather than compete chip-for-chip, Intel wants customers to buy into a platform story that spans CPUs, GPUs, and software.
How the GPU D-Series Compares to Nvidia and AMD
Direct, apples-to-apples benchmarking against Nvidia’s or AMD’s current accelerators has not been published, and reporting on the announcement is explicit that no independent comparative performance results have been disclosed. What can be compared is philosophy and positioning rather than raw numbers. Nvidia’s data-center GPU lineup, which dominates both AI training and inference deployments today, pairs very high compute density with HBM-based memory systems that prioritize bandwidth over capacity. AMD’s Instinct accelerators follow a similar pattern, also leaning on large HBM pools to compete directly with Nvidia on bandwidth-sensitive workloads.
Intel’s approach inverts that priority. By using LPDDR5X instead of HBM, the GPU D-Series trades bandwidth for capacity and, by extension, cost and power efficiency. That is a meaningful point of differentiation in a market where HBM supply has been tight enough to reshape pricing across the entire memory industry; Tech Insider has previously reported on how HBM4 shortages pushed prices up 117% as AMD and others chased supply. If Intel can deliver usable inference performance without relying on HBM at all, it sidesteps one of the biggest supply bottlenecks currently constraining the entire AI hardware industry.
| Dimension | Intel GPU D-Series | Nvidia data-center GPUs | AMD Instinct accelerators |
|---|---|---|---|
| Memory type | LPDDR5X | HBM3/HBM3e-class | HBM3-class |
| Design priority | Memory capacity, low power | Peak compute, bandwidth | Bandwidth, compute parity with Nvidia |
| Cooling approach | Air-cooled PCIe | Increasingly liquid-cooled | Mixed air/liquid |
| Primary target | Memory-intensive inference | Training and inference | Training and inference |
| Software ecosystem | Emerging, open infrastructure focus | CUDA, mature and dominant | ROCm, improving but smaller |
| Disclosed independent benchmarks | None published | Extensive, third-party verified | Extensive, third-party verified |
The biggest variable in that table is the last row. Nvidia’s dominance is not just architectural, it is also a function of CUDA’s maturity and the sheer volume of third-party benchmarking and real-world deployment data available for its chips. Intel’s GPU D-Series has none of that yet, and building it will take time, developer tooling, and trust from enterprise buyers who have spent years optimizing around Nvidia’s stack.
Intel’s Long Road Back to AI Relevance
The GPU D-Series does not arrive in a vacuum. Intel has tried, and largely struggled, to establish a foothold in AI acceleration for years. Its Habana Labs acquisition produced the Gaudi line of AI training chips, which saw limited enterprise adoption despite aggressive pricing against Nvidia. Its Data Center GPU Max series, codenamed Ponte Vecchio, launched with high expectations for supercomputing deployments like Aurora at Argonne National Laboratory but never built meaningful commercial momentum in the broader AI inference market. Against that backdrop, the GPU D-Series represents at least a strategic pivot: rather than chasing Nvidia on training performance, Intel is trying to define a new category around inference economics where its chip design philosophy, air cooling, LPDDR5X, modest power draw, can plausibly be a selling point rather than a limitation.
That pivot also lines up with a broader industry shift Tech Insider has tracked closely. Microsoft’s own silicon strategy reflects similar inference-first thinking; the company’s Maia 300 chip claims a 40% edge over Nvidia on certain workloads by targeting specific inference use cases rather than general-purpose training. Data center operators more broadly have also been forced to rethink memory strategy as HBM supply constraints bite, a dynamic that shows up in Nvidia, AMD, Google, and Meta’s joint data center spec v0.7, an effort aimed at standardizing infrastructure across competing chip vendors.
Market Impact: What This Means for the AI Chip Race
Intel’s data-center GPU business has been a tiny fraction of Nvidia’s in recent years, and that is unlikely to flip overnight because of one product announcement. What the GPU D-Series does change is the competitive narrative. Instead of a three-way race defined entirely by who can pack the most HBM onto a chip, Intel has introduced a fourth axis: can you serve the same inference workload more cheaply by trading bandwidth for capacity. If that argument gains traction among cloud providers and enterprises running cost-sensitive inference at scale, it pressures Nvidia and AMD to defend their pricing on inference-specific SKUs rather than letting training-class chips double as the default inference option.
It also matters for the memory supply chain. A credible, high-volume accelerator that uses LPDDR5X rather than HBM would reduce one source of demand pressure on an HBM market that has already seen dramatic price increases this year. Tech Insider has also reported on how Nvidia’s own China market share has slipped as competitors gain ground, with Nvidia’s China AI chip share dropping to 34% as Huawei climbed to 28%, a reminder that Nvidia’s dominance, while still overwhelming globally, is not uncontested in every market or every workload category.
The Reference Card Question
One detail worth flagging for anyone evaluating the GPU D-Series closely: Intel’s 480GB figure is a maximum capacity, not necessarily what ships on day one. Reporting cited in the research for this story indicates the reference PCIe card may initially carry a lower 160GB configuration, with higher-capacity 480GB versions potentially left to original design manufacturer partners to build out. That distinction matters for buyers comparing on-paper specs against what they can actually order at launch, and it is a detail Intel has not directly clarified in its own newsroom materials.
Similarly, the specific model number “GPU D270,” which has circulated in some secondary coverage, does not appear in Intel’s own announcement materials and should be treated as unconfirmed until Intel publishes official product nomenclature.
Historical Context: Intel’s Pattern of Late, Ambitious AI Bets
Intel’s AI accelerator history is a story of being early to announce and late to ship at competitive volume. Gaudi 2 and Gaudi 3 arrived with aggressive price-performance claims against Nvidia’s H100, yet enterprise adoption remained limited outside a handful of large customers. Ponte Vecchio, Intel’s most ambitious GPU compute effort before Crescent Island, was delayed multiple times before shipping in a narrower form than originally planned. The GPU D-Series arrives in a market that has only grown more concentrated around Nvidia since those earlier efforts, with AMD as the clear second player and everyone else, including Intel, fighting for a distant third position.
What is different this time is the narrower, more defensible positioning. Rather than trying to be a general-purpose Nvidia alternative, Intel has picked a specific niche, memory-intensive inference serving, where its design tradeoffs can be framed as advantages rather than compromises. That is a more modest ambition than previous Intel AI chip launches, and modest ambitions are sometimes easier to actually deliver on.
What Comes Next: Five Predictions
- Intel will likely disclose pricing closer to the second-half 2026 sampling window, and early pricing will probably undercut comparable Nvidia and AMD inference SKUs on a per-gigabyte-of-memory basis, mirroring Intel’s historical playbook with Gaudi.
- Independent benchmarks will emerge within months of sampling, and the real test will be tokens-per-second-per-dollar on production-scale language models rather than the peak FP8 figure Intel has published.
- Software maturity will be the biggest adoption bottleneck. Without a CUDA-equivalent ecosystem, enterprise buyers will likely wait for proof of easy integration with existing inference frameworks before committing meaningful budget.
- Expect cloud providers, rather than individual enterprises, to be the first major GPU D-Series customers, since hyperscalers have the engineering resources to validate a new accelerator architecture that smaller companies typically lack.
- If the LPDDR5X approach proves viable at scale, expect Nvidia and AMD to respond with their own lower-cost, capacity-focused inference SKUs rather than ceding the category, much as both companies have historically matched competitive threats with parallel product lines.
Risks and Open Questions
Several unknowns remain that will determine whether the GPU D-Series becomes a meaningful product or another Intel AI chip that generates headlines without volume. Pricing is entirely undisclosed, and inference economics only work if the per-card cost is genuinely competitive once memory capacity is accounted for. Software support is another open question: Intel’s oneAPI and related tooling have improved but still lag the breadth of frameworks that run natively and efficiently on Nvidia’s CUDA stack. And the gap between the 160GB reference configuration reportedly shipping first and the 480GB maximum Intel is promoting could create confusion among buyers evaluating the product against competitors with simpler, single-configuration lineups.
There is also the broader question of whether “memory capacity over bandwidth” is actually the right tradeoff for the inference workloads enterprises are running today. Long-context applications and multi-tenant agentic AI serving, the exact use cases Intel is targeting, are precisely the workloads where insufficient bandwidth could create latency problems that offset the benefit of having more memory available in the first place.
FAQ: Intel GPU D-Series and Crescent Island
What is Intel’s Crescent Island chip now called?
Crescent Island is the former code name; the shipping product is officially named the Intel GPU D-Series.
What is the peak compute performance of the Intel GPU D-Series?
Intel states up to 1.3 peak petaflops of FP8 compute, a theoretical maximum rather than a published benchmark result.
How much memory does the Intel GPU D-Series have?
Up to 480GB of LPDDR5X memory, though some reporting suggests the initial reference card may ship with a lower 160GB configuration.
What architecture powers the GPU D-Series?
Intel’s Xe3P architecture, the company’s next-generation GPU design built specifically for data-center workloads.
How much power does the chip use?
The GPU D-Series operates within a 350-watt envelope and uses air cooling rather than liquid cooling, which simplifies deployment in standard data-center racks.
When can customers buy the Intel GPU D-Series?
Intel has said customer sampling is expected in the second half of 2026. Pricing and general availability dates have not been disclosed.
Is the Intel GPU D-Series meant to compete with Nvidia’s top AI chips?
Not directly. Intel is positioning the product for memory-intensive inference workloads rather than the large-scale training tasks where Nvidia’s highest-end accelerators dominate.
Why did Intel choose LPDDR5X instead of HBM memory?
LPDDR5X is cheaper and more power-efficient per gigabyte than HBM, though it offers lower bandwidth. Intel is betting that for inference workloads bound by memory capacity, that tradeoff is worthwhile, and it also sidesteps the HBM supply shortages affecting the broader industry.
Related Coverage
- Build a Mini-ITX Gaming PC: 13 Steps, 310mm GPU Fit [2026]
- Honda Super-ONE Debuts at $36,900 With Ohme Charging [2026]
- Google Fitbit Edge Debuts at $179.99, Ships Oct. 27 [2026]
- Intel CPU Prices Jump Up to 20% as AMD Preps Hikes [2026]
- Build a Gaming PC: 13 Steps, Cut GPU Temps 19°C [2026]
