Amid a flurry of product announcements and updates at this year’s Open Compute Project Global Summit, Intel is at the show to deliver an update on Crescent Island, the company’s forthcoming AI inference GPU. After outlining the chip’s architecture back at Hot Chips 2026, the company is offering a short update on the product today, outlining some of the chip’s final performance specifications along with giving the cards based on it a proper name. Going forward, Crescent Island will be sold as part of Intel’s GPU D-Series.
Speeds and Feeds: Crescent Island Nears Release
As a brief refresher, Crescent Intel’s is first commercial AI-focused data center GPU since Ponte Vecchio in 2021. With two of Ponte’s successors canceled over the last few years – Rialto Bridge in 2023 and then Falcon Shores in 2025 – Intel announced Crescent Island at least year’s OCP Global Summit as an interim product to target the AI inference market with a heavily cost and power-optimized part. The resulting chip is not meant to be on the same performance level as flagship chips such as NVIDIA’s Blackwell Ultra or AMD’s MI455X, but then it will be a lot smaller and cheaper as well.
Most recently, Intel delivered an update on Crescent Island back in August, where the company offered a fairly detailed look at the architecture behind the chip. Based on an enhanced version of their Xe3 GPU architecture that they are calling Xe3P, Crescent Island heavily fortifies the architecture with a far larger number of XMX matrix engines, and a far larger LPDDR5X memory subsystem to keep those matrix blocks fed. In effect, Intel is banking on a massive matrix math version of Xe3 that is backed by a large LPDDR5X memory pool of up to 480GB being able to compete favorably with other inference accelerators on the market today.
For this year’s OCP Global Summit, in turn, Intel is finally publicly locking down some of the specifications for the forthcoming accelerator. In particular, we finally have the theoretical total compute and total memory bandwidth figures for the first card based on Crescent Island, which gives us enough information to paint a fairly detailed look at how the chip is configured.
| Intel GPU D-Series Specifications | |
| Intel GPU D270 | |
| Xe Cores | 32 |
| XMX Engines | 256 |
| GPU Clockspeed | ~2.5GHz |
| Peak Matrix Throughput (FP8) | 1.3 PFLOPS |
| Peak Vector Throughput (FP64) | 10.2 TFLOPS |
| Memory (Max) | 480GB LPDDR5X (ODMs) 160GB LPDDR5X (Intel) |
| Memory Bandwidth | 1.5TB/sec (~9375Mbps x 1280-bits) |
| ECC | Yes |
| L2 Cache | 32MB |
| I/O | PCIe Gen5 x16 |
| TDP | 350W |
| Form Factor | PCIe CEM (Air Cooled) |
| Architecture | Xe3P |
| GPU | Crescent Island |
In regards to performance, Intel is touting 1.3 PFLOPs of FP8 matrix performance. With 256 XMX matrix engines on the chip itself, this gives us a clockspeed of roughly 2.5GHz.
As for memory bandwidth, the flagship card will offer 1.5TB/second of memory bandwidth, which comes courtesy of its hefty 20 chip, 80 channel LPDDR5X memory bus, with the memory clocked at roughly 9375MT/second.
While Intel had previously discussed the architecture of the chip, the total performance and bandwidth had been up in the air. In practice, these figures are very close to what most of the industry has been expecting for the accelerator given current LPDDR5X memory speeds as well as the kind of clockspeeds that similar accelerator architectures have been hitting. In that respect Intel is delivering right in the neighborhood of where they should be – meaning that Crescent Island is not surpassing expectations, but it is not failing to meet them, either.
As to whether 1.3 PFLOPs of FP8 will be sufficient for Intel’s target market of data center inference operations, that remains to be seen. For comparison’s sake, AMD’s similar PCIe CEM form factor MI350P accelerator can offer a theoretical peak of 2.3 PFLOPs of dense FP8 matrix math; so a Crescent Island card comes in some 40% lower. With that said, comparing different architectures based on theoretical throughput alone rarely pans out in the real world, so a lot of the GPU’s competitiveness is going to hinge on its overall architectural efficiency compared to AMD’s CDNA 4 architecture. Especially if Intel can do meaningfully better in prefill, which is traditionally where GPU-based architectures bottleneck in data center-scale workloads.
The bigger differentiator there will likely be power and cooling. The flagship Crescent Island card is a 350 Watt card that can be air cooled, whereas MI350P’s highest performance configuration is 600 Watts. Even more so than inference performance, Intel has been touting Crescent Island’s efficiency on a performance-per-watt basis, so for Intel’s intended market the card does not necessarily need to be the fastest option, so long as it is a more power efficient option than other PCIe cards.
Finally, Intel has once again confirmed that we will see multiple memory configurations with Crescent Island cards. The company’s own card will ship with 160GB of LPDDR5X memory (20x8GB chips). However, ODM partners will be able to offer versions with up to 480GB of memory, an extreme amount of memory enabled by 24GB LPDDR5X chips and a far larger capacity than any other PCIe form factor accelerator announced thus far.
Meanwhile, Crescent Island cards are finally getting a series name. For Intel’s new series of data center AI inference cards the company is launching a new brand name for them: the Intel GPU D-Series.
The first (and so far, only) card in the series is the GPU D270, pictured above. The image, a render, does not reveal anything we do not already know about Intel’s target market for the cards. But it does reinforce that this is a double-wide PCIe card meant to go into traditional PCIe-based data center servers.
Notably, Intel is not keeping the Data Center GPU Max (or Flex) branding here, which is what Intel has used for its Ponte Vecchio and Arctic Sound-based PCIe cards. Despite being aimed at much of the same HPC market as the Data Center GPU Max family (and effectively its successor), the GPU D-series is its own product through and through.
The D-series suffix does invite some comparisons to another Intel D-named product, the Xeon D lineup. However, those chips were aimed at networking and edge devices, whereas the 350 Watt GPU D270 is very much a data center part. Plus the Xeon D brand was effectively put on a retirement glide path in 2025 with the launch of the Xeon 6 SoC series of chips (Granite Rapids-D), which are replacing Xeon D in that market. So the best way to interpret the new name is that Intel is starting anew, at least as far as GPU hardware goes.
Final Words
Wrapping things up, with today’s announcement Intel is slowly laying the remaining groundwork for the full launch of the GPU D-series. At this point the company has all but fully revealed its specifications and theoretical performance – now it is just a matter of when it will ship, and how much Intel will charge for it.
But the answers to those questions will have to wait for another day. For now, Intel is confirming that they are sampling cards to customers. However the company has not yet announced a launch date for the cards, or even confirmed that they will launch in what remains of this year. So AI server vendors looking for an alternative for PCIe inference cards will have to look towards 2027 for the release of Intel’s GPU D-series.