480GB LPDDR5X 350W air-cooled envelope
Intel has officially put a commercial name on its eagerly awaited Crescent Island architecture, unveiling the Intel GPU D-Series alongside its inaugural accelerator: the Intel GPU D270. While NVIDIA and AMD remain locked in an arms race over kilowatt-guzzling, liquid-cooled HBM racks, Intel is taking a decidedly pragmatic approach to the inference market. The company is trading brute-force memory bandwidth for massive, cost-effective memory capacity that drops straight into existing air-cooled PCIe servers.
At Computex, I had a chance to chat with Anil Nanduri, Vice President and Chief AI Product, GTM. While that wasn’t the right time to talk about Crescent Island, the OCP Summit proved to be the ideal stage for the launch. Customer sampling for the GPU D270 is officially underway, and a closer look at the silicon shows exactly how Intel plans to carve out a lucrative niche in the agentic AI and large-model serving space.
Xe3P Architecture: All Matrix, Zero Ray Tracing
At the heart of the GPU D270 sits Intel’s updated Xe3P architecture. While Xe3P will eventually trickle into future client and PC products—with the current generation of Intel consumer products using Xe3 for Panther Lake, the Intel Core Ultra Series 3—the Crescent Island implementation is a purebred data center compute engine (reminiscent of how Chipzilla’s lost Arctic Sound accidentally turned up super late).
Intel’s engineers took a scalpel to the traditional GPU die, completely stripping out conventional 3D graphics pipelines and dedicated ray-tracing hardware to reclaim valuable silicon real estate for dense matrix math. The resulting chip packs 32 Xe3P cores and 256 XMX (Xe Matrix Extensions) engines, backed by 32MB of unified L2 cache.
Running at a conservative ~2.5 GHz clock to stay within the optimal efficiency curve, the D270 delivers up to 1.3 PFLOPS of FP8 matrix throughput alongside a modest 10.2 TFLOPS of FP64 vector compute. This isn’t a card built for legacy double-precision weather simulations or rendering frame buffers; it is a purpose-built linear algebra engine designed to chew through transformer weights and long-context KV caches.
The 480GB LPDDR5X Bet: Capacity Over HBM Bandwidth
The headline feature is the memory subsystem. Instead of fighting for expensive, supply-constrained HBM3E or HBM4 stacks, Intel wired Crescent Island with a wide 1280-bit memory interface running LPDDR5X with ECC.
There is an important distinction between Intel’s reference board and what server partners will bring to market:
Intel Reference Card: Ships with 160GB of LPDDR5X. This is an exceptional amount of memory, but Intel is letting partners add significantly more.
ODM Partner Cards: Will offer configurable memory options up to a staggering 480GB of LPDDR5X on a single card (or three times the baseline capacity). This will enable even larger models for your inferencing pleasure.
That 1280-bit bus yields 1.5 TB/s of memory bandwidth. On paper, 1.5 TB/s is a fraction of what multi-terabyte HBM accelerators push, and actually sits right around the bandwidth of a single high-end workstation GDDR7 card. Why does it work?
Because the bottleneck in modern AI inference—especially Mixture-of-Experts (MoE) models, multi-session serving, and massive context windows—is often VRAM capacity, not peak bandwidth.
When only a fraction of parameters are active per token in an MoE architecture, or when an enterprise needs to keep a 300-billion-parameter model resident in memory without sharding it across four or eight $20,000+ GPUs, having up to 480GB on a single board changes the Total Cost of Ownership (TCO) equation overnight. According to Intel, a single D270 can host models up to 300B parameters, while a standard 8-GPU server node can comfortably serve 1-trillion-plus parameter models entirely in local memory.
Intel GPU D270 (“Crescent Island”) Specifications
| Specification | Intel GPU D270 Details |
| Architecture | Intel Xe3P (“Crescent Island” — Compute/Inference Optimized) |
| Compute Units | 32 Xe3P Cores / 256 XMX Matrix Engines (~2.5 GHz) |
| L2 Cache | 32MB Unified L2 Cache |
| Peak AI / HPC Throughput | 1.3 PFLOPS (FP8 Matrix) / 10.2 TFLOPS (FP64 Vector) |
| Memory Capacity | 160GB LPDDR5X (Intel Branded) / Up to 480GB LPDDR5X (ODM Partners) |
| Memory Interface & Bandwidth | 1280-bit Bus (ECC) / 1.5 TB/s |
| Form Factor & Interface | Standard PCIe 5.0 x16 CEM Add-in Card, Passive Air-Cooled |
| Board Power (TDP) | 350W |
350W and Air-Cooled: Built for the Real Data Center
Perhaps the smartest piece of engineering on the GPU D270 is its 350W thermal envelope.
Hyperscalers might be happy ripping out data center floors to install custom plumbing for 1,000W to 1,400W OAM modules, but the vast majority of enterprise data centers and regional cloud providers run on standard, air-cooled 19-inch racks. By keeping the D270 in a passive, air-cooled PCIe 5.0 x16 slot form factor at 350W, Intel allows IT buyers to upgrade existing server infrastructure without touching a single pipe or power distribution unit.
On the software front, Intel is unifying the D-Series under its common Xe software stack with day-one upstream support for PyTorch and vLLM, aiming to make moving open-weight models onto Crescent Island frictionless.
Intel isn’t trying to out-muscle NVIDIA’s flagship Blackwell Ultra racks or AMD’s Instinct MI455X on raw training FLOPS with the D270, and frankly, trying to do so right now would be a fool’s errand.
Instead, Crescent Island hits the market exactly where it hurts: VRAM starvation and infrastructure cost. By pairing 32 stripped-down, matrix-heavy Xe3P cores with up to 480GB of cheap, plentiful LPDDR5X in a 350W PCIe envelope, Intel has created a high-density “tokens-per-watt-per-dollar” workhorse.
With customer sampling happening right now and volume deployments slated to ramp into 2027, the GPU D270’s ultimate success will come down to two things: aggressive pricing from Intel and its ODM partners, and rock-solid vLLM software execution out of the gate.
