Intel's Crescent Island GPU, now the "GPU D-Series", aims at cheap, dense AI inference
Intel is taking on high-volume inference with a 480 GB, air-cooled LPDDR5X accelerator that trades raw peak FLOPS for huge capacity, low power and lower cost per token.
Two days ago at the OCP Summit 2026, Intel removed the last bit of mystery around its Crescent Island accelerator, formally naming it the Intel GPU D-Series ("Intel GDDR D", short for GPU DRAM D) and publishing a full specification. The result is a single, air-cooled PCIe card tuned to move a massive amount of memory at low power and low cost, aimed squarely at high-volume large-model inference rather than training.
- 32 Xe3P cores with 256 vector engines and 256 XMX AI engines
- Up to 1.3 PFLOPS FP8 (1,300 TFLOPS) and 1,280 TOPS INT8, up 2.6x vs. Gaudi 3
- Up to 480 GB of LPDDR5X on-package memory with 1.5 TB/s bandwidth
- Up to 350 W TDP, delivered as an air-cooled, rack-mountable PCIe card
- FP4 (introduced for the first time) and FP8 support plus a 32 MB unified L2 cache
The headline trade-off is memory capacity. Where its rivals lean on high-bandwidth memory, Intel instead packs up to 480 GB of LPDDR5X directly onto the package, giving the card a 2.5x higher memory density per watt than Intel's earlier Jaguar Shores platform and far more resident capacity than a single HBM-based accelerator. More resident memory means fewer model shards and less off-chip traffic, which is exactly what keeps a large model fast and affordable to serve.
Why that matters for inference
Running a big model in inference is less about raw FLOPS and more about keeping weights close to the compute and avoiding expensive data movement. By using the same memory as the rest of the system, the D-Series keeps the data flow "as efficient as a CPU with a big memory" while the Xe3 core cluster and dedicated XMX block do the actual math. The net effect is a target of 2x throughput, 2x scale-out capacity and 3x lower power versus a CPU-plus-GPU box.
It is a deliberately different bet than its nearest competitors, such as AMD's HBM-based Instinct cards. Intel is optimising for the cost and power of serving enormous models at scale, not the raw training FLOPS that dominate marketing slides. That positions the D-Series as one more credible answer to the same question now being asked by many hyperscalers: how do you stop relying on a single supplier for the silicon running your inference?
If you run or buy AI systems, the D-Series is a sign that affordable, dense inference hardware is becoming a real category. Servers and racks built around it could cut inference cost per token and power draw, and the air-cooled design means it can fit into existing data centers without exotic cooling. Customer sampling is expected to begin this year, so expect vendor platforms and pricing to emerge in the coming months.
Bottom line
Intel's Crescent Island, now the GPU D-Series, is a clean statement of intent: win the high-volume inference market with a simple, air-cooled card that trades peak FLOPS for huge memory capacity, low power and lower cost. Whether it lands depends on its software stack and ecosystem, but on paper it is one of the most direct challenges yet to the current inference silicon default.
Sources: Intel Newsroom (announces the Intel GPU D-Series); Tom's Hardware (Crescent Island details and D-Series naming); TechPowerUp (Intel names Crescent Island as GPU D-Series).
Corrections: Report an error at corrections@wireshaman.com.