Intel names its Crescent Island inference GPU the GPU D Series, with up to 480GB of memory and 1.3 petaflops of FP8 in a 350W air cooled card
Intel GPU D-series, code-named Crescent Island - 480 GB LPDDR5X memory - 1.5 TB/s of memory bandwidth - 1.3 petaflops of FP8 - 350 W TDP
Context
A power efficient inference card that fits ordinary air cooled servers gives data centers an alternative to Nvidia and AMD for running large models, if Intel can deliver.
Andreas Schilling is an editor at the German hardware site Hardwareluxx, and his post summarizes Intel's announcement at the Open Compute Project summit on October 12. Intel had outlined the Crescent Island architecture at Hot Chips earlier this year. The 480GB figure applies to partner built cards, while Intel's own card carries 160GB.
Deep dive
On October 12, Andreas Schilling posted the headline specs for Intel's GPU D Series, code named Crescent Island: 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth, 1.3 petaflops of FP8 and a 350 W TDP.
According to ServeTheHome, the first model is the GPU D270, built on Intel's Xe3P architecture with 32 Xe cores, 256 XMX engines, a clock speed of about 2.5GHz, 32MB of L2 cache and PCIe Gen5 x16. It is an air cooled PCIe card, and ServeTheHome noted that AMD's comparable MI350P runs at up to 600 watts in its top configuration.
TechPowerUp reported that Intel's branded card ships with 160GB, while ODM partners can fit up to 480GB, and that Intel says one card can run models of up to 300 billion parameters, with an eight card setup handling trillion parameter models. It added that the chip supports formats from FP4 to FP64 and includes four video encoders and four decoders.
Intel says customer sampling is underway with server makers, neoclouds and software vendors. Pricing, a launch date and independent performance results have not been disclosed, and ServeTheHome said buyers should look to 2027 for release.
What to expect next
Watch for pricing and a launch date, and for independent benchmarks of cost per token against Nvidia and AMD inference cards.
Also watch which server makers and neoclouds announce D270 systems, whether partners ship the 480GB configuration, and whether Intel's software stack is mature enough for customers to move inference workloads onto it. A 2027 launch would put it up against newer Nvidia and AMD parts.