Monday, February 16th 2026
Intel Confirms Data Center GPU IP After Xe3P with "Xe Next"
Intel has once again confirmed its product roadmap for the next generation of graphics IP, despite rumors that there might not be anything new after the launch of the Xe3P. According to a new presentation from Intel's Anil Nanduri, Vice President of GTM & Product Management AI Accelerators, responsible for data center AI accelerators, Intel will have a dedicated graphics portfolio after the recently introduced "Crescent Island." Labeled as "Xe Next," the next generation of GPU IP is still under development and currently without an official name. It is expected to follow the rollout of the Xe3P-based "Crescent Island" inference-focused GPUs. With "Xe Next," the data center lineup is seemingly expanding into two production lines: inference-focused "Crescent Island" GPUs and the long-awaited "Jaguar Shores" GPU for AI training and HPC workloads.
For AI training and HPC workloads, Intel's "Jaguar Shores" design is reportedly being finalized and will be completed by the middle of this year, with production starting shortly after. Assuming a perfect design, we could see Intel's "Jaguar Shores" GPU late this year or early 2027. It will likely involve a mix of Intel's internal production in collaboration with TSMC or Intel's pure 18A design, assuming yields are manageable. We have already seen that Intel confirmed HBM4 memory will be used in these accelerators, and SK hynix appears to be the main partner supplying Intel with this memory. As a rack-scale design, Intel will compete with NVIDIA and AMD in the highest-margin business.For consumers, the company is already shipping Xe3-based graphics in the form of "Panther Lake" iGPU in form factors like Arc B390/B380. While not exactly a "Celestial" generation of Arc graphics, it increases compute density, widens the vector and matrix units, and expands cache capacity for higher sustained throughput at similar power levels compared to Xe2 "Battlemage." For consumers, we expect to see a mix of the Xe3P-LPG for the core GPU IP and Xe4 "Druid" architecture for media and display duties. This mix and match of upcoming IP will be headed towards "Nova Lake" processors later this year, which is expected to debut in late 2026.
Source:
Intel on X
For AI training and HPC workloads, Intel's "Jaguar Shores" design is reportedly being finalized and will be completed by the middle of this year, with production starting shortly after. Assuming a perfect design, we could see Intel's "Jaguar Shores" GPU late this year or early 2027. It will likely involve a mix of Intel's internal production in collaboration with TSMC or Intel's pure 18A design, assuming yields are manageable. We have already seen that Intel confirmed HBM4 memory will be used in these accelerators, and SK hynix appears to be the main partner supplying Intel with this memory. As a rack-scale design, Intel will compete with NVIDIA and AMD in the highest-margin business.For consumers, the company is already shipping Xe3-based graphics in the form of "Panther Lake" iGPU in form factors like Arc B390/B380. While not exactly a "Celestial" generation of Arc graphics, it increases compute density, widens the vector and matrix units, and expands cache capacity for higher sustained throughput at similar power levels compared to Xe2 "Battlemage." For consumers, we expect to see a mix of the Xe3P-LPG for the core GPU IP and Xe4 "Druid" architecture for media and display duties. This mix and match of upcoming IP will be headed towards "Nova Lake" processors later this year, which is expected to debut in late 2026.

9 Comments on Intel Confirms Data Center GPU IP After Xe3P with "Xe Next"
scratchedsacrificed on the alter of AI data centres.Man this is THE worst timeline, intel was suppose to ermm "destroy the Sith, not join them".
2026 is stupid.
If the ram demand doesn't normalize there won't be any gddr available to make graphics cards that are worth getting at reasonable prices.
The problem is that C770 head is in real danger to be axed..