Intel's record-breaking GPU unveiled

At the Open Compute Project Global Summit, Intel unveiled the GPU D-series with up to four hundred eighty GB of memory, designed for demanding AI inference workloads.

Dethroning Nvidia's previous record

The D-series surpasses Nvidia's Blackwell Ultra B300, which offers two hundred eighty eight GB of HBM3e memory, as the data center GPU with the highest memory capacity.

LPDDR5X over GDDR7

Intel chose LPDDR5X memory, commonly found in laptops, instead of GDDR7 used in modern graphics cards, prioritizing memory capacity and power efficiency for AI inference.

Supporting massive language models

The accelerator supports large language models containing up to three hundred billion parameters on one card, while an eight GPU system accommodates models exceeding one trillion parameters.

Intel's inference strategy

Anil Nanduri, vice president of AI Products and GTM at Intel Data Center Group, said the D-series focuses on making high capacity inference practical to deploy and economical to scale.

“With Intel GPU D-series, we're focused on making high-capacity inference practical to deploy and economical to scale.”

Performance and data center fit

The D-series rates at one point three petaflops for FP8 computation and one point five TB per second memory bandwidth.

Put every prompt in its place

Intel plans to connect the accelerator with Xeon processors for coordinating tasks and handling workloads that do not require GPU acceleration.Nicolas Dubé said the goal is matching different workloads with the right intelligence and economics.

“Put every prompt in its place.”