Intel reportedly plans 480GB AI GPU with LPDDR5X memory

What happened

Intel is reported to be planning a graphics card with 480GB of memory aimed at AI workloads, according to TechRadar. The reported specification combines 1.5TB/s of memory bandwidth, 1.3 petaflops of FP8 performance and a 350W power draw. Pricing and availability have not been disclosed.

The most unusual detail is the memory type. The headline reporting says the card will use LPDDR5X rather than GDDR7, the memory more commonly associated with high-end graphics cards. The memory choice departs from the usual GPU design playbook, which typically prioritises bandwidth per chip over total capacity.

The story is currently based on a plan rather than a confirmed product launch. Until Intel publishes commercial details, the specification should be treated as reported rather than final.

Why it matters

For AI teams, memory capacity is often the constraint that decides what can run on a single accelerator. Large models and long-context workloads need to hold substantial data close to the processor, and a card with this reported capacity would change how many systems are needed for certain tasks. Bandwidth and power draw determine how efficiently that data can be used, so the three figures should be read together rather than as isolated headline numbers.

The reported use of LPDDR5X also points to a wider question for organisations buying AI infrastructure: how hardware is balanced between capacity, speed, power and cost. Leaders planning digital and AI investments will need to assess these trade-offs against their own workloads, rather than assuming that the highest-performance memory is always the right answer. Without pricing or availability, procurement and budgeting teams cannot yet make firm comparisons.

By the numbers

• 480GB of reported onboard memory capacity for the planned Intel GPU
• 1.5TB/s of reported memory bandwidth
• 1.3 petaflops of reported FP8 performance
• 350W of reported power consumption

The Renascence take

Most readers will focus on the 480GB figure, but that is the least useful number on the sheet for experience-led decisions. Capacity only creates value when the workload, the data pipeline and the people operating the system are designed around it. The principle underneath is simple: a service is shaped by how quickly and reliably information reaches the point of use, not by how much is theoretically stored nearby.

The headline is about capacity, but the real question is whether a service can keep data flowing to the people and processes that depend on it. Customer-obsessed operators should ask vendors how memory is accessed under real load, what latency customers will actually experience, and what it costs to run at 350W around the clock. Treat any new AI hardware as a service design decision, not a spec-sheet upgrade. Wait for pricing and availability before committing, and pilot against the experience you need to deliver.