AI accelerators are usually introduced with a giant compute number. Intel’s new pitch puts a different constraint first: fitting more of a model into each card.
At the Open Compute Project Global Summit in October 2026, Intel disclosed detailed specifications for its GPU D-series, formerly code-named Crescent Island. The announced design combines up to 480 GB of LPDDR5X memory, 1.5 terabytes per second of memory bandwidth, 1.3 petaflops of peak FP8 compute and a 350-watt power envelope. Intel’s OCP technical presentation lists the underlying 32 Xe cores and 256 matrix engines.
Why memory capacity matters for inference
Training creates a model; inference runs it to answer prompts, classify data or operate an agent. During inference, model weights, the growing key-value cache for long conversations and batches of concurrent requests all compete for accelerator memory.
If a model does not fit on one card, operators must partition it across several accelerators. That can increase cost and introduce communication overhead. Intel says a 480 GB partner-designed card can support models with up to 300 billion parameters, while an eight-GPU system can handle more than one trillion parameters. Actual requirements depend heavily on numerical precision, quantisation, context length and serving software.
The 480 GB headline needs a footnote
Intel’s own small print says its branded PCIe card has 160 GB of LPDDR5X. The platform design allows original-design-manufacturer partners to build cards with options up to 480 GB. In other words, 480 GB is the top ecosystem configuration, not necessarily the memory fitted to every Intel-branded product.
LPDDR5X is widely used where power efficiency and capacity matter. Its 1.5 TB/s bandwidth is lower than the multi-terabyte-per-second figures associated with the newest high-bandwidth-memory accelerators. Intel is betting that some inference workloads will value larger capacity, conventional air cooling and a lower 350-watt envelope more than maximum bandwidth.
The standard PCIe form factor is also part of the pitch. Intel says operators can install D-series cards in existing air-cooled data-centre environments instead of adopting specialised liquid-cooled racks. That could simplify deployment for enterprises whose power density and facilities were not designed around the highest-wattage accelerators.
Capacity can change the shape of a cluster
More memory per card may let an operator use fewer accelerators for a given model, but it does not automatically mean a faster or cheaper service. A smaller accelerator count reduces interconnect traffic and can simplify scheduling; lower bandwidth can slow token generation when the workload repeatedly moves weights through memory. The winning balance depends on batch size, model architecture and latency targets.
The 350-watt envelope could also matter beyond the server. Air-cooled racks are easier to place in existing facilities, and a lower card power may leave more of a site’s electrical budget available for additional machines. Those advantages have to be measured at the rack and application level, where CPUs, networking, cooling and idle time all contribute to energy use.
Specifications are not production economics
Customer sampling is underway, but Intel has not supplied public pricing, broad availability dates or enough independent benchmark data to test its economic claims. Peak FP8 throughput does not reveal tokens per second, time to first token, latency under load, utilisation or energy per response. Software support and model optimisation can matter as much as silicon.
Tom’s Hardware’s technical summary likewise treats the announcement as an entry into a crowded inference market, not evidence that the product has displaced established accelerators.
The significant development is the design priority. Intel is arguing that the next phase of AI infrastructure will be measured less by a single training benchmark and more by how economically systems keep large models serving many users. The D-series specifications make that case on paper; deployments and independent testing will decide whether the bet works.
Leave a Reply