Intel just detailed its next datacenter GPU at COMPUTEX 2026. The chip, known internally as Crescent Island, arrives in a PCIe card. Most flagship accelerators favor socketed designs these days. This one stands apart in another way too.
It skips HBM. It skips GDDR. Intel chose LPDDR5x memory instead. The same type found in premium laptops and phones. And it packs a lot of it. Up to 480 GB. That capacity exceeds what NVIDIA’s current top GPUs offer. Memory prices have jumped more than threefold in the past year. LPDDR5x helps keep costs in check.
Bandwidth tells a different story. Intel shared no official figures yet. A wide 1024-bit bus might deliver around 1.2 TB/s. Compare that to the 20 TB/s or more that NVIDIA and AMD push with their latest silicon. Token generation speed hinges on memory throughput. Or it used to.
The industry has shifted. Inference now splits into distinct phases. Prefill handles the initial prompt. It proves compute intensive. Users notice the delay before an AI chatbot begins replying. Decode follows. It generates tokens one by one. The two stages place different demands on hardware.
NVIDIA spotted this split early. Last summer it unveiled the Rubin CPX. The accelerator carried 128 GB of GDDR7. It promised up to 30 petaFLOPS in NVFP4. Video encode and decode blocks came integrated. The plan involved offloading massive prefill jobs to CPX cards while Rubin GPUs managed decode on HBM4 memory. Code assistants and long-context models stood to benefit. Agents have driven token counts higher. The approach looked logical. The Register reported on the concept.
Yet NVIDIA changed course. By March it dropped CPX from public roadmaps. The company acquired Groq and pivoted to LPX racks built around Groq 3 LPUs. Those systems target low-latency decode. SRAM inside the LPUs delivers massive bandwidth for token generation. Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing, later told reporters the CPX idea still held merit. It could appear in future generations. Tom’s Hardware covered the removal.
The need for strong prefill performance never vanished. Enter Intel. Crescent Island uses the Xe-3P architecture. That brings native support for FP8 and FP4 data types. The card draws 350 watts and cools with air. No exotic liquid systems required. Intel has hinted the GPU will work with NVIDIA Dynamo. That framework disaggregates prefill and decode across clusters of accelerators.
But Intel holds other cards. It invested $350 million in SambaNova earlier this year alongside partners. In April the companies outlined a disaggregated inference platform pairing Xeon 6 processors, SambaNova RDUs and, in one configuration, NVIDIA GPUs. That rack-scale system went live this week. It crams 36,864 CPU cores into a 100 kW enclosure while chasing agentic AI workloads. The Register detailed the deployment.
Nothing stops Intel from mixing its own Crescent Island GPUs with SambaNova hardware. Open-source tools such as LLMd offer an alternative to Dynamo. The combination could create vendor-neutral inference clusters. Cost matters here. High-bandwidth memory remains expensive and supply-constrained. LPDDR5x brings capacity without the premium.
Analysts question whether 1.2 TB/s will suffice even for prefill. Compute density could compensate. Intel has released few FLOPS numbers so far. Real-world benchmarks will decide if the design delivers on its promise. Still, the bet looks shrewd. NVIDIA focused resources on decode acceleration after the Groq deal. Intel spotted the opening.
Broader industry moves add context. Intel’s CEO Lip-Bu Tan has steered the company closer to NVIDIA since taking the helm last year. Xeon 6 processors now serve as host CPUs in NVIDIA’s DGX Rubin NVL8 systems. The two firms cooperate on multiple fronts even as they compete in accelerators. Intel’s newsroom announced the integration.
Meanwhile hyperscalers hunt every efficiency gain. Long-context models keep growing. Million-token prompts no longer seem exotic. Video generation and complex coding agents push the boundaries further. A PCIe card with massive cheap memory suddenly appears practical. It slots into existing servers without chassis redesigns. Power and cooling stay manageable.
Success won’t come easy. Intel’s track record in discrete GPUs has been uneven. Software maturity matters as much as silicon. The company must prove its oneAPI tools and drivers handle disaggregated inference smoothly. Partners will watch closely before committing rack-scale deployments.
Yet the timing feels right. NVIDIA’s pivot left a gap. Crescent Island aims squarely at it. If the numbers hold up in testing, Intel could carve out a niche in prefill acceleration. The rest of the stack, whether SambaNova, Xeon or even mixed NVIDIA cards, fills in the decode side. Customers gain flexibility. They avoid locking into a single vendor’s full inference solution.
Memory pricing volatility adds urgency. When HBM costs triple, alternatives gain appeal. LPDDR5x scales with consumer electronics volumes. That supply chain offers stability hyperscalers crave. Bandwidth gaps remain real. Clever scheduling and model partitioning may close some of the difference. Early indications suggest Intel believes they can.
The coming months will bring more data. Intel has promised additional disclosures on performance and software support. Watch for benchmark results on representative long-context workloads. Those will reveal whether Crescent Island truly becomes the GPU NVIDIA’s Rubin CPX nearly was. For now the architecture stands as a pragmatic response to shifting inference economics. And the market may reward that pragmatism.
WebProNews is an iEntry Publication