At Hot Chips 2026, Intel disclosed a Crescent Island datacenter GPU with 480 GB of LPDDR5X DRAM, a capacity the company says tops every other GPU and AI accelerator on the market.

The 480 GB pool lets the chip host Deepseek-V4 Flash entirely in VRAM — a feat no other non-HBM accelerator can match, Intel says.

What compute does Crescent Island actually deliver?

Intel built the chip on its new Xe3P IP, placing it within the Celestial datacenter GPU family. The Xe3P core changes over Xe3 are concrete and quantifiable:

• Xe3P quadruples XMX matrix units versus Xe2 and Xe3, and adds FP8 and FP4 datapaths
• Xe3P raises L1/SLM cache to 512 KB per Xe core, up from 384 KB in Xe3
• Xe3P doubles the register file to 1 MB per core
• Xe3P keeps the XVE ALU count at SIMD16, contradicting earlier rumors of a doubling

Crescent Island packages 32 Xe3P cores alongside 32 MB of last-level cache. The 32 MB LLC sits below NVIDIA Blackwell and AMD RDNA3/4, but the analysis presented at Hot Chips argued AI workloads remain memory-bandwidth-bound, so a larger cache buys little.

What is the actual memory bandwidth?

Intel declined to disclose the figure. When pressed, the company replied: "We are not disclosing that information at this time."

Based on the 480 GB capacity and leaked PCB photographs, only one memory bus width fits: 1280 bits across 20 LPDDR5X modules — 12 on the front of the board, 8 on the rear. At the LPDDR5X-9600 data rate, that bus would deliver over 1.5 TB/s of throughput.

How does FP64 stack up against competing accelerators?

Crescent Island carries 2,048 full-rate FP64 FMA units across the die. At an assumed 2.5 GHz clock, the chip delivers approximately 10 TFLOP/s of FP64 — substantially above competing accelerators in this class, most of which omit full-rate FP64 entirely.

The HPC angle matters. Intel's next-generation CPUs — Venice and Diamond Rapids — will carry even higher FP64 throughput, more memory bandwidth, and larger memory capacity, making them attractive alternatives for HPC buyers who don't need a discrete GPU.

Assuming Intel clocks the die at approximately 2.5 GHz and quadrupled matrix-operation rates without shifting datatype ratios, the projected compute stack lands at:

• 10.2 TFLOP/s FP64 vector
• 20.5 TFLOP/s FP32 vector
• 41 TFLOP/s FP16 vector
• 328 TFLOP/s TF32 via XMX
• 655 TFLOP/s FP16/BF16 via XMX
• 1.3 PFLOP/s FP8 via XMX
• 2.6 PFLOP/s FP4/MXFP4 via XMX

Crescent Island lands roughly 30% ahead of NVIDIA's RTX PRO 6000 Blackwell on matrix compute and over 5x on FP64. The RTX PRO 6000 retains a 6x lead in FP32 and a 3x lead in FP16 vector throughput.

When does Crescent Island actually ship?

Intel announced at OCP 2025 that Crescent Island would sample by year-end. The limited detail on display at Hot Chips, combined with Intel's rocky history in datacenter GPUs, raises the prospect of a slip into 2027.

Who is the customer?

The LPDDR5X architecture targets a buyer who prioritizes memory capacity per dollar above all other considerations — a market segment Apple currently serves alone. The full-rate FP64 pipeline also opens doors in HPC, where competing accelerators have throttled double-precision throughput.