Intel Data Center GPU Max
Data centre GPU for HPC and AI workloads
The Intel Data Center GPU Max series is a discrete graphics processor designed for high-performance computing and artificial intelligence. Built using a multi-tile chiplet design, it features over 100 billion transistors and up to 128 Xe Cores. The architecture prioritises large cache capacity and high memory bandwidth to accelerate scientific simulations and machine learning tasks.
- Contains over 100 billion transistors
- Uses TSMC 5 nm and Intel 7 process nodes
- Features up to 128 Xe Cores
- Includes up to 408 MB of L2 cache
- Powers the Aurora supercomputer
What it is
The Intel Data Center GPU Max series represents Intel’s entry into the high-end discrete graphics market for data centres. These chips target heavy computational loads such as high-performance computing (HPC) and artificial intelligence (AI). Unlike consumer graphics cards, they focus on double-precision floating-point throughput and massive memory bandwidth rather than gaming features. The design eliminates fixed-function texture units found in other GPUs, acting instead as a parallel processor for compute tasks. It supports ray tracing acceleration and matrix operations through dedicated hardware extensions.
How it came about
Intel developed this architecture to compete with established players in the data centre GPU market. The project, codenamed Ponte Vecchio, marked a significant shift in Intel’s manufacturing strategy by combining multiple process nodes in a single package. The company aimed to establish its Xe architecture across integrated, mid-range, and data centre segments. This effort involved creating a new physical design that utilised advanced packaging techniques to connect different silicon tiles. The goal was to deliver breakthrough results in scientific discovery and AI inference.
Inside the chip
The chip uses a chiplet-based layout with five different process nodes. Compute tiles containing Xe Cores are fabricated on a 5 nm process and sit atop a base tile made with Intel 7 technology. This base tile serves as an input-output die and houses a large L2 cache. Each Xe Core includes a 512 KB L1 cache, which also functions as shared local memory. The L2 cache capacity reaches up to 408 MB, though some variants enable less. This two-level caching system aims to reduce latency penalties from missing the L1 cache. The architecture includes Intel Xe Matrix Extensions for AI vector and matrix math.
How it was made
Manufacturing involves a complex assembly of tiles using embedded bridges and 3D stacking. TSMC produces the compute tiles, while Intel fabricates the base tile. The base die connects to high-bandwidth memory, PCIe interfaces, and peer GPUs. This approach allows Intel to mix process technologies optimal for different functions within one package. The physical design results in a dense component with over 100 billion transistors. The advanced packaging supports high bandwidth between the compute elements and the memory subsystem.
Where it went
These processors power large-scale systems like the Aurora supercomputer at Argonne National Laboratory. They handle workloads such as molecular dynamics simulations and 3D brain visualisations. The chips support up to 128 GB of high-bandwidth memory, making them suitable for data-intensive AI training and inference. They connect via PCIe or Intel Xe Link to scale across multiple servers. The hardware accelerates scientific animation and visualisation tasks through built-in ray tracing units.
Versions and family
The series includes the Max 1550 with 128 Xe Cores and the Max 1100 with 56 Xe Cores. The Max 1100 operates at clocks up to 1.55 GHz and has a 300 W thermal design power. It features 108 MB of enabled L2 cache and 48 GB of HBM2e memory. The larger Max 1550 model offers higher core counts and memory capacity for more demanding tasks. Both variants share the same underlying chiplet architecture and caching hierarchy.
Why it matters
The introduction of this GPU series marked Intel’s serious push into the discrete data centre graphics market. By adopting a chiplet design, Intel demonstrated a new path for scaling performance beyond monolithic dies. The large cache sizes influenced trends in handling memory bandwidth constraints in compute architectures. Its deployment in top-tier supercomputers highlighted its role in advancing scientific computing capabilities. The design choices reflected a balance between raw throughput and memory latency management.
Questions people ask about Intel Data Center GPU Max
What is the Ponte Vecchio chip?
Ponte Vecchio is the codename for the Intel Data Center GPU Max series. It is a discrete graphics processor built for high-performance computing and AI workloads. The chip uses a multi-tile design with over 100 billion transistors.
How many cores does the Max 1550 have?
The Intel Data Center GPU Max 1550 contains 128 Xe Cores. These cores serve as the basic building blocks for parallel processing tasks. The smaller Max 1100 model has 56 Xe Cores.
What memory type does it use?
The GPU uses high-bandwidth memory (HBM2e). The Max 1100 variant includes 48 GB of this memory, while larger models support up to 128 GB. This provides theoretical bandwidths exceeding 1 TB/s.
Who manufactures the Ponte Vecchio tiles?
Intel uses a mix of foundries for the chiplets. TSMC fabricates the compute tiles on a 5 nm process. Intel manufactures the base tile using its Intel 7 process technology.
Good reading and sources
There is no English Wikipedia article on Intel Data Center GPU Max yet, so the facts on this page come from the sources listed above: intel.com/content/www/us/en/products/details/discrete-gpus/data-center, chipsandcheese.com/p/intels-ponte-vecchio-chiplets-gone-crazy, nextplatform.com/compute/2021/08/24/intels-ponte-vecchio-gpu-better-no. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.
Like Intel Data Center GPU Max
All graphics processorsMore from Intel
All of them
WorldMOS Technology 8502
An 8-bit CPU from MOS Technology that powered the Commodore 128CPU · 8-bit
🇺🇸 United StatesRapidCAD
A 486DX in a 386 socket, with a dummy FPU chip to match.CPU · x86
🇺🇸 United StatesTolapai
Intel embedded SoC with Pentium M core and security accelerationSoC · x86