Technology, Process and Cost
NVIDIA GB10 APU TSMC CoWoS-R
NVIDIA GB 10 and TSMC CoWoS-R : detailed physical analysis, floorplan analysis, and cost structure
SPR26029SCOPE
- Analyzed devices: NVIDIA GB1O
- From module: NVIDIA DGX SPARK
PHYSICAL ANALYSIS
- Detailed photos in optical and SEM views
- Precise thickness measurements
- Package opening
- Delayering
- Material identification
Technology
- TSMC advanced node
- CoWoS-R
Type of Analysis
- SEM View
- TEM View
- Floorplan
- Cross section
- Delayering
- Optical View
- X-ray View
- CT scan
MANUFACTURING PROCESS FLOW
- Recreation of manufacturing process flow
- Step by step process flow
Technical analysis
- Module​
- Component​
- Process​
- Wafer​
- Material​
- IP
- Packaging
COST ANALYSIS
Available in the second version, in July 2026
- Supply chain evaluation
- Simulation of device cost
Type of Analysis
- Die cost
- Yields
- Package assembly cost
- Raw wafer cost
- Wafer Front-End cost
- Die cost
- Dicing & Probe test cost
- Package assemble cost
- Final test cost
Product's objectives
- Provide technology data, manufacturing cost and selling price.
- Physical analysis of the package and die using X-ray, optical microscopy, and scanning electron microscopy (SEM).
- Clear view on the technology and supply chain.
- Package size, die size, die delayering to reveal the manufacturing
- Floorplan to reveal the IP architecture.
- A complete cost analysis and selling price estimation of the NVIDIA GB10.
NVIDIA’s GB10 Grace Blackwell Superchip marks a major step in bringing data-center-class AI computing to compact desktop systems. Designed for Project DIGITS, later commercialized as DGX Spark, the chip targets AI developers, researchers, and data scientists who need to prototype, fine-tune, and run large generative AI models locally before scaling workloads to cloud or data-center infrastructure.
A key strength of GB10 is its unified memory architecture. With 128 GB of coherent unified system memory, systems based on GB10 can run AI models with up to 200 billion parameters locally, while two connected systems can support models up to 405 billion parameters. This memory-centric design addresses one of the main bottlenecks in generative AI workloads: the ability to handle large models without relying exclusively on remote infrastructure.
The chip also reflects NVIDIA’s strategy to extend the Grace Blackwell ecosystem beyond hyperscale data centers. By pairing GB10 with the NVIDIA AI software stack, including CUDA, TensorRT, PyTorch support, development kits, and pre-trained models, NVIDIA enables a consistent development path from desktop prototyping to cloud or data-center deployment.
- Overview
- Company Profile, Supply Chain, Market
- Physical Analysis
- Related Reports
NVIDIA, TSMC
