🔍 Read the full analysis: How 3D Packaging Brings AI Chip Dies And HBM Together on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A November 22 explainer from ThorstenMeyerAI.com describes how advanced AI accelerator packages place compute dies and high-bandwidth memory (HBM) on a silicon interposer, with some designs also stacking compute dies vertically. The article says packaging capacity and HBM supply are constraints alongside chip performance, while several details—including physical package dimensions—are not publicly specified by chipmakers.
ThorstenMeyerAI.com published an explainer November 22 describing how advanced AI accelerators combine compute dies and high-bandwidth memory (HBM) in a single package, using silicon interposers and, in some designs, vertically stacked compute dies. The development matters because accelerator performance and deployment depend not only on the processors, but also on the ability to move data to them and manufacture the complex packages at scale.
The explainer breaks down the package from the bottom up: a package substrate carries power and signals, while a silicon interposer provides dense connections between the compute dies and memory stacks. HBM consists of DRAM chips stacked above a base die and joined by vertical connections called through-silicon vias (TSVs). This arrangement places memory close to the processors and gives them many connections in a compact area.
The article compares designs using figures and descriptions attributed to the companies. It says Nvidia’s Rubin GPU has two reticle-limited compute dies and 336 billion transistors, citing Nvidia’s architecture blog. AMD describes its MI350 Series as having eight vertically stacked accelerator complex dies and two I/O dies; the article gives the MI355X a total of 185 billion transistors. AMD says its MI450X and MI455X use CoWoS-L packaging, and has described stacked 3D hybrid-bonded compute dies for those products.
Memory and package specifications are also changing. The article says the JEDEC HBM4 standard, published in April 2025, allows stacks of four, eight, 12 or 16 layers and up to 64 GB per stack. It cites TSMC’s descriptions of CoWoS packaging and reports the company’s CEO said on October 15, 2026, that it was in real-world production of five-reticle-size interposers while working on larger designs. These are company statements and product specifications, not independent measurements by the explainer’s authors.
Packaging Sets AI Accelerator Limits
The package architecture helps explain why adding faster compute dies alone may not make an AI system proportionally faster. Those dies must receive data quickly, which makes memory bandwidth and the wiring between memory and processors central to performance. The explainer cites Nvidia product-page figures showing bandwidth rising from 900 GB/s for V100 to 19.2 TB/s for Rubin. These are product-specific figures across different generations; they do not by themselves establish how much faster a workload will run.
Manufacturing capacity is another constraint. Large interposers and stacked dies require specialized assembly, and the article describes advanced packaging as a production bottleneck in its own right. It also cites SK hynix’s expectation that HBM demand will exceed supply through 2030. That forecast is the memory maker’s assessment, not a guarantee of future shortages. If supply remains tight, operators may face limits on how many accelerators they can obtain, regardless of demand for AI computing.
The rack view makes the infrastructure implications concrete. The article says Nvidia’s GB200 and GB300 NVL72 systems contain 72 GPUs and 36 CPUs per rack, with liquid cooling and dedicated networking and power equipment. Such systems make chip packaging part of a broader deployment challenge involving memory, cooling, rack design and component availability.
Products Worth Considering

Bose QuietComfort Ultra Headphones Gen 1, Lunar Blue

Bose SoundLink Plus Portable Bluetooth Speaker, Black

Sony WH-1000XM4 Wireless Noise Cancelling Headphones (Silver)
As an affiliate, we earn on qualifying purchases.
From HBM Stacks to Full Racks
The explainer uses a 3D model to illustrate layer order and component counts, but says its dimensions are schematic. According to the author, Nvidia and AMD do not publish the dimensions in millimetres of the dies, full packages or memory stacks discussed, so the model does not estimate them. That distinction matters: the model is an explanatory visualization, not a physical-scale rendering or teardown.
HBM stacks and silicon interposers address related but separate design needs. TSVs connect the DRAM layers vertically within a memory stack. The interposer then links the stacks to compute and other dies across the package. TSMC describes CoWoS-S as using a large silicon interposer to accommodate functional dies and HBM. The article says TSMC’s CoWoS-L uses smaller silicon bridges joined in a carrier layer for larger package designs.
At rack scale, the article cites Nvidia’s DGX documentation and AMD’s CDNA 4 white paper for component counts, and an Open Compute Project listing for the GB200 NVL72 rack outline. It describes a liquid-cooling system in which coolant runs through cold plates attached to processors. These examples show how a package design fits into a system, but they do not establish that every accelerator or server uses the same layout.
““large silicon interposer area to accommodate various functional top dies, including logic chiplets, with high-bandwidth memory (HBM) cubes stacked over it””
— TSMC, as quoted in the explainer
Dimensions and Supply Outlook
Several limits remain on what can be verified from the material. The 3D model’s dimensions are schematic, and exact physical dimensions for the dies and packages are not provided. The source also combines specifications and statements from Nvidia, AMD, TSMC, SK hynix and other organizations; these are attributed company disclosures rather than an independent teardown or production audit.
The scale and duration of any HBM or packaging shortage are also uncertain. SK hynix’s forecast signals its view of demand and supply, but the explainer does not give a quantified shortfall, a market-wide baseline or a confirmed timeline for relief. Product availability and the performance customers will see in specific workloads are not established by the package diagrams or bandwidth figures alone.
Capacity and Product Rollouts
The next developments to watch are production capacity for advanced packaging and HBM, and the specifications of the next accelerator platforms as they become available. TSMC’s reported work on larger interposers and AMD’s descriptions of future stacked-die products point to ongoing design and manufacturing changes; the article does not provide delivery schedules for those advances.
For buyers, the practical test will be whether suppliers can deliver complete systems at scale, including memory, cooling and networking—not just compute dies. Updated product documentation, independent system testing and clearer supply figures would help determine how the package designs translate into actual performance and availability.
Key Questions
What does HBM mean in an AI accelerator?
HBM means high-bandwidth memory. It is made from DRAM layers stacked vertically and connected using through-silicon vias. The stacks sit close to compute dies in the package to provide a high-bandwidth path for data.
What does the silicon interposer do?
The interposer is a dense wiring layer that connects compute dies, memory stacks and other components inside a package. It provides many more short connections than a conventional circuit board can accommodate in the same area.
Why does packaging affect AI performance?
Compute dies need a steady flow of data. Memory bandwidth and package connections can limit how effectively processors are used, so faster compute alone does not determine system performance.
Are the explainer’s 3D model dimensions to scale?
No. The article says the model shows layer order and component counts, but uses schematic sizes because the companies do not publish the relevant physical dimensions in millimetres.
Is an HBM shortage through 2030 confirmed?
No. The explainer attributes that expectation to SK hynix. It is the company’s forecast that demand will exceed supply through 2030, not confirmation of the size or duration of a market-wide shortage.
Source: ThorstenMeyerAI.com
