AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: How AI Chip Components Come Together In Three Dimensions on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

A November 22 explainer from ThorstenMeyerAI.com uses 3D models to show how AI accelerators combine compute dies, high-bandwidth memory, silicon interposers and cooling inside a package, then connect many packages in a rack. The report points to memory bandwidth and advanced packaging capacity as constraints alongside compute, but some product specifications and supply forecasts are company statements, and the models use schematic dimensions.

ThorstenMeyerAI.com published a technical explainer on November 22 showing how AI accelerators are assembled from compute dies, stacked high-bandwidth memory, silicon interposers and cooling hardware—and how 72 GPUs can be combined in a single rack. The report’s 3D models illustrate why memory bandwidth and advanced packaging, not just processor performance, affect how quickly AI systems can run and how many can be built.

The explainer depicts an accelerator as a multi-part package, rather than a single piece of silicon. A package substrate carries power and signals to the system module. Above it, a silicon interposer provides dense wiring between compute dies and nearby high-bandwidth memory (HBM). The HBM consists of DRAM chips stacked vertically and linked through silicon vias, or TSVs. A cold plate on top carries liquid cooling to the processor components.

The article compares products from Nvidia and AMD, drawing on their published specifications. It says Nvidia’s Rubin GPU uses two reticle-limited compute dies and lists 336 billion transistors, citing Nvidia’s architecture blog. For AMD’s MI355X, it describes eight vertically stacked accelerator complex dies and two I/O dies, alongside eight 12-high HBM stacks; AMD lists 185 billion transistors for the product. These configurations differ, and the figures are manufacturer specifications rather than independent measurements.

The model then zooms out to rack scale. Nvidia’s GB200 and GB300 NVL72 configurations are described as having 72 GPUs and 36 CPUs, arranged across compute and switch trays with power shelves. Nvidia’s guide says coolant runs through cold plates attached to CPUs and GPUs; the report cites 120 kilowatts of cooling capacity for a GB200 NVL72 rack. The models show component order and counts, but not exact physical dimensions: the author says chip and package sizes are schematic because the companies do not publish those measurements.

At a glance
reportWhen: Published November 22, 2026
The developmentThorstenMeyerAI.com published a layer-by-layer explainer of AI accelerator packaging, using 3D models to connect chip components to rack-scale systems.

Packaging Shapes AI System Supply

The report’s central point is that AI performance depends on moving data to compute as well as on the compute itself. HBM bandwidth helps feed large accelerator dies, while the interposer’s dense connections allow memory and compute to communicate across a package. Adding more or faster compute may not produce the expected system-level gains if memory access, packaging or cooling cannot keep pace.

Those requirements also affect supply. Advanced interposers and vertically stacked memory need specialized manufacturing, and the source says packaging capacity is a production constraint in its own right. SK hynix has forecast that memory demand will exceed supply through 2030, according to the report; that is a company outlook, not a settled guarantee. If supply remains tight, access to accelerators could continue to depend on the availability and cost of several components, not only on GPU production.

Rack-scale systems make the consequences visible for data-center operators. A 72-GPU installation brings substantial power, cooling and networking needs in addition to the chips themselves. The technical details matter because they help explain why deploying AI capacity requires coordinated supplies of memory, packaging, servers and cooling infrastructure.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

From Chip Package to Rack

The report places the designs in a broader progression toward larger and more complex packages. A reticle is the maximum area a lithography scanner can print in one exposure. TSMC describes CoWoS-S interposers as reaching up to 3.3 times reticle size, or about 2,700 square millimeters. For larger designs, CoWoS-L combines smaller silicon bridges in a carrier layer. The article cites TSMC’s chief executive as saying on October 15, 2026, that the company had entered real-world production of five-times-reticle-size packaging and was working on larger sizes.

HBM has also advanced by stacking more DRAM layers while reducing the thickness of individual chips and the gaps between them. The article cites JEDEC’s April 2025 HBM4 standard, which allows stacks of four, eight, 12 or 16 layers and up to 64 GB per stack. Its comparisons of bandwidth—from 900 GB/s for Nvidia’s V100 to 19.2 TB/s for Rubin—are manufacturer figures from different products and generations, not a guarantee of application performance.

At rack scale, the source cites Nvidia’s system guide and AMD’s technical material for component counts, and an Open Compute Project listing for a GB200 rack’s stated dimensions. The 3D models are the author’s, built in open-source CAD software with layer order and counts represented, but schematic dimensions. The article says its purpose is to make the construction and dependencies easier to see, rather than to provide manufacturing drawings.

““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 ThorstenMeyerAI.com’s report

Dimensions and Supply Forecasts

The article’s models are schematic rather than dimensionally exact. It says Nvidia and AMD do not publish the millimeter dimensions of the dies, packages or memory stacks, so the models do not establish their precise physical sizes. The cited product configurations and bandwidth figures also come from company materials and are not presented as independent tests.

Future supply remains uncertain. SK hynix’s expectation that memory demand will exceed supply through 2030 is a forecast, and the report does not quantify how much packaging capacity is available or how quickly it may grow. It also does not establish how shortages, manufacturing yields or system costs will change for customers. The article describes several current and planned configurations, but specifications and availability can change as products move through production.

Capacity and Product Rollouts

The next developments to watch are production capacity for advanced packaging and HBM, along with vendor updates on accelerator availability and rack configurations. TSMC’s stated work on interposers larger than five times reticle size could expand the packaging options for future systems, but the report does not provide a schedule for the larger sizes it says are in development.

As Nvidia, AMD and memory suppliers release further specifications, system builders will be able to assess how package designs, memory capacity, bandwidth, cooling and rack power change across generations. For now, the report offers a structural account of how those parts fit together; the scale and duration of any supply constraints remain dependent on manufacturing outcomes and future demand.

Key Questions

What is an AI accelerator package made of?

In the configurations described by the report, a package combines compute dies, HBM stacks, a silicon interposer and a substrate, with cooling hardware attached in the server. Exact designs vary by manufacturer and product.

What does HBM do?

High-bandwidth memory places stacked DRAM close to the compute dies. TSVs connect the memory layers vertically, while the package’s interposer provides dense connections between memory and compute.

Why does the interposer matter?

The silicon interposer supplies many fine connections between components in a package. The report says this specialized packaging can constrain production as designs grow larger and more complex.

How many GPUs are in an Nvidia NVL72 rack?

The report describes the GB200 and GB300 NVL72 configurations as containing 72 GPUs and 36 CPUs, distributed across compute trays, with separate switch trays and supporting equipment.

Are the report’s 3D models exact scale?

No. The author says the models represent layer order and component counts, but use schematic dimensions because exact die, package and memory-stack measurements are not published in the source material.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

The 4.8 Staircase: What the Market Actually Believes About Claude’s Next Release

Market predictions suggest a high chance of Claude 4.8 release by July, but no official announcement has been made. Here’s what is confirmed and what remains speculative.

External GPUs Perfect For AI Tasks In 2026

In 2026, external GPUs (eGPUs) are emerging as key tools for AI workloads, offering flexibility and power for laptops and mini PCs. Learn what’s new.

Mobilised, Not Spent: What’s Left of Europe’s €200 Billion AI Offensive

Europe aims to mobilize €200 billion for AI, but only a fraction is committed, and actual spending lags behind. The real challenge remains unaddressed.

2026 Gaming PC Essentials: The 8 Best Motherboards

A detailed overview of the eight best gaming motherboards for 2026, highlighting features, compatibility, and what builders should consider.