The Building Blocks Of An AI Chip, Explained In 3D
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: The Building Blocks Of An AI Chip, Explained In 3D 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

ThorstenMeyerAI.com published a 3D explainer on Nov. 22, 2026, showing how an AI accelerator combines compute dies, stacked HBM, an interposer, substrate and cooling. The article links those package layers to rack-scale systems and identifies memory bandwidth, advanced packaging and component supply as constraints; some product and supply details are company statements, not independent findings.

ThorstenMeyerAI.com published a 3D explainer on Nov. 22 showing how an AI accelerator is assembled from compute dies, high-bandwidth memory, a silicon interposer, a package substrate and liquid cooling. The report moves from individual chip layers to 72-GPU rack systems, arguing that memory bandwidth and advanced packaging—not just raw compute—help determine how quickly AI hardware can be built and used.

The explainer’s 3D model presents the parts in their physical order, from the package substrate at the base to a cold plate at the top. Between them, a silicon interposer carries dense connections between compute dies and memory stacks. The source says its model uses the correct layer order and component counts but schematic dimensions: Nvidia and AMD do not publish the relevant measurements in millimetres, so the model does not estimate them.

In the model, HBM (high-bandwidth memory) consists of DRAM chips stacked above a base die and connected vertically by through-silicon vias, or TSVs. The interposer provides wiring that a conventional circuit board could not accommodate at the same density. The report describes Nvidia’s Rubin GPU as using two reticle-limited compute dies and cites Nvidia’s product page for 19.2 terabytes per second of memory bandwidth. The article compares that figure with 900 gigabytes per second for the V100; those are product specifications from different generations, not a measure of performance on a particular AI task.

The rack view shows how packages scale into larger systems. According to the cited Nvidia DGX GB user guide, a GB200 or GB300 NVL72 rack has 72 GPUs and 36 Grace CPUs across 18 compute trays, alongside nine NVLink switch trays, management switches and power shelves. Nvidia’s documentation also specifies 120 kilowatts of cooling capacity for a GB200 NVL72 rack. The source says the next-generation Vera Rubin NVL72 is listed by Nvidia as available now, with 72 Rubin GPUs and 36 Vera CPUs.

At a glance
reportWhen: Published November 22, 2026
The developmentThorstenMeyerAI.com published a layered 3D explainer of AI accelerator packages and rack systems, highlighting packaging and memory as potential constraints on scaling.

Why Memory and Packaging Limit Scale

The 3D breakdown matters because an accelerator’s advertised compute capability is only one part of its operation. Models need data delivered to processors quickly; memory bandwidth and the connections between chips affect how effectively compute resources can be fed. The source uses the rise in cited bandwidth—from the V100’s 900 GB/s to Rubin’s 19.2 TB/s—to illustrate how manufacturers have increased memory throughput across product generations. It does not establish that bandwidth alone determines the speed of a model or the performance of a complete system.

Packaging also affects how many accelerators can be produced. Large interposers and stacked memory require specialized manufacturing, and the article describes advanced packaging as a production bottleneck. It cites SK hynix’s expectation that HBM demand will exceed supply through 2030. That is a company forecast, not a confirmed outcome. If supply remains constrained, businesses may face limits on system availability and costs, but the source does not quantify the effect on prices or delivery schedules.

At rack scale, cooling and networking add further requirements. A 72-GPU system needs more than a collection of processors: it also depends on switch trays, power infrastructure and liquid cooling. That makes the package-to-rack view useful for readers evaluating why AI capacity depends on a chain of components and manufacturing steps, rather than on compute chips alone.

Samsung 55-Inch M70H Series 4K HDR Mini-LED Smart TV with HDMI Cable and 12-Month CPS Warranty

Samsung 55-Inch M70H Series 4K HDR Mini-LED Smart TV with HDMI Cable and 12-Month CPS Warranty

Roborock Qrevo S Pro Robot Vacuum and Mop with 18,500 Pa Suction

Roborock Qrevo S Pro Robot Vacuum and Mop with 18,500 Pa Suction

De'Longhi Magnifica Evo Automatic Espresso Machine with Milk Frother, Silver

De'Longhi Magnifica Evo Automatic Espresso Machine with Milk Frother, Silver

As an affiliate, we earn on qualifying purchases.

From Interposer to Rack

The source outlines two approaches to fitting more compute into an accelerator package. Nvidia’s Rubin is described as having two reticle-limited compute dies. AMD’s MI350 Series uses eight vertically stacked accelerator complex dies and two I/O dies, according to AMD; the article gives the MI355X transistor count as 185 billion. AMD says its next-generation MI450X and MI455X use stacked 3D hybrid-bonded compute dies. These are vendor descriptions, and the designs differ, so the counts should not be treated as direct performance comparisons.

The interposer’s size is tied to the limits of chipmaking equipment. TSMC says its CoWoS-S packaging can use an interposer up to about 2,700 square millimetres, or 3.3 times a reticle. For larger packages, CoWoS-L connects silicon bridges in a carrier layer. The source reports that TSMC’s chief executive said on Oct. 15, 2026, that the company was in real-world production of five-reticle-size packaging and was working on nine- and 14-reticle sizes. These statements describe manufacturing capability, not how many units are available to customers.

HBM stacks are also changing in height and density. The article cites SK hynix on thinning DRAM chips for 12-layer stacks and Samsung on a seven-micrometre gap in its 12-high HBM3E. It also references JEDEC’s April 2025 HBM4 standard, which allows stacks of four, eight, 12 or 16 layers and up to 64 GB per stack. The explainer’s central contribution is to place these separate specifications in one visual sequence, from memory tower to complete rack.

What the Model Cannot Establish

The 3D model is explicitly schematic in size; the source says the companies do not publish measurements for the dies, package or memory stacks in millimetres. It shows layer order and counts, not a dimensionally exact reconstruction. Nor does it independently test the cited bandwidth figures or compare real-world model performance across products.

Several points remain dependent on company statements: product configurations, packaging capabilities, cooling specifications and SK hynix’s forecast that HBM demand will outstrip supply through 2030. The source does not give production volumes, customer allocation, pricing or delivery data. It also does not quantify how much any packaging or memory constraint affects the availability or cost of a particular AI service.

Capacity Depends on More Than Chips

The next developments to watch are deployment of newer accelerator systems and whether manufacturers can expand HBM and advanced-packaging output to meet demand. Nvidia lists the Vera Rubin NVL72 as available now, according to the source, but the article provides no shipment totals or independent confirmation of broad availability. TSMC’s stated work on larger CoWoS-L formats and memory makers’ product road maps are also relevant to how future packages may be assembled.

For now, the explainer provides a visual account of the hardware stack rather than a new benchmark or supply assessment. Readers seeking evidence of changes in capacity will need further reporting on manufacturing volumes, system deliveries and actual deployment. The source does not specify when those figures will be available.

Key Questions

What does the 3D explainer show?

It maps an AI accelerator from its substrate and silicon interposer through compute dies and HBM stacks to a cooling plate, then shows how accelerator packages fit into a rack. The model uses schematic dimensions.

What is HBM?

High-bandwidth memory is DRAM arranged in vertical stacks and joined using through-silicon vias. Placing the stacks close to compute dies allows dense, high-speed connections, according to the explainer.

Why does the interposer matter?

The silicon interposer provides dense wiring between the compute dies and memory. The source describes advanced interposer packaging as a possible manufacturing constraint, though it does not provide production-volume data.

Does the article prove HBM will be in short supply through 2030?

No. It reports SK hynix’s forecast that demand will exceed supply through 2030. That is a company prediction; the article does not independently verify future supply or quantify possible price effects.

How many GPUs are in the NVL72 rack described?

The source says Nvidia’s GB200 and GB300 NVL72 systems contain 72 GPUs and 36 CPUs per rack, alongside switch trays, power equipment and liquid cooling.

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

How Microsoft’s School AI Pact And California’s Laws Address Human Oversight

Microsoft is offering school districts an AI contract with human-review terms as California enacts workplace and health care safeguards.

5 AI Voice Cloning Software Guides For A Closer Look At 2027’S Tools

A source-based look at five AI voice cloning guides, from ElevenLabs workflows to Python deployment, and what buyers should verify.

DeepSeek V4.1 Flash 1017: Near-Frontier Agent Performance, Lower Cost

DeepSeek says its V4.1 Flash 1017 update improves agent benchmarks and lowers task costs. Independent testing and current-model comparisons remain limited.

A 2026 Hardware Review Checklist For AI Automation

A 2026 checklist outlines hardware categories for AI automation setups and urges buyers to verify specifications, compatibility and fit.