Skip to content

Packaging, testing, and advanced packaging ​

After reading this section, you will be able to name the four steps a wafer goes through after it is finished, and explain why one of the bottlenecks for AI chips lies in "packaging" and "HBM".

After the wafer is done: test, dice, package, test again ​

Line them upAfter the wafer leaves the fab — click in order
  1. 1?
  2. 2?
  3. 3?
  4. 4?

Click them one by one, in order.

Wafer testing comes in two kinds: one tests only a small number of dedicated test structures on the wafer (parametric testing), while the other performs full functional testing on every single chip (also called die sort). A prober presses a "probe card" covered with tiny probes onto the wafer. In the past, bad chips were marked with an ink dot; now the record is kept in a file called a wafermap. Chips with some functions broken can sometimes still be sold as cheaper models with reduced capacity.

Why test before packaging? Because packaging itself is expensive—pick out the bad ones first, and you don't have to pay packaging costs for them. (When yields are very high, there is also a practice of packaging without testing first, called "blind packaging".)

A die is a small piece cut from the wafer, carrying a complete copy of the circuit. The English plural has three spellings: dice, dies, and die; a photo of a die is called a die shot.

What packaging does ​

Packaging is the last step of manufacturing: putting the die into a casing to protect it from damage and corrosion. According to Wikipedia, it has three jobs:

  • Connection: connecting the chip to the circuit board through pins, solder balls, and so on. An advanced processor may have thousands of connection points in a single package.
  • Protection: against impact, chemical contamination, and light exposure.
  • Heat dissipation: a cheap plastic package can dissipate about 2 watts of heat, while a comparable ceramic package can handle up to 50 watts.

Packaging also affects speed: Wikipedia notes that the delay caused by packaging can account for nearly half of the total delay in high-performance computers.

How is the die connected to the package? Two basic methods:

  • Wire bonding: using extremely thin metal wires to connect the pads at the chip's edge to the package; the material is shifting from gold to copper.
  • Flip chip: growing small bumps on the chip's pads, then flipping the chip over and soldering it directly onto the substrate—the connections are shorter and more numerous.

Advanced packaging: chips no longer rely only on "getting smaller" ​

Transistors are getting harder and harder to shrink, and more and more expensive (as covered in the previous section); another path is: putting several chips together more tightly.

Look it upIn advanced packaging…

Click one to see.

Wikipedia also records a detail: after TSMC built its fab in Arizona, USA, in 2023-9, analysts pointed out that chips produced there still had to be sent back to Taiwan for packaging.

HBM: stacking memory right next to the GPU ​

AI chips constantly read and write massive amounts of data, and memory speed often becomes the bottleneck. The approach of HBM (high-bandwidth memory) is:

  • several layers of DRAM dies stacked on top of each other, connected vertically through through-silicon vias and microbumps;
  • placing them right next to the GPU via an interposer;
  • Using an extremely wide interface: the first generation already had 1024 bits per stack, while traditional graphics memory (GDDR) has only 32 bits per channel.

Generation-by-generation progress (bandwidth per stack):

GenJEDEC standard dateBandwidth per stack
HBM2013-10128 GB/s
HBM22016-1256 GB/s
HBM2Eend of 2018up to 307 GB/s
HBM32022-1-27about 819 GB/s
HBM3E2024 volume productionabout 1 to 1.2 TB/s
HBM42025-4up to 2 TB/s (2048-bit interface)

Example: NVIDIA's H100 uses 5 stacks of HBM3, totaling 80 GB and 3 TB/s of bandwidth. The first GPU to use HBM was AMD's Fiji (Radeon R9 Fury X) in 2015-6; HBM development began at AMD in 2008, with partners including SK Hynix, UMC (interposer), and Amkor and ASE (packaging).

Who makes it: According to Wikipedia, the largest HBM makers in 2025 were SK Hynix, Samsung, and Micron; Chinese Wikipedia states that the two Korean companies together hold over 95% of global HBM capacity. TSMC manufactures the base logic die for HBM and formed a chip alliance with SK Hynix in 2024.

It crowds out ordinary memory: Micron notes that the wafer-capacity conversion ratio between making HBM and making DDR5 is 3 to 1. Wikipedia records that ordinary memory prices have risen more than 200% since early 2025. This also explains the market figures we will see in the next chapter: in 2026, memory chips are projected to grow far more than any other category.

Quiz timeWhy is HBM so much faster than traditional memory?

Check yourself ​

  • I can name, in order, wafer testing, dicing, packaging, and final testing
  • I can explain what a chiplet, an interposer, and a through-silicon via each are
  • I can explain why HBM is fast, and why it drives up ordinary memory prices

The next chapter looks at the money and people of the whole industry: how big the market is and where the money flows.

gaodong.aiAboutPrivacyContact

Independent tutorials, not affiliated with any product mentioned. Product names and trademarks belong to their owners. Check official pages for prices and policies.