Skip to main content

AMD, Nvidia graphics could double in performance with new ‘stacked’ tech

A new technology introduced by Taiwan Semiconductor Manufacturing Company (TSMC) could boost the power of graphics cards by Nvidia and AMD without making them physically larger. The technology is called wafer-on-wafer, and mimics 3D NAND memory technology used in modern solid-state drives by stacking layers vertically rather than spreading the hardware horizontally across the printed circuit board, which would require additional physical space. 

So what’s a wafer? Unlike your favorite snack, it’s a thin slice of polished semiconductor material that serves as the foundation for a crisscross of layered copper wires that convey electricity, and the transistors that are the heart of the processor. The wafer and mounted components are cut by a diamond saw into single chips and placed into the physical processor package you see when you break open the desktop. 

Recommended Videos

Right now, graphics chips produced by Nvidia and AMD rely on a single wafer. But TSMC, the largest dedicated independent semiconductor foundry on the planet, discovered a way to stack two wafers in a single package. The upper wafer is flipped over onto the lower wafer, and then both are bonded together. Moreover, the upper wafer contains the in/out connection piercings (aka thru-silicon vias), thus the duo is packaged using flip-chip technology. 

Get your weekly teardown of the tech behind PC gaming
Check your inbox!

According to TSMC partner Cadence, the technology could see two sets of wafers connecting to each other in a cube-shaped package using what’s called an interposer, an electrical interface that routes one connection to another. More than two wafers could be stacked vertically as well, with all but one wafer sporting the in/out thru-silicon vias connections. 

While this is a lot of tech talk, it’s basically describing how graphics chips can be scaled vertically, not horizontally, using TSMC’s technique. Not only can you cram more cores into a single graphics chip, the communication between each wafer would be extremely quick.

Thus, instead of tweaking an architecture and rebranding the product as a new family, manufacturers could potentially stack two or more current GPUs on a single card as a product refresh. The operating system would detect it as a single card, and not as a multi-GPU configuration. 

With 3D NAND, memory cells are stacked vertically and connected together via makeshift data elevators. This method enables manufacturers to provide additional storage capacity while staying within the same physical constraints. This design is faster, too, given the data travels up and down the memory tower rather than hunting down its destination using horizontal “city streets.” 

The problem with stacking processor wafers may be in the overall manufacturing yields. One of two wafers could pass, but because the other wafer is bad, both would be discarded. This method could prove to be too costly on low-yield products and would need to be used on production nodes with high manufacturing yields, like TSMC’s 16nm process technology. 

TSMC introduced its wafer-on-wafer technique during its symposium in Santa Clara, California. The company also revealed a partnership with Cadence for 5nm and 7nm+ process technology for high-performance and advanced mobile computing. 

Kevin Parrish
Former Digital Trends Contributor
Kevin started taking PCs apart in the 90s when Quake was on the way and his PC lacked the required components. Since then…
Nvidia might dominate the GPU market in January
The RTX 4090 graphics card sitting on a table with a dark green background.

Nvidia, AMD, and Intel are all set to update their lists of the best graphics cards sometime in January. Now, according to HKEPC, a Hong Kong media outlet, we know that the Nvidia RTX 5080 is reportedly set to launch on January 21. Meanwhile, AMD may take more time to ramp up its GPU availability, giving Nvidia the chance to dominate the GPU market in early 2025.

With CES 2025 mere days away, we already know that January is going to be one busy month for PC enthusiasts. Nvidia's got the RTX 50-series, AMD is launching the RX 9000 series, and Intel will add one more GPU to its Battlemage lineup. Only the last one has a confirmed release date of January 16, but it seems that Nvidia will be just one week behind -- if the leak is to be believed.

Read more
Nvidia’s new GPUs show up in prebuilts, but the RTX 5090 is missing
iBUYPOWER RTX for AI PCs side view of pre-built on sale hero

Nvidia's upcoming RTX 5080 and RTX 5070 Ti just appeared in several iBUYPOWER gaming PCs. This is the first U.S. retailer to list Nvidia's RTX 50-series in prebuilt systems. The listings are interesting, with performance figures that really don't add up. Still, the biggest question is: Where's the GPU that's bound to beat all the current best graphics cards? Yes, we're talking about RTX 5090.

The listings have already been taken down, but they were preserved by VideoCardz. A total of five systems were listed by iBUYPOWER, but they all contained the same two GPUs -- either the RTX 5080 or the RTX 5070 Ti. Both cards are said to come with 16GB of memory, and we expect them to be announced on January 6 during the CES 2025 keynote held by Nvidia's CEO, Jensen Huang.

Read more
The best PC gaming feature of 2024 didn’t come from Nvidia, AMD, or Intel
Forza Horizon 5 on the Sony InZone M10S.

One of the great benefits of PC gaming is the ability to take matters in to your own hands. Not enough games support your favorite feature? Unhappy with the frame rate your PC is producing? There's all sorts of applications that can let you tinker and optimize your PC gaming experience.

Many of these are rather niche, but there's one that I would recommend just about every PC gamer install. It's called Lossless Scaling, and if you haven't already heard of it, I'm about to make your day.
How Lossless Scaling scaled up

Read more