6 Innovative Approaches to Overcome Transistor Scaling Challenges
Transistor scaling is hitting tough limits, from rising heat to power loss and shrinking space. Experts in the field share six practical approaches, including gate-all-around designs, backside power delivery, chiplets, and vertical building. These ideas show how chipmakers can keep improving performance when traditional scaling falls short.
The perspectives in this article were written by the named contributors and submitted through Connectively. Connectively's automated system selected them and used AI to write the headline, introduction, section headings, summary sections, and illustration, and to make minor edits for clarity. Articles are assembled and published automatically. Report errors to corrections@connectively.us. How we make our content
AZ
AY
KS
OM
RL
Use Gate-All-Around for Cooler Phones
Lilach BullockAI Implementation Consultant and Fractional CMO, Lilach BullockThe challenge with shrinking transistors is leakage and heat, because current slips through even when a transistor is switched off and the chip runs hotter. The approach I'd watch is gate-all-around design, where the gate wraps around the channel and controls the current more tightly. If you're explaining this to buyers, leave gate length out of it. Tell them the chip wastes less power while it sits idle, so it makes less heat and the battery lasts longer. Describe it as a phone that stays cooler in a pocket, because a buyer can picture that.
Route Power From the Backside
Ace ZhuoCEO | Sales and Marketing, Tech & Finance Expert, TradingFXVPSTo scale further, wiring up transistors has become harder than scaling up transistors themselves. For very small metal pitches, the thin copper wires end up using a lot of their cross-section area to scatter electrons as well as to host a barrier "liner" layer. This results in wiring up to scale much faster than the transistor scaling itself. In particular, supply rails for power end up competing with signal wiring for front-side metal, and can end up dominating the IR "droop" that determines how much current you can run on a core.
I see this in our latency sensitive trading racks where the 'advertised' boost clock looks great but in sustained all-core load tests frequency drops precipitously. For trading order execution speed, this matters far more than raw peak speed.
My suggestion would be to go with backside power delivery, using nano-TSVs to power the transistors in the chip from the back side of the die, leaving the front side to be used for signal routing only. I mentioned that Intel had a 6% frequency increase in their trials of their new PowerVia backside power delivery technology, as well as greater than 10% increase in utilization of the die surface, and I see similar benefits for TSMC's similar technology. This will end up being a larger real benefit than squeezing a few nm out of the gate. And go with alternative metals for the interconnects on the smallest sizes of metal (ruthenium is one such metal) -- copper's liner overhead ends up being way too much on these smallest sizes. Fix the wiring and power distribution, and the gain in the transistor itself will actually show up in real app-level testing, rather than just in a set of benchmarks.
Solve Heat Through System Co-Design
The challenge I see most directly is heat. For decades, smaller transistors also meant lower power per transistor, so chips got faster without getting hotter. That relationship broke down years ago. Transistors keep shrinking, but leakage current and power density don't fall at the same rate, so packing more of them into the same area concentrates more heat in less space. Today the practical limit on many chips isn't how many transistors you can fit, but how much power you can deliver and how much heat you can remove.
We feel this at the data center level. Each new GPU generation brings more compute, but also higher power draw per chip and per rack, which pushes the industry toward liquid cooling and much heavier power infrastructure.
The approach I'd point to is moving part of the scaling problem from the transistor to the package. Chiplets and advanced packaging split a large design into smaller dies, each built on the process that suits it best, and connect them with very high bandwidth. Stacking high-bandwidth memory right next to the compute is a good example: it cuts the energy spent moving data, which is now a big share of total power in AI workloads. Newer transistor structures like gate-all-around help with leakage too, but they won't solve the problem alone.
My advice is to treat it as a system problem rather than a transistor problem. The biggest gains now come from designing the chip, packaging, memory, cooling and software together, so that every watt goes into useful work.
Split Functions Across Optimized Chiplets
Kruno SulićFounder & SaaS Product Builder, ClipriseOne specific challenge with scaling down transistor size is that once features get extremely small, you run into leakage current and much tighter manufacturing variability. In simple terms, electrons become harder to control, so transistors may switch less predictably, waste more power at idle, and generate more heat than the size reduction alone would suggest. That is where the old assumption of "smaller automatically means better" starts to break down.
The most practical innovative approach I would recommend is not relying on transistor shrink alone, but combining selective scaling with chiplet-based design and advanced packaging. Instead of forcing every function onto one monolithic die at the newest node, you split the system into smaller specialized blocks. The parts that truly benefit from the most advanced process, like compute-heavy logic, can use it, while analog, I/O, cache, or support functions can stay on more mature and cost-efficient nodes.
That approach helps overcome the main limitation in three ways. First, it reduces yield risk, because smaller dies are generally easier to manufacture reliably than one very large leading-edge die. Second, it improves economics, because you are not paying the highest node premium for blocks that gain little from it. Third, it can improve performance per watt by shortening interconnect distances at the package level and letting designers optimize each block for its actual job rather than pushing one process technology beyond its efficient range.
From a product-building perspective, this is similar to modular software architecture. When one layer becomes too expensive or fragile to keep scaling linearly, the smarter move is often to separate the system into optimized components with clean interfaces. In semiconductors, that modularity is becoming one of the most realistic paths to continue meaningful performance gains even as pure transistor scaling delivers smaller returns.
Balance Leakage Control, Speed, and Yield
Omer MalikCEO, ORM SystemsOne of the biggest challenges in shrinking transistor size is controlling current leakage as the gate becomes smaller. At extremely small dimensions, quantum tunnelling and short-channel effects make it harder for a transistor to switch off completely. This increases standby power consumption and heat generation, limiting the performance gains manufacturers can achieve. A smaller transistor only creates real value if it delivers better computing performance without introducing unacceptable power or reliability costs.
I recommend gate-all-around (GAA) transistor architecture, particularly designs using stacked nanosheets. Unlike conventional FinFET structures, GAA transistors allow the gate to surround the channel on all sides, improving electrostatic control and reducing unwanted leakage. Nanosheet designs also offer flexibility in adjusting channel width to balance performance and power efficiency. The main tradeoff is manufacturing complexity, including tighter process tolerances and more demanding fabrication steps.
I would evaluate this approach using three practical measurements: leakage current at a defined operating voltage, switching performance under comparable workloads, and manufacturing yield. For example, a design that improves transistor density but requires substantially more cooling or produces lower fabrication yields may offer limited commercial benefit. Device architecture, production cost, and thermal management must be evaluated together.
The next breakthrough in transistor scaling will come from controlling power and current more precisely, not simply making transistors smaller.
Build Upward When Physics Hits Walls
Runbo LiCEO, Magic Hour AII'm Runbo Li, co-founder and CEO of Magic Hour. The biggest challenge with scaling down transistors isn't manufacturing precision. It's quantum tunneling. When you shrink a gate below a few nanometers, electrons stop behaving like obedient particles moving through circuits. They start leaking through barriers they shouldn't be able to cross, like ghosts walking through walls. That means your transistor can't reliably switch between on and off. And if you can't do that, you don't have computing. You have an expensive heater.
I think about this the same way I think about hitting walls in AI video generation. When you push a system to its physical limits, you don't keep brute-forcing the same axis. You change the dimension entirely. That's what needs to happen with chip design.
The most promising approach is gate-all-around (GAA) transistor architecture. Instead of a gate sitting on top of the channel like a lid on a pot, you wrap the gate completely around the channel, like a fist gripping a rope. Samsung started shipping GAA-based 3nm chips, and TSMC is moving that direction for their 2nm node. By surrounding the channel on all sides, you get dramatically better electrostatic control. Leakage drops. Power efficiency goes up. You buy yourself several more generations of scaling before physics shuts the door again.
But here's what excites me more. The real innovation isn't just the transistor shape. It's the shift toward heterogeneous integration, stacking different types of chiplets vertically and connecting them with through-silicon vias. You stop trying to cram everything onto one impossibly small 2D plane and start building upward. Apple's M-series chips already hint at this philosophy, combining CPU, GPU, and memory on a unified package.
The lesson applies everywhere, not just semiconductors. When you hit a wall, don't drill harder into the same wall. Find a new dimension to build in. That's how two people build a platform with millions of users, and it's how the chip industry keeps Moore's Law on life support for another decade.