High-tech products typically rely on economies of scale and massive up-front investment. This creates odd situations. Tesla hides its high-performance mode behind a feature flag. The hearing aid manufacturer Unitron uses the same practice to lock down advanced Speech in Noise algorithms. These devices are effectively derated from their maximum hardware capability. Beyond consumer outrage, this occasionally creates markets for modding.
A lot of manufacturing cost has to be invested up front. In R&D and in building incredibly complex production lines. The fewer variations a factory has to produce, the cheaper it is to construct. So manufacturers reduce the number of pathways by making reusing the same parts across different models and sometimes making upgrades pure software changes. They end up effectively selling what is theoretically more valuable hardware in a cheaper product. This makes economic sense for the manufacturer. But it also creates a gap. Usually this gap is occupied by hobbyists, which aren’t purely rational economical actors. Sometimes though demand is so large, that profit oriented actors take notice.
Computer hardware manufacturers often use derating. With the AI boom demand for hardware has skyrocketed and with it. Nvidia’s RTX 4090 GPUs have almost the same silicon chips on them as their RTX 6000 Ada. The 6000A however has double the VRAM. That makes the 6000A an interesting chip for running AI models1.
Unlike the other examples, this isn’t a simple software flag. An upgrade requires a custom board with the bigger RAM capacity, desoldering the chips from the original GPU and soldering it onto the custom board. Finally it needs custom firmware for the chips to recognize the additional memory. This is an involved process with specialized equipment and quite a bit of R&D2. All to upgrade a $1'700 GPU to a GPU worth $6'900 GPU (prices in MSRP).
The equipment isn’t very expensive. With $3,000 one can acquire solid equipment for the job: a BGA rework station (to desolder, align and resolder chips), a good microscope and a hot air gun. Finally, of course, it demands quite a bit of skill. What can be more difficult is procuring the parts needed for the upgrade. The shopping list includes a custom “Clamshell” PCB, the extra memory chips, a new cooling solution and some other consumables like stencils. The parts cost about $480. But with the expertise and the supply chain in place, this is a very scalable and attractive business model. Modders would likely make somewhere around $3,000 profit per upgraded card3. The supply chain has been in place in China for a few years now, and it seems they keep finding new and more profitable ratchets4.
The larger shops might no longer be doing these upgrades as with the massively increased 4090 prices, the profit margins have shrunken. The at least 12 months it lasted, this was an incredibly attractive business model5.
The tech enthusiast mainstream is critical of manufacturers' derating measures, especially when they’re purely in software. Often though, it allows for a cheaper end product, which benefits a broader set of consumers.
Modders put a lot of energy into reverse engineering and circumventing such deratings. Ultimately, it isn’t a net productive system. Manufacturers spend capital on these measures. Modders then work hard to reverse them. This creates a big opportunity cost. It is an interesting quirk of capitalism: resource efficiency is encouraged, but only within the boundary of a company.
-
The 6000A has 10% more CUDA cores than the 4090. So the 6000A is slightly more powerful even after the upgrade. For many workloads though, the main constraining factor is VRAM though. Since a lot of AI workloads are memory-bound. On the other hand the 4090 has slightly faster RAM, which is a small bonus (GDDR6X instead of GDDR6). Especially for these memory bound workloads. ↩︎
-
This post is a great, short summary of the process and has a video of how this is done. It shows the perspective of an enthusiast having picked up on what was happening at likely industrial scale in China for a long time already. ↩︎
-
Since there was massive excess demand in 2024 for the higher memory 6000A, the price one would be able to fetch for an upgraded 4090 was probably not that far from an original 6000A. Performance is very similar on both modded and OEM version. 6000A’s were selling way above MSRP too. Reliability of the modded versions likely drew the price down a bit. As a rough estimate I’m assuming the following Profit = $6'900 * discount - $1'700 - parts - labour. Where the discount is 20%, the cost of parts is $480 and labor is estimated at $340. ↩︎
-
There is a longer history to GPU modding supply chains in China. An earlier, well-documented, industrial-scale example are Mobile-on-Desktop GPUs in 2021. Nvidia had introduced Lite Hash Rate limiters to make their Desktop GPUs less attractive to crypto miners. The mobile GPUs didn’t have this restriction, however, which made it attractive for shops to repackage the mobile cards onto desktop PCBs with better cooling. In 2024, US export controls barred professional-grade GPUs from being exported to China. This created huge demand for alternative AI chips. Chinese shops started disassembling RTX 4090s to pack them more tightly, making them viable for datacenter use. This repackaging was likely done at a large scale. They were desoldering the main chip and RAM and resoldering them onto a better PCB package. The VRAM upgrade was just the next evolution. ↩︎
-
Currently though the RTX 4090 is going for around $2100 used. Prices have been stuck on a similar level for the last year. So there is still demand for the 4090. Even as the market for its 5090 successor has moved down to now be around $3300 new and around $2900 used. The upgrades were first confirmed in August 2024 in a Chinese cloud datacenter. Meaning at that point these upgrades had likely been going on for a long time. In November 2023 there was clear evidence for modding of the 4090. These upgrades must have driven some of the demand for the cards. Economically it stopped making economic sense at the beginning of last year though. So it doesn’t seem plausible that these upgrades are still creating direct upwards price pressure in the market. ↩︎