They do secondary binning; the best ones are overclocked (and more importantly, warranted to perform stably at that overclock) compared to nVidia's specifications.
Frequency gains are mostly linear; a 10% increase in clock speed does indeed yield 10% more frames rendered per second.
Importantly, they don't tend to be over-volted to achieve this overclock even though their coolers are more than capable of dealing with that, but overvolting decreases the usable life of the GPU where "standard" overclocking, generally speaking, does not.
This is a legacy thing, NVIDIA now does binning themselves. It started with Turing and the "A" chips, now it's evolved with Ampere. Partners will buy chips that are already pre-binned into "bad/good/exceptional" tiers (bin 0, bin 1, bin 2). As a result, partners generally no longer apply any special binning to most models of cards, since that mostly covers what they need to know.
And for the Ampere generation, they hardly built any of the lower-tier card models, and those chips have to go somewhere. So they end up using low-tier chips in FTW3/Strix/Gaming X and other high-tier products too. So buying a high-tier model is not any particular guarantee of chip quality anymore either - you can buy a 3090 FTW3 and get a bin-0 chip in it. The only exception is true halo products like the Kingpin/HOF/etc targeted at the XOC market which will definitely get Bin-2 chips and maybe additional binning on top of that. But partners shipped 27 quadrillion FTW3 and Strix and other "flagship" cards and couldn't put good chips in every one of them.
For the mass-market (non-XOC market), everything goes onto the same PCB and you test the finished product and the better ones get the "OC" model sticker. That's it, that's entire binning process for most partners/most of the products on the market. You can have a FTW3 with a bin-0 chip in it. Why do anything more, if you know it's going to sell in 0.1 milliseconds anyway?
The partners that still do binning, typically on limited segments like the Kingpin/HOF/etc, no longer actually test the chip performance, they test the electrical characteristics and then apply statistical analysis to figure out which of the inter-pin voltages/etc are statistically predictive of chip quality. But there is less and less of this because it's expensive and time-consuming, and the differences don't tend to produce significant gains in finished product performance as much as they used to, due to GPU Boost algorithms getting better, etc.
Much like Silicon Lottery exiting the CPU binning business, it's become a bit of a time-waster and partners no longer really bother with it like they used to.
But again, that gets back to the problem with AIB partners: "so what, exactly, would you say you do here?". They made sense in a world where they were allowed some freedom to experiment and service niche product segments/etc. With super tight control, the only thing they are allowed to do anymore is pump up the TDP to try and squeeze out a win, and bigger coolers. Can't do certain kinds of smaller coolers that might compete with the server business either. They just crank up the TDP and slap a bigger cooler on it and crank the price up 20%, it's not really all that significant a benefit to consumers anymore.
They no longer bin. They don't modify the silicon or software in any way. NVIDIA even provides the memory as part of the bundle with the GPU sale. A lot of them use NVIDIA's reference PCB design, and the only common change to that design is custom VRM, which is pointless in a world where overclocking is dead and everyone is undervolting. Cooler design is all pretty functionally-interchangeable across brands, you need something, but a Gaming X is not really different from a FTW3 in any meaningful way. Everything meaningful is done by NVIDIA now.
They are supposed to "buffer inventory" but in practice every time there's a glut they come back to NVIDIA with their hands out looking for refunds for their leftovers and kickbacks on the inventory they keep. So heads they win, tails NVIDIA loses. And it's not like their prices are awesome the rest of the time either.
They no longer even act as a price anchor... 10-series launch, only EVGA and Zotac put out cards anywhere near the "official" MSRP (in hideously limited quantities) and they were still $20-30 over. Everything else was actually higher than the FE cards, in the $725-800 range. Same for the 30-series where NVIDIA cards were the only remotely affordable ones. So these days it's NVIDIA price-anchoring the partners, partners want a 20% higher street price over the NVIDIA design for what seemingly amounts to a higher factory TDP configuration and a bigger cooler. (Which is of course tied to the BOM cost which NVIDIA controls, but... the price structure kinda doesn't matter to the consumer.)
Most of them don't even build the cards themselves... EVGA for example contracts out their assembly to third-party factories in Taiwan, they don't own the factory. So they basically just take the spec from the customer, and give it to the engineers. Well, at least their secretary does that I guess.
So what, exactly, do partners do around here anymore? I realize that a lot of this is the result of product strategy by NVIDIA, some other parts are just the direction that tech has gone over the years (automatic boosting, undervolting, etc), but, regardless, consumers don't really benefit from the things that AIBs are able to offer nowadays. Partners provide almost no real value-add anymore. Functionally they are middlemen, and middlemen get squeezed to zero by the market.
Ironically, in true Office Space fashion, probably the strongest argument is that they deal with the fucking customers so NVIDIA doesn't have to, EVGA could still distinguish itself on warranty/etc. But that's really the only meaningful distinction between AIB partners anymore and it was really only a couple of them that had worthwhile service. With EVGA gone, that argument starts to fall apart too.
