GPUs The Great NVIDIA Switcheroo | GPU Shrinkflation April 8, 2025 Last Updated: 2025-04-08 We look at how NVIDIA has downsized essentially all of its gaming GPUs in terms of relative configuration compared to each generation’s flagship The Highlights This article expands upon our "RTX 4080 problem" by looking at the entirety of the RTX 50 series, including how the RTX 5070 looks an awful lot like a prior 50-class or 60-class GPU.NVIDIA is giving you the least amount of CUDA cores for a given class of GPU than ever before.GPU prices have crept higher across the board, but NVIDIA's, in particular, have lost step with what we came to expect from generations of GPU launches. Table of Contents AutoTOC Our fully custom 3D Emblem Glasses celebrate our 15th Anniversary! We hand-assemble these on the East Coast in the US with a metal badge, strong adhesive, and high-quality pint glass. They pair excellently with our 3D 'Debug' Drink Coasters. Purchases keep us ad-free and directly support our consumer-focused reviews! Intro NVIDIA is giving you the least amount of CUDA cores for a given class of GPU than ever before.  Today an RTX 5070 is comparable to a GTX 950 (watch our review) in some ways when you run some numbers. An RTX 5080 isn’t distant from a 2060 (read our review) in some considerations. The relationship between the number of CUDA cores the flagship has and the number of CUDA cores the lower-tier GPUs has been getting worse basically across the board. The amount of money you have to spend, even adjusted for inflation, to buy the GPUs has been staying flat or rising.When this happens in any other product category it’s called shrinkflation. Editor's note: This was originally published on April 3, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication. Credits Host, Writing Steve Burke Writing, Research Jeremy Clayton Video Editing Vitalii Makhnovets Writing, Web Editing Jimmy Thang Chart - Percentage of CUDA Cores Relative to Generational Flagship We’ve already talked about the pricing issues and availability, but what we’re doing now is revisiting a topic that we ran about 2 years ago in a video called the RTX 4080 problem, in which we explored why no one was buying 4080s (watch our review) at the time. It wasn’t just the money but because the relationship of what you got for the money. We’re taking the concepts where we broke out the pricing, the components, the die area, etc. and applying it to the 50 series and, in short, it has not gotten better.   We just have 2 main charts to go through in this article but they’re really interesting. Now that NVIDIA has shipped everything except for the 60-class card, we’ve got a good amount to look at. The real goal of this is to explore the relationship between the money and what you get for it, but we’re also going to compare some of the cards against prior generations and doing some inflation adjustments.   We started working on this piece for the 5080 launch and then realized it’s going to get worse. So, we waited for the 5070, which is now here (unfortunately). Let’s get into the data for this. This chart compares the percentage of the Flagship CUDA core count that each configuration is. Due to architectural changes, we’re not interested in the raw count of CUDA cores, but the percentage occupancy of the maximum config for the flagship die. Our chart plot tracks each same-named GPU class across NVIDIA generations on a percentage scale representing how many CUDA cores each has relative to a larger configuration. The GPUs are all shown relative to the CUDA core count of that generation’s top gaming, non-Titan GPU – the 5090, 4090, 2080 Ti (watch our review), 1080 Ti (read our revisit), and so on – which we’re calling the Flagship-class. If you see 100% anywhere, that means it is equal in CUDA core count to the flagship. We went with the 3090 (watch our review) for the 30 series rather than the late-arriving, cash-grab, full die 3090 Ti (watch our review). We had to make a judgment call. The Flagship-class is plotted relative to the largest die’s maximum possible core count. The GTX 780 Ti (watch our revisit) is one of the few exceptions where NVIDIA made a flagship with the full non-cut-down die.  The GTX 780 (watch our revisit) had 80% of the CUDA cores that the flagship 780 Ti did. The RTX 2080 (watch our review) brought that down to 68% of full CUDA count, then just 59% for the 4080, but it gets worse. The 5080 is a mockery of an 80-class card with only 49% of the flagship-class CUDA count configuration. We don’t care if the die is different or not for this chart, just the config. The 3080 temporarily bucked the trend at 83%, which was great. This correlates with its incredibly good value and performance at launch with positive reviews. Back in our review of the 3080, we said, “The card performance overall is impressive. It’s a big recovery from the 20 series when we reviewed it and called 3 of the cards a waste of our time because they were 1080 Tis and then complained for 55 days about how there was no RTX and the cards were named RTX. So this was a big turnaround for NVIDIA.” That also, however, aligns with the reviews we and others gave to the 3090 and 3090 Ti. For example, in our 3090 Ti review, we stated, “For us, hard pass on this. 8-12% for $2,200 is insane.” And with an overclock back then, we were able to nearly equate the 3090’s performance with the OC 3080. That’s how close they were. The odd 80 Ti/Super class from the 20 series to 40 series occupy the space between the 80 class and the flagships. There’ll likely be another between the 5080 and 5090.  The Super refreshes really should be called the “oopsies” edition GPUs. NVIDIA rolls these out when they make an “oopsies” on price and public sentiment, using Supers to meet halfway on price. Our hope is that the 5080 and 5090 gap ends up again as an “oops, let’s fix this” rally from NVIDIA with a mix of the 2080 Super’s (watch our review) or 4080 Super’s relatively sane pricing along with the 3080 Ti’s (watch our review) aggressive configuration. That might start to help fix this a little bit.  The 70 Ti/Super class drops hard. The 1070 Ti (watch our review) had a 68% CUDA core configuration for this class, falling to just 41% for the 5070 Ti (read our review). By this logic, the 1070 Ti offered far more GPU relative to the 1080 Ti than the 5070 Ti is to the 5090. Next, we’ll expose NVIDIA’s grand switcheroo between the 70 and 80 class GPUs.  From the 770 series to the 3070 (watch our review), the CUDA core count of the 70-class cards once reliably was between 53-59% of the flagship’s CUDA core count.  Then the 4070 bore only 36% of the CUDA cores the 4090 had, and the card falling in the 50-60% range was now the 4080. Moving into the present, the RTX 5070 has an anemic 28% of the flagship’s configuration. If you were to extend the 70-class line out on its previous trend, you’d arrive around the same position as where the 80-class is now. Strictly speaking in proportions and if we want to do funny percent math, the 3070’s (watch our review) core allocation relative to its respective flagship was 100% higher proportionality against the 5070’s. The 60-class is where it gets really bad. That 28% figure for the 5070 is lower than almost every 60 class configuration. The 60 class traditionally occupied the 30-40% range with a high outlier in the 20 series at 44%. This tracks with the fact that we were softly positive on the 2060 at its release – more positive than the 2080. The 3060 returned to the low 30% range, but the 4060 got slashed to 19%. Here’s what we said of the 4060’s worse cousin, the 4060 Ti, “The RTX 4060 Ti 8GB is one of the worst GPU launches from NVIDIA that we’ve ever covered.” And that brings us to the 50-class. The 19% on the 4060 is where the 50-class has sat multiple times. NVIDIA covered this segment during the 20 series with the 16 series GPUs, which we didn’t plot for sake of simplicity. Moving forward, the 3050 was a 24% configuration, and it’s no wonder why the 4050 got canned – it would be like scraping the bottom of the barrel so hard that you just get splinters. But what’s crazy is that the 5070 barely clears the 27% config of the old GTX 950. That’s just sad. Now that we’ve established the trends, let’s keep all of that in mind and analyze pricing in the same way. Chart - Inflation Adjusted Prices Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) This line plot tracks the launch price of all the same GPUs in each class, adjusted for inflation from the month of each GPU’s launch to January 2025. Right away, we see that the flagship class has changed massively. The 780 Ti, 980 Ti (watch our revisit), and legendary 1080 Ti fall within a consistent $100 spread. The 980 Ti was slightly cheaper at a $650 launch price, which is $865 after the inflation adjustment. The 1080 Ti sits at $912, in stark contrast to the massive jump of the 2080 Ti at $1,510 adjusted. That’s a 66% cost increase gen-on-gen for the customer. It was technically available for $1,000, but in very limited quantities, and the vast majority went for $1,200, which is what we adjusted from. The price went up again for the 3090, with a slight relief in the 4090, before jumping again to the $2,000 mark with the 5090. It undoubtedly costs more for NVIDIA to make a 5090 than it did to make a 1080 Ti, but there’s no argument that more than double the retail price is painful for a consumer. The 80 class has also risen, though not to the same extreme degree as the flagship class. The GTX era 80s had inflation-adjusted prices between $734 and $886. There was a slight bump to just over $1,000 in the 20 series, followed by relief to the mid-$800s at the 3080, before rising insurmountably to the 4080. When taken alongside the CUDA core configurations, all of this underscores both just how good the 3080 was and how terrible the 4080 was. The 3080 had a spike in core allocation and a return to “normal” pricing, while the 4080 fell off core config cliff and the price went up at the same time. Coming to the present, the 5080 is back at the same relative price as the 2080, but at a much worse relative CUDA core count. The 80 Ti/Super class is an oddball – as if NVIDIA hasn’t been able to decide whether it’s better as a later, better value 80 class alternative like in the 20 and 40 series, or if it should be a weirdly positioned, poor value cash-grab like the 3080 Ti. The 70 Ti/Super class has risen in price across the generations that it’s existed, from roughly $500 at its introduction in the 10 series to $849 in the 40 series. AMD Radeon GPUs were competitive in this price bracket back in the 10 and 20 series days, which is likely the reason why we see this aggressive pricing during that time period. From the 30 series onward, NVIDIA’s dominance has allowed this class of card, specifically, to sit comfortably between the 80 and 70 classes. The 70 class has managed to stay relatively flat from one end of the chart to the other. The all time low price was in the GTX 900 series at $440, and the high point was the 20 series at about $750. That’s a large swing, but it’s stayed relatively flat since then. The 60 class paints a similar picture. The inflation adjusted price line is generally flat overall with a slight downward trajectory since the 20 series, but in that same time the core config has gone into the dumpster. We don’t know anything about a theoretical future 5060, but we’d bet it won’t be a pleasant addition to this data set. Finally, the 50 class hasn’t seen much action, but it hasn’t seen many releases in recent years – probably because the 4060 took its actual place. Judging by the 3050, NVIDIA is probably unwilling to launch a GPU for under $250 again, let alone the $145 mark of the 1050 (watch our review). Additional Segmentation Over the years, the means of product segmentation have migrated. Product segmentation isn’t inherently an evil thing, and especially in the world of silicon where the costs are enormous to make any of these products, but it can be applied in ways which just don’t feel good as a consumer. Segmenting the 1080 Ti at 11GB versus the Titan cards at 12GB didn’t feel particularly bad. It was obvious what they were doing, but the affected user base was much smaller. Some of the other ways NVIDIA has historically segmented its products include splitting double precision out into only the highest-end cards, which at one point included Titans. Another is by forcing users over to Quadro for verified drivers as an additional layer of liability reduction for big organizations. Neither of these two segmented features are noticeable to the vast majority of end users, so it doesn’t feel as bad to the consumer. Over time, that has drifted to VRAM increasingly, which now means there is a new developing class of users.  For the gaming audience, we get situations where a $750 video card can find itself in situations of unplayable stuttering and latency nearing 800 ms PCL due to VRAM overload and swapping. Joining the scientific user base that once needed double precision, or now might need various machine learning capabilities, there is now the segmented customer base of so-called “creators.” Not just YouTubers, but anyone making 3D art, games, or similar media. These users are being pushed into the 90-class, which is further diminishing the capabilities of the highest-end gaming cards or pushing those high-end gaming consumers into price categories of professionals who use their GPUs to make money. It’s easier to shrug it off knowing it’ll make back the time, even if it’s still unpleasant. “Arbitrary” Naming Back in our RTX 4080 Problem video, we talked about how all of this is predicated on the assumption that the names mean anything. Like Whose Line Is It Anyway, sometimes it feels like the names are made up and the prices don’t matter. We’ve been open about our opinions about this changing over the years: At one point, it did feel like names were somewhat arbitrary. It is just a name, and it’s ultimately the specs and price that matter. But the shift came over the last couple generations, where we came to appreciate that what’s in a name is important. NVIDIA has used the 80-class cards to establish an expectation in customers, and regardless of whether NVIDIA intends it to still be perceived as the high-end as opposed to some mid-range card (which it is now), the fact is that their consumers do perceive the 50 name as intended to be high-end. This is sort of a death of the author scenario, but then NVIDIA doesn’t want to name a $1,000 video card a “5070.” That creates new problems.  NVIDIA has died as the author, and the consumer is now in control over what these names mean. To quote someone in the industry, it’s the “perception of reality” versus the reality.  If NVIDIA wants to establish a reality where an 80-class card is half of a 90-class card, they can do that; however, if the end users perceive that a 5080 should be a true high-end device, that’s all that actually matters. NVIDIA is also responsible for this. The company spent a decade establishing the 80-class cards as the top-of-the-line, behind only the Ti class cards. It has now bifurcated those two lines and created a large gulf between them. And this is getting worse with 5070 cards that are now more similar to older 50-class cards. And so while the name itself is technically arbitrary as compared to the specs, the name matters. It defines an expectation.  Let’s explore that philosophy a bit more. If Toyota suddenly starts shipping rebadged Yugos that it calls Camrys, that’s going to cause problems with the customer base. That’s what NVIDIA is doing. If the AMC Gremlin is sold under literally any name, it’s going to cause problems. The point is, the RTX 5080 is a Yugo. Or a Gremlin. Or a Ford Pinto. And NVIDIA has spent a decade branding it as a supercar (and it was at one point a supercar).  Conclusion Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. Zooming back out, we think the overall picture is clear. NVIDIA has downsized essentially all of its gaming GPUs in terms of relative configuration compared to each generation’s flagship. All of the lines go down. The chart from earlier had a lot of words to say one thing: Line go down = bad. We don’t want the line to go down. We want the line to stay the same or go up. The 80 class is now in line with former 70 class GPUs and the 70 Ti/Super class is now in line with former 60 Ti class territory. The last 60 class card was configured like a 50-class of yore.  Some might argue that the 4090 and 5090 being such monsters skews the comparisons, but we think that’s more of a perception issue based on NVIDIA's success at pushing the cost of the high-end higher. NVIDIA’s flagship GPUs have been very large pieces of silicon since the 20 series, and the CUDA config cutting had barely begun at that point, and the MSRP wasn’t as high as it is now. The price of NVIDIA’s GPUs has generally gone up over time, even accounting for inflation that does things like turn the former $700 1080 Ti into a $912 GPU in today’s money. But then you look at $900 GPUs in today’s money and that’s a 9070 XT. And the 9070 XT isn’t positioned where the 1080 Ti was. The closest GPU might be the 5080 at $1,000 and that also doesn’t feel like a 1080 Ti by price. Flagships, however, are the worst, rising from that level to $2,000 with the 5090. Non-flagships haven’t risen quite as much, but it’s still significant. In this case, line go up = bad. For the consumer, anyway. The 70-series here is one of the most textbook examples of shrinkflation. While the price point has stayed fairly consistent for a few generations, remember that the relative CUDA core configuration has dropped by a huge amount during that time. It’s gone from $610 with a 56% configuration in the 30 series, down to $550 with an embarrassing 28% core configuration in the 50 series. NVIDIA is giving you a half-size slice of the GPU pie with the 5070 than it did with the 3070, but it’s charging you basically the same amount of money for the privilege. All of the GPUs are victims of the configuration cutting we talked about. Even the technically-cheaper-than-they-used-to-be 70 and 60 class cards are providing less of a share of the capabilities of their respective flagships than they used to.  And AMD isn’t immune to this, of course. We have an entire article dedicated to the company’s fake MSRPs that delves into this. NVIDIA, however, holds 90% of the market, and it’s important for you to understand how your money is disproportionately losing value when it’s spent with NVIDIA versus many years ago. We don’t have an answer for this. It’s sort of too big, but it’s important to know about and to start thinking about. Maybe enough people will pay attention to this so that it will help them make informed purchasing decisions.