This is an update to my original (now deleted) posts, with additional suggestions included.

Image 1:
– It’s very clear that GPU architecture has improved over time, with the newest series offering, on average, better performance for the MSRP (adjusted for inflation).

  • There are diminishing returns in terms of performance, especially at the high end. I believe this is because people who want the absolute best are often willing to pay any price.

Images 2 & 3:
– It seems that actual prices adjust over time based on GPU performance to keep older series competitive.

  • Image 2 is a little hard to read, so I included a log-scale version in Image 3.

Notes:
– All GPUs are compared against the RTX 5090. So, if a GPU shows 50% performance, it means it benchmarks, on average, at half the performance level of the 5090.

  • All benchmark data is from UserBenchmark, cross-checked with other sources where appropriate. I understand concerns exist regarding UserBenchmark’s accuracy, but these are mostly relevant when comparing different manufacturers or CPUs, which is not applicable here.

  • The "current low price on Amazon" reflects what I found in a quick search better deals may be available.

Posted by Large_Cantaloupe8905

22 comments

  4. Add AMD so the comparison is a bit better.
    Performance vs Watts as well

  5. These axes makes a lot more sense than your previous graphs.

    That said, and I know a lot of people have you grief about this, but it was easier to follow each generation when they were connected by line.

  7. Is that performance number with or is it without *multi* frame generation for the 5090? If it is without, then it is more impressive cost:performance than I thought… If it is with, then I personally think the graph needs to make it clear.

  9. It would be very interesting if you could add a chart that compares each card to its series‘ flagship. So 50 series gets compared to 5090 in relative performance, 40 series to 4090, 30 to 3090Ti and so on.

  10. My 4070ti felt like a raw deal when I bought it (it was) but seeing this gen I’m not that mad about it.

  11. Can you please share the raw data? I would really want to play a round with the data! Please!

    Also I think the x axis should be inversed to see the best value ones in one quadrant easily 

  13. I had a 3080 Ti and most recently upgraded to a 4080 Super.

    I upgrade my PC every 5-6 years.

  14. I recently paid below MSRP for a 5080 here in the UK 🙂

    Built the whole system (9800x3d + 5080) for less than 1800 🙂

  15. Huh… And I was happy with my 3050. The better GPUs must be insane

  16. Would’ve been interesting to see current prices as well, maybe an average of used but in mint conditions

  17. What isn’t factored is relative EE design and modern PCI 5.0 SNR requirements vs legacy cards.

    5090 trounces everything because it’s 750mm2 and 170SM on a current 4N node.

    The closest thing to this was the 2080 TI per die size (ignoring yield rate), but NVIDIA purposely used an older TSMC 12 (byproduct of TSMC 16) node for a better price point at the time. Whole 20 series lineup was overly large relative to both 10 (TSMC 16) and 30 series (SS8).

    40/50 follow a more legacy 600-10 series run per GPU tier. Where 80 class falls into 300-400mm2 full die.

    Power requirements are much higher these days. 2080 TI was a 250w design. Now your mid range 5070 has same 250W TDP.

  18. Would have loved to see one that shows frames/$ compared to the performance of the top consumer card at launch and for it to include the 10 series.

    Every card on this graph is a scam.

  19. Would be interesting to add GPU memory as a datapoint since that’s a limiting feature on many of these cards.

  20. The non-FE cards here in Canada is so expensive that it is pretty much extortion. Yeah, 5090s are in stock, but they are the 5090s that are like 1k above MSRP.

  21. The 3060ti was such an amazing card. Matching 2080 performance for less than half the price. The 4060ti was an absolute joke. While the 5060ti isn’t great either, at least it’s cheaper.

  22. You should try normalizing the data based on a known “good reference” card. Take the most popular and/or best price/performance card of the past 10 years and normalized everything around that card.

Comments are closed.