Nvidia RTX 4070 Ti with memory mod easily beats RTX 4080 in Superposition benchmark

Nvidia RTX 4070 Ti with faster memory mod
(Image credit: TecLab on YouTube)

The benefit of using faster video memory on Nvidia GeForce RTX 4070 Ti Super graphics cards has been demonstrated by Brazilian YouTubers. Hardware tinkerer Paulo Gomes and the overclocking team at TecLab both memory-modded RTX 4070 Ti Super graphics cards and overclocked the GPU (using the stock cooler) to boost performance. Using the Unigine Superposition 8K benchmark to check their handiwork, it is gratifying to see that they could both achieve significantly better 3D benchmark scores than the higher-tier RTX 4080.

Tech site VideoCardz unearthed the two new Portuguese language videos, and noted that Paulo Gomes works with Manli, while TecLab is sponsored by Galax. As rivals, of sorts, this led to Gomes and TecLab competing over who could push the RTX 4070 Ti Super the furthest with the help of a memory mod.

These memory mods required more than ‘simply’ de-soldering and re-soldering faster VRAM chips on the respective graphics card PCBs. Work also had to be done to implement support for the new faster memory chips. Thankfully, the rival teams could rely on the support and expertise of their respective sponsors to get the modded cards up and running without firmware or software issues.

For their competitive testing, the teams agreed to use the stock coolers that their Manli and Galax graphics cards came with. Both firms equip quite beefy triple fan cooling shrouds, so no one had an obvious advantage here. The test comparison platform chosen was Superposition. This attractive benchmark with screen-space ray-traced global illumination and dynamic lighting uses the Unigine 2 engine and targets “extreme hardware stability testing.” Superposition also conveniently includes GPU temperature and clock monitoring.

Swipe to scroll horizontally

Card model

Superposition stock settingsSuperposition with memory mod and OC

RTX 4080 reference

8,525 (22.4 Gbps)

NA

Manli RTX 4070 Ti Super / Gomes

7,212 (21 Gbps)

8,870 (24 Gbps)

Galax RTX 4070 Ti Super / TecLab

7,028 (21 Gbps)

9,133 (26 Gbps)

Above you can see the collated competition results, showing that both Gomes and TecLab managed to achieve better Superposition scores than the higher-tier RTX 4080 (stock). If we focus on the TecLab results, we note that its memory mod and OC precipitated 30% better performance in Superposition. The result means that the modded Galax also performed over 7% faster than the stock RTX 4080 in the benchmark.

The latest ‘Blackwell’ RTX 50 series rumors suggest we could be getting a new GeForce line with GDDR7 memory delivering up to 28 Gbps at launch, with the memory maker roadmaps targeting up to 40 Gbps. Don’t expect the first RTX 50 cards until the last quarter of 2024 when the top RTX 5080 and 5090 models are likely to launch first.

Mark Tyson
Freelance News Writer

Mark Tyson is a Freelance News Writer at Tom's Hardware US. He enjoys covering the full breadth of PC tech; from business and semiconductor design to products approaching the edge of reason.

  • Alvar "Miles" Udell
    nVidia: The way it's meant to be played.*

    *In a way totally controlled by us to enforce strict performance segmentation and to allow us to show greater generational gains than would be possible if we did not intentionally handicap our products.
    Reply
  • Roland Of Gilead
    Alvar Miles Udell said:
    nVidia: The way it's meant to be played.*
    Or..

    nVidia: They way we are going to play you (suckers)
    Reply
  • kiniku
    Love my 4070 Ti Super.
    Reply
  • 35below0
    Pointless to argue when the GPU is not available and has to be crafted manually. Too few have the skill. And there's no warranty.

    Almost like this is a bs article designed to flame nvidia. Not that they don't deserve to be knocked, but they're running a business. They're extracting maximum financial gain from their designs vs overheads, logistics costs, R&D, etc. etc. not max performance gain.
    Extracting maximum performance in a way they cannot market or sell is cool for modders and GPU enthusiasts, but not a good idea for nvidia at all.
    Reply
  • CelicaGT
    35below0 said:
    Pointless to argue when the GPU is not available and has to be crafted manually. Too few have the skill. And there's no warranty.

    Almost like this is a bs article designed to flame nvidia. Not that they don't deserve to be knocked, but they're running a business. They're extracting maximum financial gain from their designs vs overheads, logistics costs, R&D, etc. etc. not max performance gain.
    Extracting maximum performance in a way they cannot market or sell is cool for modders and GPU enthusiasts, but not a good idea for nvidia at all.
    It's just proving out what everyone suspected, as a 4070 Ti owner I thought it was pretty cool. A more extensive benchmark suite would have been more interesting, something more representative of the average users gaming catalogue.
    Reply
  • bit_user
    The article said:
    Using the Unigine Superposition 8K benchmark to check their handiwork ...
    Ha! They had to push it well into territory where it's bandwidth-starved for memory bandwidth to be a real limiting factor.

    I'm not invested in the matter enough to investigate further, but if anyone went to the source, please let us know if they also showed performance at other resolutions.
    Reply
  • CelicaGT
    bit_user said:
    Ha! They had to push it well into territory where it's bandwidth-starved for memory bandwidth to be a real limiting factor.

    I'm not invested in the matter enough to investigate further, but if anyone went to the source, please let us know if they also showed performance at other resolutions.
    I'm not overly fluffed to dig into it further either. All I can say is that at the cards intended resolution (1440p) VRAM overclocks do virtually nothing. GN(?) and others did a dive into this and found VRAM bandwidth only started to be a limiting factor at 4K and above. It's one of the reasons I caved and bought mine (It was on sale too.) as my monitors are both 1440p.
    Reply
  • hotaru251
    35below0 said:
    And there's no warranty.
    i'd argue 98% of GPU owners never use it once in their life.

    35below0 said:
    in a way they cannot market or sell
    they could easily sell the gpu modded for those who lack the skill.
    Reply
  • bit_user
    hotaru251 said:
    i'd argue 98% of GPU owners never use it once in their life.
    I never buy aftermarket warranties and the main benefit I believe I get from manufacturer warranty is protection against failures in the first 48 h of usage (see also: the infamous "bathtub curve").

    So, what I'd do is buy a GPU, stress test it for a couple days, then mod it and just accept the risk the mod goes wrong or there's some long-term failure. Heck, I just modded a motherboard in a way that's guaranteed to void the warranty, so...

    hotaru251 said:
    they could easily sell the gpu modded for those who lack the skill.
    Yeah, kinda like how you could buy pre-delidded CPUs.
    Reply
  • 35below0
    hotaru251 said:
    i'd argue 98% of GPU owners never use it once in their life.
    98% of GPUs never RMA'd you say? I don't think so.

    Warranty is useful insurance to have. Doesn't matter whether it's used or not.
    bit_user said:
    I never buy aftermarket warranties and the main benefit I believe I get from manufacturer warranty is protection against failures in the first 48 h of usage (see also: the infamous "bathtub curve").
    That's largely what it's for. Even with a high quality manufacturing process, some products will be duds. Those typically fail very early on.

    hotaru251 said:
    they could easily sell the gpu modded for those who lack the skill.
    This is completely besides the point.

    What you are asking now is that nvidia introduce another GPU in between the 4080 Super and 4090. They would have already done so if it made sense financially. It doesn't.

    Just because it's possible to prove a concept like this modded 4070 can work, doesn't prove it can be made to pay.
    Reply