I own both cards, and I've run both of them in Superposition on exactly the same hardware. I know
experimentally what the performance delta is between the two of them when using less than the 970's 4gb VRAM limit. Theoretically; 1178MHz/773MHz = 1.53, so the GTX 970 should be a minimum of 53% faster, plus some additional benefit from faster memory... but in practice, it does worse than 53%:
(Superposition DirectX results) | M4000 | GTX 970 | Difference |
720p low | 13578 | 18966 | 37.85% |
1080p med | 5115 | 7532 | 47.25% |
1080p extreme | 1415 | 1991 | 40.71% |
4k optimized | 2278 | Over VRAM | N/A |
Using the M4000 as a proxy because I don't have an 8gb GTX 970, it's pretty hard to find situations that would occur naturally where the 970 is VRAM limited but (M4000's FPS x 1.53) would still be a good experience. That had me suspecting they were running some unreasonable settings that you would not normally encounter to get that doubling of benchmark scores... which digging into the sources, is exactly what they did. They cranked the settings to
"8K Optimized", and scored 6.5FPS without the mod, and a still-unplayable 11.8FPS with.
It's interesting to see it built and tested and there's some edge cases where you see a benefit, but it's not a massive swing, and the linked video only seems to have seen a benefit in games that came out 8+ years after the card released and much less than a doubling. Nvidia has definitely made some really bad calls on their newer cards, but the case they intentionally under-VRAM'ed the 970 as part of a nefarious plot is weak.
That's due to the memory bus design, not the chips, so this would have the same issue where the last 1/8th of the VRAM is slow. Doesn't look like they push the VRAM hard enough to encounter it in the videos the article links to.