What do you want in a Switch 2?

[Pemalite]

Bofferbrauer2 said:

Pemalite said:

That 135% more bandwidth and 60% more flops only increased performance by 9% at 1440P.
That is pretty insignificant.

You are still proving my point. That Architecture is more important than flops or bandwidth.

Overclock that 960 heavily and it will match/exceed the 7970 even at 1440P.

Not to mention the 960 was always more of a 1080P card anyway.

As for benchmarks themselves, Anandtech is a more legitimate source.
https://www.anandtech.com/bench/product/1722?vs=1744
.

I get your point, but at some point, having too low a bandwidth will choke the whole chip.

Also, your whole point of the bandwidth being unimportant by pointing to a chip with an oversized connection compared to one with a more sensible size got disproven a bit with the 1660 Super, which pretty much only has better bandwidth over the normal 1660 and et soundly beats the older version by about 15% due to it.

You are missing the point. The 7970 is offering diminishing returns despite a substantial increase in computational throughput and memory bandwidth.

And in some benchmarks the 960 is beating it.
The 7970 is not a good GPU today for 1440P gaming anyway, but it's still pretty good at 1080P.

Here the 960 is besting the 7970 in many benchmarks.
https://www.guru3d.com/articles_pages/msi_geforce_gtx_960_gaming_oc_review,17.html
https://www.guru3d.com/articles_pages/msi_geforce_gtx_960_gaming_oc_review,16.html
https://www.anandtech.com/bench/product/1722?vs=1744

Also need to keep in mind that the 4GB variant of the 960 is faster than the 2GB and the 7970 has a GHZ edition variant, some review sites don't list the differences/variant.

Either way... It's a testament of how superior Maxwell is to first generation Graphics Core Next on all fronts, it can beat Graphics Core Next with less Flops and less Bandwidth.

Bofferbrauer2 said:

I get your point, but at some point, having too low a bandwidth will choke the whole chip.

Also, your whole point of the bandwidth being unimportant by pointing to a chip with an oversized connection compared to one with a more sensible size got disproven a bit with the 1660 Super, which pretty much only has better bandwidth over the normal 1660 and et soundly beats the older version by about 15% due to it.

Actually... Allow me to elaborate.

Bandwidth is unimportant, as in... The raw black and white numbers you see listed everywhere. (A little like Teraflops.)
The real-world bandwidth is important... Just like Teraflops.

The reason why Maxwell can pull ahead of Graphics Core Next is because it employs a myriad of bandwidth saving technologies like delta colour compression which gave it a potential 50% more bandwidth even at identical memory bandwidths, Pascal improved that by another 20-25% on top of it.

Combine that with tiled based rasterization and improved culling... And Maxwell was able to achieve allot more than Graphics Core Next at the same GB/s advertised bandwidth numbers... It's also why the Switch with it's pitiful 25.6GB/s of bandwidth is doing so well.

Obviously the more you increase a GPU's functional units, the more bandwidth you need, hence the GTX 1660 series.