Hello, one is blue and one is green!Can't get more Apples-to-Apples than 2 AMD x86 APUs with Zen 2 CPUs & RDNA 2 GPUs.
Tommy McClain
Hello, one is blue and one is green!Can't get more Apples-to-Apples than 2 AMD x86 APUs with Zen 2 CPUs & RDNA 2 GPUs.
Tommy McClain
Can't get more Apples-to-Apples than 2 AMD x86 APUs with Zen 2 CPUs & RDNA 2 GPUs.
Tommy McClain
Hello, one is blue and one is green!
At a basic high level that's exactly what they are. But that oversimplification glosses over the customisation that both parties have done to those GPUs and CPUs. Using TFlops as a measurement can't take any of that into account just as it can't take into account the differences in IO speed and the impact that will have. Or changes to caches etc. Cerny summed that up in the expo he did.
Just as MIPs are useless for CPUs, Flops are useless for GPUs. It's a number you throw out to people who have a very basic understanding of what they are talking about. The only real way to measure modern GPUs is by how they perform in real world scenarios.
Keep looking at all the minutia & you'll never be able to compare them. Look at the same games running on both systems & you'll just have somebody say you can't look at them because one of them was the lead system & they did the bare minimum on the port. Or one vendors dev tools are behind another vendor. Or one's dev environment is to the metal or the other has too much hardware abstraction. Eventually none of them can be compared. So why the hell are we here discussing all of this shit? LOL
Tommy McClain
In that case we just have to get smarter about how many levels of meta we can deal with before the variance is too great for any realistic measurement to be useful. And measurements like MIPs, Flops are too basic to contain any useful information. Leave those to the n00bs
Would you have that mindset if the PS5 had more FLOPs? Cause I know from last generation it was the one thing got shoved down everyone's throats.
Tommy McClain
A long time ago we estimated that a 2070 approximate performance and feature set is an ideal console within a specific price point.
That’s exactly where a 5700xt sits. Arguably the only features missing is dlss. It could be 8 TF or 10TF as long as it’s performing or putting out 2070 benches that’s all that matters. Honestly I don’t know where PS5 will land. It has the CUs of a 5700 and the clockspeed above the 5700XT anniversary. It gives it a wide range of performance.
That being said I do think there is a complex around any digit that isn’t a 10. Neither 5700 or 5700xt are at peak above 9.75 TF.
If you walk away with 5700xt performance with a super fast SSD and a Zen2 with an amazing price point I don’t understand the argument around whether it’s 8.2 or 10.2. It’s not relevant. You’re getting massive amounts for your money. The only thing that would make this less desirable is only if their competitor manages to ship the same price point in which more people would desire more power but likely not in the face of giving up a game library.
The larger TF differential is going to be put towards resolution and that’s where most of it’s benefit will lie.
Teraflops isn't going to matter, it's going to be bandwidth limited.
Pro has higher fill rate but runs most games at a lower resolution, sometimes while even using optimizations like checkerboarding (making the actual rendered resolution/bandwidth requirements even lower). 3d rendering is like a chain, only as strong (fast) as it's weakest link. I would imagine that Pro is more often than not bandwidth constrained, given is fillrate advantage, while at equal settings 1x would be bottle necked by something else. Both are held back by the CPU, though.I'm wondering about that as well.
But the PS4Pro manages extremely well with only 218 GB/sec and compares very favorably against the X1X with it's 336 GB/sec BW. Originally, I thought the Pro was going to be really bandwidth started and it isn't. Maybe dev's have to work around that, I don't know, but it doesn't seem to be a big issue.
Hmm haven’t heard such a rumour before. I suspect if true this would only be for the graphics pipelineFitting rumor with the subject: PS5 CUs and RDNA3 CUs have independent clock domains depending on its activity.
On the other hand if PS5 geometry engine capabilities bring really deferred vertex shading the tflops number lose greatly its value because of huge efficiency gains in the pipeline. Would like to read 3dilettante take on this...
Hmm haven’t heard such a rumour before. I suspect if true this would only be for the graphics pipeline
Min 1,15
Hmm. I’ll wait for more (other sources) or the RDNA3 presentation. Unfortunately I haven’t found his information to ever be correct.
Min 1,15