Never looked at it that way before. Quite telling of a story.
The Great PS4 Missing AF Mystery *spawn
- Thread starter Shortbread
- Start date
its even closer if you factor the GPU up clock on Xbox, dangerous territory these assumptions but we can adjust for CU performance a little if we assume PS4 is default at 800mhz and we know Xbox was lifted to 853mhz the numbers are
(1920 * 1080) / 18 = 115200
(1600 * 900) / 12 = 120000 / 853 * 800 = 112544
(1920 * 1080) / 18 = 115200
(1600 * 900) / 12 = 120000 / 853 * 800 = 112544
Encyclopedia Brown is closed for business.
You could create a synthetic homemade test to test AF bandwidth demand across a library of games by downclocking vram until you start getting performance degradations with AF on at various levels and again with AF off, find the difference. Match settings as close to console versions as possible. It could be done very comprehensively if one wanted to spend a dozen hours testing variables. Preferably conducted on AMD 7800 hardware. Then you'd start to get a general idea of the cost of AF.
Last edited:
I like this approach. Take some time to do but at least it's real data to work with
Digital Foundry vs console texture filtering
It's about time! This needs to blow up, whoever is responsible for this mess needs to sort that out.
It's about time! This needs to blow up, whoever is responsible for this mess needs to sort that out.
D
Deleted member 11852
Guest
And their conclusion: we're not sure.
Epic fail.
Epic fail.
Well that was anti-climactic.
We already came to the same conclusion in our own discussion!
hesido
Regular
What they did with emulating the situation is interesting. Of course, even the bandwidth starved APU showed very little performance loss with AF enabled. That said, it could be CPU limited due to DX11 and the AF could be coming for "free". Too bad they didn't test how much CPU bound their test were by upping the resolution.
Is that a random blog article from a random blogger about the subject spreading FUD and misinformation?
My thoughts as well. None of the games were running low level drivers so it's possible that the overhead allowed for latency hiding of free AF. The real question is when DX12 games are released and what the impact will be then.
Not to mention none of the APU benchmarks actually hit locked frame numbers, something console games must target much more vigorously.
Pure speculation, but it would be funny if the drivers for DX11 PC had _always_ incorporated the cost of AF somehow. And not enabling or enabling would make virtually no difference.
D
Deleted member 11852
Guest
Drivers simulate consumed bandwidth for disabled features? Ok..
Instead of DF looking at it purely from a hardware perspective (issue)... they should have spent a little bit more (MORE) time on seeing if certain developers have better (updated) SDK/toolchains when compared to others. Or simply see, if certain features are broken (or become broken) when working (compiling / porting code) between multiple platforms. Not everything is hardware related DF....
forumaccount
Newcomer
Really they should just take a development team that made a choice they didn't understand and ask them directly. If they don't respond then you don't really have a conclusion. Keep in mind that DF has a really bad track record of correctly identifying things like level of AF and AA methodology in shipped titles*. Their articles would just be a lot better in general without this kind of back seat guessing. They're definitely at their best when they are sitting down with developers and asking mostly intelligent questions.
*( Just going off my own games that they've reviewed, at least. This may not be a common opinion. )
too bad they didn't test with slower DDR3 memory. If it is a memory-bandwidth issue, and I use an APU that is even slower than the xbox one gpu, I would prefer to also use lower-clocked memory. DDR3 1600 or DDR3 1333 might show different behaviors.
E.g. APU tests from the past confirm that the jump from 1600 -> 2133 is a big one for the APU. But 2133 -> 2400 isn't that big. So with 2400 the APU might have more than enough bandwidth for most situations.
I still think it is the memory-bandwidth so. If you don't get your memory access patterns right on one shared memory-pool you loose much bandwidth. The loss is even higher on gddr5 memory (because of the bursts). It might not be much, but if the developer has the chance to just deactivate a single feature and everything is back on track, that the developer will most likely choose this way. Later, when they might have more time (after launch) they could tweak their engine here and there and save some bandwidth with better access patterns and turn the feature back on with a patch.
