2x rate FP16.
- Status
2x rate FP16.
Don't they end up with lower power consumption due to reduced bandwidth/storage, assuming it's straight up replacing current FP32 workload ?
(but in any case, not everything is being switched to FP16)
But that's not 8.4TF, really. Is it?
upnorthsox
Veteran
Speaking of the 8.4TF, has there been any speculation on whether Scorpio's 6TF is actually at 16FP like the 8.4TF for the ps4pro?
No way this is going to be the case. Most surely 6TF at FP32 and perhaps 12TF at FP16 if they decide to include that feature (not sure if it's a vega thing or not).
upnorthsox
Veteran
Why is this? If they could've claimed 12TF don't you think they'd have done it by now?
This comes down to the definition of a FLOP if we get technical on the wording. Floating Point Operation. There are two
16 and 32. We take 32 as the default and thus we don't append @32 to it. So the term 8.4TF@16FP is valid IMO.
But Sony have not claimed 8.4TF as well because it's only really useful for some cases.
Battlefield 1 will be interesting, more consistent 60 in multiplayer will definitely be a noticeable improvement from the current PS4 version. Sad to see no mention of Bloodborne, The Order, Until Dawn and others. You'd think Sony would have pushed for all major second/first party titles to be updated but i guess that's not the case.
Yay! Knack!
Actually, this is great, most of the games I own will get patched. The only important missing game is Witcher 3
Lots of VR games there, too...
loekf
Regular
Yay! Knack!
Actually, this is great, most of the games I own will get patched. The only important missing game is Witcher 3
Knack ? I suspect Marc Cerny did the Pro upgrade during his evening hours.
Is there really anybody out there who plays Knack ... still ? Cerny should make a VR version. Oh wait... Playroom VR ?
If fully exercising 2x FP16, it would mean fully exercising the packed registers, and loading 2 sets of 16-bit data from memory rather than 1 32-bit set in the same time frame doesn't change the naively calculated bit count. The memory question might depend heavily on representation and use case.
Some operations are unlikely to significantly change in power consumption (2 half-sized additions might yield similar power consumption + possible overhead). Multiplication could be cheaper, but it's been noted elsewhere that it's the data movement that is the larger contributor and 2x16 is the same amount of data as 1x32. It's also not clear exactly how this is being implemented and what limits there are to the operations. Tweaks that might change how flexible the support is could cost power.
On the other hand, the other benefit is reducing occupancy--which means those CUs spend less time stalled. This can mean there might be scenarios where this could depend on how effective the GPU's clock gating is in saving power in stall situations.
Cerny did by simply giving the headline rate. It's not too different from quoting the 32-bit rate, since it's also not realistic to expect that to be fully exercised either. It's just a hardware parameter.
http://www.gamasutra.com/view/news/283611/Inside_the_PlayStation_4_Pro_with_Mark_Cerny.php
Can't wait to play 16-bit games at 1000fps
And get accused of misleading people, and only releasing a console half as fast as they claimed?
That would have been a terrible mistake for MS to make.
There are dozens of us! Well I still do speed runs on hard mode, every 6 months or so.
I'm trying to imagine Mark Cerny telling developers with a straight face it's super easy to patch for Pro, and then he wouldn't patch Knack. For that alone, it had to be done regardless of the user base. Still, there's a lot of people buying a PS4 for the fist time, that's probably a good reason to patch old games, it gives some exposure to forgotten games on PSN.
I remember some devs said the games that are problematic to patch are open world. I guess having to go back, retest everything, performance tuning, makes it much more expense there than with simple linear games.
I felt I should expand further on a point I forgot to get back to concerning straight-up replacing the FP32 workload.
I do agree that this can be a gain, if the FP32 load of N operations is converted to an FP16 variant that processes in half the time and then immediately stops.
I would be curious where things are on the balance, since going into Pro mode generally means increasing the problem size and tweaks to increase utilization.
Devs frequently make optimizations in what is stored in memory to conserve bandwidth and increase hit rates, so that component may not change as much. Delta color compression, however, is new addition that reduces memory's power contribution--although this is with faster memory and a larger workload.
It might be interesting to see how much more efficient the Pro is when running a non-Pro game.
Writing a game with all of the Pro's efficiency features, but then explicitly keeping the workload at the same level of the PS4 version would be an interesting comparison as well. However, I don't know what good that would do to put the extra work in to create a Pro code path that refuses to differentiate itself from the standard PS4 path.
Agreed, MS needs to tread lightly here. People still quote and meme the shit out of their XBO launch.
All in all, they still need to focus on having games and building their library. That's a bigger priority in terms of marketing. It's also easier to communicate than 12TF @ 16FP.
edit: last but not least, occam's razor, they are still too far away to begin communicating on the technical details of their system.
- Status
Similar threads
