AMD Mantle - John Carmack, Tim Sweeney, & Johan Andersson Open Discussion
Linus Tech Tips
·Linus Tech Tips
·2014-05-07
·
3,064 words · ~15 min read
0:07
hi guys um Jensen's gone right um I wanted to ask I want to ask what you
0:12
guys thought of mantel um just to see since you guys are considered luminaries
0:17
and large developers uh what you guys
0:21
think oh you guys start okay I mean I
0:25
know my opinion on it is I in terms of
0:28
action I question Nidia have a response
0:31
to mandle and I think that's unequivocally know they should you know
0:35
it would be I think a horrible mistake if some reason you know somehow NVIDIA
0:38
got panicked by this and made some lower level NVIDIA API you know you already
0:43
have very good low-level access through the the GPU extensions that Nidia has
0:48
always rolled out know ahead of anywhere else give you as much performance as you
0:52
want to trade the inconvenience of doing Invidia proprietary stuff on them now
0:58
mantle is inde AMD has talked many times
1:01
in the past about close to the metal type architectures and it only ever
1:06
became interesting because of their dual console if it was just a way to to do on
1:11
the PC be a little bit more uh you know lower level for them I couldn't have
1:15
cared less I the landscape does matter
1:19
that they have both the major console WIS with a similar architecture that you
1:22
can get on the PC developers will be making systems architectural changes
1:27
that favor those um so I think it it's not a stupid thing
1:32
for AMD to be doing at this point there's some benefits that they can reap
1:36
from it it could have implications for steam steam box type things there um but
1:42
I they're probably if Microsoft and Sony both embraced it that would be very very
1:47
powerful for AMD but it doesn't look like they're going to at least Microsoft
1:51
from the you know words that they're mouthing so I I mean if I was still
1:57
doing all of the you know all of the B major technology coding for I for nextg
2:02
game stuff I probably would not be embracing mantle right now uh but it
2:08
would be there would be days when it would be extremely tempting but when I
2:11
would dispassionately step back and look at it I probably wouldn't come down on
2:15
the side of saying it's it's worth the effort there obviously others come to
2:19
different conclusions sure uh there's some good
2:23
ideas and the we really liked the idea
2:27
of having low level low overhead access to the GPU I mean if you look back at
2:32
what's in both directex and openl there's a lot of overhead in those apis
2:37
and the multiple layers of them and the fact that they date back to you know the
2:41
old STI model of Hardware rendering which is very very different than the
2:45
modern Shader based programming model that we have now you know where you have
2:50
potentially unified memory and a lot of
2:53
processing power available in the GPU and lots of ways of sharing data that go
2:56
beyond just calling a ton of API functions for everything little thing
3:00
you want to do uh so there are good ideas there I hope that uh really helps
3:05
the openg go committee and Microsoft shape their future apis I me if you ask
3:10
me would I much prefer to have a low-level API for accessing the GPU the
3:14
answer is yes but five of them each for
3:17
different uh hardware architectures and vendor and operating systems absolutely
3:23
not that is the wrong direction for the industry to go um and to have yet
3:29
another AP and the PC that's still different than the standard PC DirectX
3:33
API and it's different than the open GL op es that exists on Mac and Android and
3:38
iOS and it's different than the PlayStation 3 PlayStation 4's lowlevel
3:43
rendering API and it's different than what Microsoft provides in Xbox I I
3:47
don't think it's a a good
3:51
idea yeah
3:56
so no one of the key things about it is
4:00
and one of the main reasons we're doing is that it's not a replacement uh
4:03
there's it's not designed that way at all uh the idea is that we we can solve
4:09
some of the long-term problems that we've actually been having on the PCS
4:12
platform all the stuff we've been talking about today of getting robust
4:16
performance consoles type type of of of
4:19
robust stable performance uh those are things we can experiment and work with
4:23
with M we do that with d3d and we do that with GL also of course uh this is
4:27
another Avenue you can see uh it's also really interesting of just essentially
4:31
opening up something that we're quite familiar with already we've been
4:34
spending well essentially last two years working on the nexan consoles and the
4:38
architectures are the same we intimately familiar with with with those architectures and it's a good fit there
4:42
just as as described and I see even
4:46
though we're not done with our work with it and is definitely not not not done
4:50
yet either to be to be shipped out uh I still see it as a as a success even
4:55
right now actually just because of these conversations and the amount of things
4:59
that actually the amount of developments and the amount of sort of both
5:04
enthusiasm and even the opposite of enthusiasm all the type of um it's been
5:10
bit stale you could say in in especially in the PC graphics Microsoft they switch
5:14
Focus for for quite some time over to to other things which they probably really
5:19
didn't need to do uh but I I think now
5:22
going forward it's a really exciting opportunity both on the on the pieces
5:26
basis and stuff we're experimenting with mental and learning on the Souls but
5:30
also mobile as we discussed before of of okay what should be the future Graphics
5:33
pipelines there there's a lot of movement in all of these things and it's it's just good to have lots of different
5:38
Avenues to experiment uh and actually hopefully deliver concrete value uh in
5:43
things but I completely agree with both John and Tim that if NVIDIA would do
5:47
their own API and then Intel would do an API and then qu would do an API that's
5:53
not a future want to be that that would be a bad future uh that would be very
5:57
very a lot of work you mark moft and Kronos does play an important part of
6:02
all of all this and um I hope M will be
6:05
quite influential in in in in many these aspects also and what we're doing right
6:09
now also specifically with with DD is is just a start um thank you yeah you
6:15
you've used the API where we've just read the spe or the marketing materials
6:20
um what would you say is the the two questions what do you think is the cost
6:24
of like in man years or whatever of implementing manal support in
6:28
Battlefield and what do you think it's the ultimate gain on PC like could you
6:31
try to quantify that it it's it is too early to quantify we've been very busy
6:36
just making sure our game works overall and actually sh sh shipping it out we're
6:40
almost done with it uh so it's too early
6:43
to say we'll have a lot more information in November mid November talk more about
6:48
it and it will be interesting to see I think it's important also to one one
6:51
thing that people well General people that are not developers it can be
6:56
difficult to understand that when it comes to bits it's actually it's not not
6:59
like the ball okay we solve this one we're done go home and the game Run as
7:03
well and ball next is like wack you you
7:06
find fix one and then you thought that that was the thing you made that four
7:09
times faster and then the next B shows up after 10% of that Improvement of
7:14
course the biggest issue is that really you have to architect your large scale
7:18
Vision to take maximum Advantage no API
7:21
just all of a sudden really makes a dramatic difference on things because if
7:25
you buil a good game engine to you know to any API even if you magically made
7:30
all that API overhead vanish it doesn't make that much of a difference it's the
7:33
possibility of do I mean there's stuff on the gcns that I'm very excited about
7:38
some of the asynchronous pipeline cues for different things that would be great
7:41
to you know to take direct direct control over but then that's you know
7:46
you either design for that or you don't design for that it will be a gradual
7:50
approach also of sure we're doing Bel before it's not like Bel before is only
7:55
dedicated for a specific API and specific platform it's it's it's it's a
7:59
rich game there you want to work with something and and deliver concrete
8:03
benefits and evaluate and then go forward and and and see where we are and
8:06
hopefully in the future will design games based on many of the concepts that
8:10
we're proving out a lot earlier now than what we can do otherwise and I think
8:14
that that's actually a good change for the entire industry even though people
8:17
sometimes do get stuck up on on like the the overall thing thinking again oh the
8:22
same discussion as I had before of okay is mobile will replace all the PC no
8:27
that's that will not happen same that will not replace all of the ex or all of
8:30
the Y that they're complimentary well actually when use m is not complimentary
8:34
but it's compliment in the in the actual industry over all uh and you want that
8:39
competitiveness not necessarily by having five different apis but having
8:43
movement uh uh I think that's really really interesting cool I thanks I
8:48
appreciate it lot can you tell us tell us how much they pay for you to use
8:53
it actually that's let's let keep
8:58
having no I can answer a couple questions uh they didn't have to pay us
9:03
anything to use it I've been pitching these type of ideas for for many many
9:06
years for every single vendor uh and it's actually talking just about my my
9:12
own egoistic view of solving my own specific problems is that we see a lot
9:17
of stalls we see a lot of performance performance gone missing in many areas
9:20
that we actually want to get at that performance we want to learn how to
9:24
program a a machine on on a on a lower level that that has actually a really
9:28
really good architecture for for that and I think that's something that will
9:32
will sort of roll out going forward also
9:36
okay so it's good as an R&D platform for developers you me with and John you
9:40
were talking about the desire to have front bu for Access or something like
9:43
that lower level API would be more likely to provide that I think it's also
9:46
good if it pushes Microsoft and open go folks to improve their drivers by
9:51
realizing that hey you can actually achieve a lot more performance on PC
9:55
what the more lowlevel techniques um I think it would only be bad if the
9:59
outcome of this is that now we have to deal with five vendor specific apis for
10:03
AG GP yeah and also the vision for for
10:07
mantle well it can play out many ways you don't really you can't really say
10:11
for for sure uh but the vision for mantle is that it will it can become a
10:16
an API also that's something we're completely open for and something is open for also it's just that that's not
10:21
the right way to go at it initially there are those Aven Ed already with
10:24
cronas and with these this is open API
10:27
it wouldn't be a lowlevel well well actually I don't yeah I don't
10:32
fully agree sure you can make an API that's completely down to the level sort
10:36
of what you like on P we worked it was a
10:40
very very thin abstraction now but it was completely specific with that architecture I do think the way modern
10:45
gpus work is is quite different from
10:48
from how the DX or you absolutely the G
10:52
design was from from the beginning there there are extensions at Le in De Land
10:55
where you can access many things but there there are commonalities between
10:59
like binding for example that's an awesome thing that has been working on
11:03
with with with G that is really the future for all GPS and if you ask the
11:07
different vendors uh even people even vendors perhaps on mobile space that do
11:11
not have it yet they believe that is the future well now I'm talking for all
11:16
mobile MERS not but it makes sense as as a modern design
11:21
for that uh and you see some some convergence towards things that are
11:26
generally the designed and then some people are some Des
11:29
will come a little bit later for that but when it's a great fit for for for
11:33
key hardware and it's great fit for developers then I think well that's the
11:37
way will go like my point about unified memory is absolutely here and we really
11:41
need to get get down to gpus are just peers on the same memory pool as the CPU
11:47
and we really need like documented texture layouts and not noticing that
11:52
everything is coming soc's also so even the GPU is a combination of tons of
11:57
different processors a significant amount of process either if it's an bet
12:00
GPU or if it's an mgpu there's quite a lot of quite independent engines there
12:05
and that's something that's hidden from us now the GPU is something that you
12:08
from the CPU you prepare your commands you just give it to them and it has a
12:12
driver that tries to figure out which goes where that's just generally good
12:16
design to expose those things of living if you're trying to do something picky
12:19
like using the hardware video decompression stuff or then you know
12:22
re-encoding there you have to V to forand new API to get that Capt even
12:28
even even like compute for example compute is well one of the best use
12:32
cases compute is for graphics which is kind of uh yeah feels a little bit weird
12:38
but but it's not like you want to use it both tightly coupled with your with your
12:43
graphics and you want to use it completely separately that CPU drives
12:47
for for loading this and stuff and you want to have background tests that that
12:50
have U well can have really highly to see and run in the background and just
12:54
eat up whatever Victor units that are free on your GPU that's also generally
12:58
good design for even in the GPU space I'm hitting on these issues where I want
13:01
GPU task scheduling and we have no concept of prioritization between GPU
13:06
SKS where know we need to be able to specify time slices and priorities and
13:10
you know it's down being a CPU scheduler for a finite re and well like they take
13:15
quite a lot of time until we find a convergence to a a definite model there
13:19
and but in the meantime it's good to have a couple of models both from
13:23
hardware and from software sort of competing or or or being tested there uh
13:27
and see what sort of what's come s that because I do think there will be quite
13:31
interesting conver there's clearly a large map of
13:36
enthusiasm in the panel for gsync um is it thus fair to say that
13:42
you'll be playing the games that you're
13:47
developing
13:51
wow well if NVIDIA they have gsync they
13:56
have I have to say they have the
13:59
you don't you don't have M at all um it
14:03
starts getting a little bit tricky in my view yeah NVIDIA had mantle
14:09
and we had gsync or we had ideally all
14:12
vendors should have alls that would be the best solution
14:15
that the best solution for the industry but I think what you're seeing if you
14:18
look at it from a practic point of view you'll have first mover first mover uh
14:22
advantages for the for the companies that invest r& the effort and I should
14:26
bring something out uh something out there uh so I'll play the game on Clos so
14:33
forious reasons is the GC software side
14:37
of things and Hardware side on the the card is that going to be licensable to
14:41
like our Intel and afd going to be able to yeah we haven't precluded the notion
14:44
of it frankly you know we're first is trying to get it going get the ecosystem
14:48
going get the display guys going we had to change the GPU architectur get it
14:51
going and then every included any in the future that sounds like a softw
14:57
thinking yeah if only that pesky GC had
15:01
thing I'm just kidding I know Tim
15:05
John yeah well we primarily use invid Hardware at and we we buy L of
15:13
it that's
15:16
surprise I've got a light boost monitor on my desktop so I mean it'll probably
15:20
update to it yeah I heard my voice call
15:25
that uh we got the the voice of God over here telling us one more
15:39
question