AMD Mantle API [updating]

[overclocked_enthusiasm]

I am still slightly confused. The original AMD "45% faster with Mantle" slide specifically stated the hardware was Kaveri 7850k and 290x. Does this imply the more powerful the GPU the more load Mantle can offload from the CPU to the GPU and thus improve overall performance? Or was this done simply because the initial Mantle patch is optimized for GCN 1.1 cards?

I guess the question is where is the sweet spot once full optimization is extended to GCN 1.0 cards? At what point will diminishing returns kick in and limit the price/performance reward? Here is what I see for AMD as positives:

1. A typical upgrade path for gamers is CPU, GPU or both. Because of this, gamers may choose to allocate all their upgrade $$ to just the GPU since the CPU may be "good enough" with Mantle.

2. New builds and/or upgrades may see AMD CPUs as more compelling than Intel based upon the price/performance with Mantle games.

3. This may provide a halo effect aroung AMD products IF other game devs view the increased buzz and benchmarking around this api as beneficial to their brand..."I want some too". This will help entrench this api which has proven historically VERY difficult to do.

Mantle means NOTHING unless it results in incrementally higher hardware sales for AMD and this ONLY happens if game devs sign on and the Catalyst team is prepared and able to handle their responsibility to provide world class drivers...including Xfire, Dual Graphics, and frame pacing.

 

[UniversalTruth]

Does this imply the more powerful the GPU the more load Mantle can offload from the CPU to the GPU and thus improve overall performance?

My guess is a no...

Guys, why don't you chill out a little bit and patiently WAIT till AMD get their act together... Haven't you noticed that with their resources they need TIME?

 

[3dilettante]

I am still slightly confused. The original AMD "45% faster with Mantle" slide specifically stated the hardware was Kaveri 7850k and 290x. Does this imply the more powerful the GPU the more load Mantle can offload from the CPU to the GPU and thus improve overall performance?

It means the weaker CPU doesn't choke on the driver as much, which allows the GPU to not sit around waiting for the CPU to stop choking on the driver. Nothing is being offloaded, the CPU just isn't getting in the way. Better CPUs don't get in the way as much.

The small GPU-limited gains and larger CPU-limited gains show what Mantle currently brings.
Mantle's low-level access and exposure of certain hardware behaviors can lead GPU-side benefits, but those were not the most expounded benefit.

Mantle means NOTHING unless it results in incrementally higher hardware sales for AMD and this ONLY happens if game devs sign on and the Catalyst team is prepared and able to handle their responsibility to provide world class drivers...including Xfire, Dual Graphics, and frame pacing.

Mantle is all about making the devs do what the drivers used to do.
The exposure of individual cards in a crossfire configuration as separate queue devices allows for the game to manage its multi-GPU workload.

 

[overclocked_enthusiasm]

My guess is a no...

Guys, why don't you chill out a little bit a patiently WAIT till AMD get their act together... Haven't you noticed that with their resources they need TIME?

The flip side of that is that product launches for a company with limited resources are doubly important to get right since they don't have the PR muscle to change or redirect the narrative once it is initially muddled or incorrectly established. Look at the confusion with Kaveri and now Mantle. I would argue patience isn't for the consumer to give but conversely it is the company's responsibility to provide clarity.

Look at it this way:

1. Kaveri launched (1/14)...confusion, no Dual Graphics, no drivers
2. Mantle patch (1/30)...confusion, which hardware best, GPU? CPU? no drivers
3. Cat driver 14.1 (2/3??)...heightened expectations to solve all the above with a beta driver

Now what if THIS happened

1. Quietly launch Kaveri into the OEM channel
2. Wait for Kaveri retail launch with:
~ Updated drivers
~ Mantle patch (with link on AMD.com with demo)
~ Update Dual Graphics page on AMD.com with appropriate 250/260 recommendations (with demo)
~ Identify AMD CPU/GPU ideal setups for reviewers to bench so there is CONSISTENCY for the average retail buyer to act upon for an upgrade or new build (with demo)

I would argue that waiting 2-3 weeks to cover all your bases with 2 MAJOR initiatives (Kaveri & Mantle) is far superior to the piece meal launch we have seen over those same 2-3 weeks that sows confusion and inconsistency. I WANT AMD to do well...they need help on the execution/PR side of their launches and it wouldn't cost them much to fix it.

 

[3dilettante]

A few weeks may be pretty optimistic for that swath of milestones.
I think more critical to AMD is that Kaveri and Mantle were already behind schedule. Kaveri is brutally so.

AMD doesn't cater to gamer confusion. Odds are it does better the less they are able to evaluate it in relation to the rest of the market.

It has made a lot of promises to shareholders and financial backers, as well as contractual obligations and schedules it has committed to.
It's the same thing as to why AMD didn't delay the release of Bulldozer and other less than winning products. It's not free to get the ball rolling and have other commit to its plans and then reneg on them without consequence, either in more immediate penalties or the blowback from yet another failure to provide accurate guidance and additional loss of good will amongst those it needs to work with.

 

[overclocked_enthusiasm]

A few weeks may be pretty optimistic for that swath of milestones.
I think more critical to AMD is that Kaveri and Mantle were already behind schedule. Kaveri is brutally so.

AMD doesn't cater to gamer confusion. Odds are it does better the less they are able to evaluate it in relation to the rest of the market.

I guess the few weeks I was referring to is the Jan 14 - now time frame we have seen for Kaveri support...they obviously had months before that to work on the issues I mentioned. I don't want to bog down the thread so I will just say in closing that while AMD may not "cater to gamer confusion", the simple fact that their products may be underwhelming is even more reason to be in control of the narrative...and drive home the most complimentary hardware/software scenarios that the retail consumer can emulate.

 

[3dilettante]

Odds are AMD made commitments to the dates it chose, and odds are those were already "this is the last delay, I swear" kinds of commitments.

Confused gamers aren't materially hurt by their uncertainty, and can't hurt AMD even if they were.
Actual partners and investors? They can be hurt, and they can hurt AMD back.

It does not seem that AMD is in a position that it can control the narrative.

 

[DieH@rd]

Star Swarm Benchmark is out on Steam
http://store.steampowered.com/app/267130/

 

[repi]

Well they say this:

EDIT - They just removed the last screenshot (which by sheer coincidence (?) was the only one showing very different FPS than what the averages show) after someone commented it was from Single Player and not Multi Player. Really??? How professional...

The FPS figures that show in the top right corner of my screenshots are from 0.5 seconds and measured differently and not as accurate. Those are not the timings I'm using, I'm using the timings based on our perf overlay graph in the bottom left (the CPU avg number) which is a much longer average CPU frame time in these areas that I selected that have stable load.

It is not an absolute perfect and extensive benchmark, but it good test case to compare performance of the different rendering backends. Just wait a day or two and you'll see plenty of PC hardware review sites have their numbers out on a much larger set of hardware and scenes

 

[repi]

By looking only at the two frame rate counters, you could say "WTF? Mantle only gives a 4fps boost!" - but then look at the graph on frame times in the lower left.

Never ever compare frame rate in absolute 'fps' numbers, it has a non-linear relationship. Use milliseconds / frame instead for comparisons or percentages. This is your first warning

 

[Davros]

The steam page says
"Download AMD's Mantle driver here when it becomes available"
and the link takes you to Catalyst 13.11 Beta 9.5
page was posted today 30th jan, why does it say when available when it is available ?

 

[Kaotik]

The steam page says
"Download AMD's Mantle driver here when it becomes available"
and the link takes you to Catalyst 13.11 Beta 9.5
page was posted today 30th jan, why does it say when available when it is available ?

Because 13.11 Beta 9.5 isn't Mantle driver and was released last year? The AMD driver page will be updated when the Mantle driver (14.1 Beta) comes available

 

[Andrew Lauritzen]

Though you should ignore sites that test multiplayer on empty servers.

Yeah testing multiplayer is obviously an issue. Have you guys considered releasing any sort of benchmark like the ones you internally used to help the press there? I'm assuming the numbers you showed were in full screen exclusive mode - are there similar gains in borderless mode too? Any gotchas there?

Great work for getting this out the door! Looking forward to trying it out myself (once there are drivers for my 7970? hrm), and glad to see some general engine improvements in there too!

I'm curious though, given the focus on AMD CPU improvements in the more "common" config, are you guys willing to hint at what percentage of your users use AMD CPUs vs Intel? Are they similar to the Steam numbers for instance or skewed one way or another?

Never ever compare frame rate in absolute 'fps' numbers, it has a non-linear relationship. Use milliseconds / frame instead for comparisons or percentages. This is your first warning

Indeed! Someone mistakenly gave me the keys to this place so I think I'm gonna make this a ban-worthy offense!

 

[overclocked_enthusiasm]

http://battlelog.battlefield.com/bf4/news/view/2955063418790832031/36/

38 pages long and counting...this community is BIG and your key demo for dGPU.

https://twitter.com/AMDGaming/status/428958040046702592

A 1000 paper cuts...

 

[Dave Glue]

ok ok, nothing wrong with that, but sounds like "if you can not beat Intel hw side, do it with the sw".

We'll see with Core i3 benchmarks and Mantle if the bottleneck Mantle relieves is primarily CPU "overall", or just CPU's with 4+ cores and poorer single threaded performance.

 

[Malo]

Star Swarm Benchmark is out on Steam
http://store.steampowered.com/app/267130/

Stardock account required? Yeah not sure I can be bothered.

 

[repi]

Yeah testing multiplayer is obviously an issue. Have you guys considered releasing any sort of benchmark like the ones you internally used to help the press there? I'm assuming the numbers you showed were in full screen exclusive mode - are there similar gains in borderless mode too? Any gotchas there?

Great work for getting this out the door! Looking forward to trying it out myself (once there are drivers for my 7970? hrm), and glad to see some general engine improvements in there too!

Would love to be able to release a good benchmark externally, but the one we use internally for multiplayer is not something we can release as our "pseudo soldiers" run on the dedicated server that one sets up on a local machine and then connect to with the machine one wants to play and measure on.

There are good gains in windowed / borderless as well, but haven't measured that extensively, main focus for now have been fullscreen with all types of configs (multi-GPU is quite a bit of work to handle explicitly & efficiently in all cases). But windowed is really important for us, both during development and when playing the game.

I'm curious though, given the focus on AMD CPU improvements in the more "common" config, are you guys willing to hint at what percentage of your users use AMD CPUs vs Intel? Are they similar to the Steam numbers for instance or skewed one way or another?

For BF4 the amount of AMD CPU users are similar to the steam numbers but slightly lower, which I think makes sense.

 

[DavidGraham]

Just tried out the Swarm benchmark. It was choppy as hell, but curiously only used 20% of my CPU (spread out on 3 cores) and no more than 50% of my GPU .. mostly ran closer to 40%. I am guessing it needs more work to optimally use system resources .

The FPS figures that show in the top right corner of my screenshots are from 0.5 seconds and measured differently and not as accurate.

That confirms what I thought earlier. Thanks.

Those are not the timings I'm using, I'm using the timings based on our perf overlay graph in the bottom left (the CPU avg number)

So you are measuring CPU frame time, not actual frame rates?

 

[Andrew Lauritzen]

There are good gains in windowed / borderless as well, but haven't measured that extensively, main focus for now have been fullscreen with all types of configs (multi-GPU is quite a bit of work to handle explicitly & efficiently in all cases). But windowed is really important for us, both during development and when playing the game.

Makes sense, was just curious as I typically prefer borderless where possible, especially in cases like BF4 that support scaling the 3D content but UI at full res - which works great btw Fully understand full screen being the main focus though and TBH I'm sort of surprised you did the work to support multi-GPU configs at all. I'm still not sure how much I care about those setups, but at least you have better access to do something slightly less stupid than driver AFR with Mantle Looking forward to seeing some analysis of that in any case.

For BF4 the amount of AMD CPU users are similar to the steam numbers but slightly lower, which I think makes sense.

Yep, that was my expectation as well but thanks for confirming.

Just tried out the Swarm benchmark. It was choppy as hell, but curiously only used 20% of my CPU (spread out on 3 cores) and no more than 50% of my GPU .. mostly ran closer to 40%. I am guessing it needs more work to optimally use system resources .

Or that's entirely the problem Mantle is trying to solve... single-threaded graphics API bottlenecks.

So you are measuring CPU frame time, not actual frame rates?

What definition of "actual frame rate" makes more sense than the CPU timings that the simulation uses?

 

[Dave Glue]

Never ever compare frame rate in absolute 'fps' numbers, it has a non-linear relationship. Use milliseconds / frame instead for comparisons or percentages. This is your first warning

Uh, yes - that's my point, hence the screenshots showing the frametime graphs right under it. I'm fully aware Beyond3D members understand this - you come to have a feeling for mindset of a forum when you've been a member for 11 years.

Yes, frametime is getting more and more popular as a standard measurement thankfully, but more mainstream (or lazy) sites still just run a benchmark and report average FPS. At least on more mainstream configs, Mantle's strength (at least this stage) seems to be the consistency of delivered frames.

My slight fear is that this may do Mantle a disservice if only the average FPS difference is reported in comparisons to DX, so hopefully AMD will do what they can to educate review sites on measuring frametime over just the seemingly typical "Put all sliders to the right, run benchmark.exe, slap average FPS number in chart".