[GameGPU] Q1 2015 GPU Market share

[RussianSensation]

Source

Fermi and Kepler owners must be upgrading to Maxwell in droves and HD5000-7000 series owners are tired of waiting! With 980Ti out, many HD5870/6970/7970/R9 290X owners will finally give in!

AMD is in serious trouble now. Their market share is reaching extremely dangerous levels whereupon NV's dominance of the market gives them even more leverage to control prices (which is bad for us gamers) and influence game development with GameWorks. All of that will make it that much harder for AMD to recover which could mean we are slowly moving to a monopoly position, similar to Intel vs. AMD in the CPU space.

With all the focus on Fiji Fury, I am starting to wonder if AMD has really lost focus of how important the mobile dGPU market segment is in the overall GPU industry. They keep putting all their eggs into the desktop basket which has shaped their GPU development into the wrong direction. With NV's bottom-up approach, their smallest units that make up the GPUs are very power efficient and thus NV can scale their graphics to various market segments. AMD's top-to-bottom approach to GPU building seems outdated in 2015.

NV has announced mobile G-Sync and it shouldn't be too long before 965M/970M/980M get refreshed with higher clocked parts come 2H of 2015. Without AMD having a strong mobile dGPU lineup for R9 300 series, they are soon going to drop < 20% market share if they have nothing to compete in the mobile dGPU segment until Pascal in 2H of 2016. I am one of the strong proponents that AMD should have focused a lot more on the mobile dGPUs, a faster growing market segment than desktop cards. Even if Fiji is competitive, it will do little to stop the bleeding considering NV's mobile dGPU stack is basically uncontested and besides Apple throwing a bone to AMD, no one is really interested in using AMD's mobile GPU chips in their laptops.

 

[NTMBK]

AMD wants to replace mobile dGPUs with APUs. When HBM2 rolls around, APUs with a pair of HBM stacks will be great for laptops.

 

[ShintaiDK]

AMD needed a new uarch and they got none because they ran out of money. And spending what they got on hopeless projects. Lacking develeoper relations is another matter.

And Fiji is turning into a farce.

 

[Silverforce11]

And Fiji is turning into a farce.

Doom and glooming.. how expected from you.

So, you care to share with us the benchmarks for Fiji?

 

[dacostafilipe]

AMD needed a new uarch and they got none because they ran out of money.

GCN is still a great architecture, there's no reason for creating a new one.

What they need is improved power management, like what they have in Carrizo (only slides at the moment, we still need benchmarks to confirm if it works!)

Can I just remind you that Maxwell is a GCN-like Kepler evolution?

 

[ShintaiDK]

Doom and glooming.. how expected from you.

So, you care to share with us the benchmarks for Fiji?

Whats next from you. Marketshare is a lie? 35% to 22.5% in a year. What do you call that?

Its not hard to estimate where Fiji will end up. But watercooling and 4GB sets its own limits.

GCN is still a great architecture, there's no reason for creating a new one.

What they need is improved power management, like what they have in Carrizo (only slides at the moment, we still need benchmarks to confirm if it works!)

GCN is outdated and needs an overhaul. And its quite clear from the numbers.

Can I just remind you that Maxwell is a GCN-like Kepler evolution?

You can start by proving it first.

 

[Silverforce11]

Please estimate away on Fiji performance, I'd like to see your logic at work.

Show me the gloom & doom.

As for GCN being a great uarch... proof:

http://www.hardwareluxx.de/index.ph...ette-directx-12-featurelevel-beherrschen.html

 

[Erenhardt]

Aha... #1 metric driving GPU purchasing decision is desktop GPU marketshare.
This shift in recent nvidia marketing tells me AMD fury is going to stomp everything currently on the market.

 

[Pandora's Box]

Doom and glooming.. how expected from you.

So, you care to share with us the benchmarks for Fiji?

It will get 5,000 fps.



...in Quake 3.

 

[DDH]

Its not hard to estimate where Fiji will end up. But watercooling and 4GB sets its own limits.

Where you would like it to end up perhaps. But any estimate you make will be utterly rubbish as you do not have any factual information

GCN is outdated and needs an overhaul. And its quite clear from the numbers.

Outdated why, because it has existed for X years? A ridiculous assertion.



RS, I have enjoyed reading your thought provoking posts these last few months, but this one seems overly pessimistic. Regardless, I have high hopes for AMD, DX12 and the future of PC gaming.

This launch of 980ti seems abrupt. Was there an official announcement? I saw a few leaks and then the flood gates seemed to open as if the NDA was released pre-emptively.

 

[ShintaiDK]

Please estimate away on Fiji performance, I'd like to see your logic at work.

Show me the gloom & doom.

As for GCN being a great uarch... proof:

http://www.hardwareluxx.de/index.ph...ette-directx-12-featurelevel-beherrschen.html

That chart is wrong. GCN doesnt support conservative rasterization for example.

4096SPs at 1 or 1.05Ghz doesnt leave much to the imagination. Its not a new uarch after all. And selling 4GB flagship cards in mid 2015...

 

[Erenhardt]

That chart is wrong. GCN doesnt support conservative rasterization for example.

It does

4096SPs at 1 or 1.05Ghz doesnt leave much to the imagination. Its not a new uarch after all. And selling 4GB flagship cards in mid 2015...

You seem to know a lot about Fiji, can you share some benchmarks with us? Could you give us memory performance numbers?
Or anything really. 😉

 

[Pinstripe]

That chart is wrong. GCN doesnt support conservative rasterization for example.

I always believed at least GCN1.2 (Tonga) does support FT12_1:
http://www.chiploco.com/amd-iceland-tonga-new-technologies-35155/

 

[ShintaiDK]

It does




You seem to know a lot about Fiji, can you share some benchmarks with us? Could you give us memory performance numbers?
Or anything really. 😉

You can try run DXcapsviewer and show that it does.

Memory speed is 512GB/sec. Scaling wise its more or less whats needed for the 4096SPs. You can freely choose if you want to extrapolate from the 285 or the 290X. The only thing that can surprise is if its somehow clocked to the moon instead of around 1Ghz.

 

[arandomguy]

There is way too much importance placed on Fiji, very few people buy graphics cards in that range. Nvidia did not gain marketshare because they are selling Titan X in droves but because they released the GTX 970 while AMD in the same time frame released the appallingly priced R9 285...

 

[zlatan]

That chart is wrong. GCN doesnt support conservative rasterization for example.

Why not? GCN has manual VS-PS interpolation, it's unrolled by the complier. I'm already used conservative rasterization on PS4 GNM, and it is possible with the newer Mantle version, but I didn't try. It should be possible on D3D12 too, but many of the multiplatform devs don't care about it.

I always believed at least GCN1.2 (Tonga) does support FT12_1:
http://www.chiploco.com/amd-iceland-tonga-new-technologies-35155/

This is not hardware specific. With manual VS-PS interpolation GCN can support any kind of interpolation form. They just need to write a complier for it.

ROVs is also not hard. It's basically an UAV ordering solution with mutex-based approach. It needs a little shared memory on the whole GPU, so the hardware can use mutex to controll the work. For OIT this will lead to slightly higher performance and lower memory footprint.

 

[Erenhardt]

Why not? GCN has manual VS-PS interpolation, it's unrolled by the complier. I'm already used conservative rasterization on PS4 GNM, and it is possible with the newer Mantle version, but I didn't try. It should be possible on D3D12 too, but many of the multiplatform devs don't care about it.



This is not hardware specific. With manual VS-PS interpolation GCN can support any kind of interpolation form. They just need to write a complier for it.

Thanks for the explanation.

 

[ShintaiDK]

You can emulate as much as you want, with epicly slow results. But DX12 mandates hardware support. Its already been covered before.

ROVs are lacking too.

290X

285

GTX 980

DX12.1

 

[dacostafilipe]

You can start by proving it first.

To be clear, I did not say that Maxwell is a GCN copy, it just took what worked great in GCN and applied it to Kepler.

Smaller resources grouping (less Cores per SM), more complex crossbars, more cache ... everything to improve architecture efficiency.

 

[zlatan]

You can emulate as much as you want, with epicly slow results. But DX12 mandates hardware support. Its already been covered before.

The interpolation is already emulated on GCN. All the work is done on the shaders. It won't be different with modified interpolation.

ROVs are lacking too.

It can be supported. ROVs is basically a PixelSync copy, and AMD support it on OpenGL.

Sure, I'm know that the devs don't care about ROVs, and conservative rasterization right now, because there are many other new things on D3D12, and they don't have time for these extra features. That's understandable, but the support can be provided with a driver update.

 

[ShintaiDK]

A summon up from how Microsoft views it:

 

[zlatan]

That picture is false. Kepler can support resource binding TIER2. I wrote the support for it to our engine last week.
Please, if you guys don't understand how is this works, than don't try to explain it.

Also GCN can support unlimited number of UAV, which is also a tested feature by me, because our engine can't activate every graphics effect on resource binding TIER2. We use more than 100 UAVs for the pipeline.

 

[frozentundra123456]

It does




You seem to know a lot about Fiji, can you share some benchmarks with us? Could you give us memory performance numbers?
Or anything really. 😉

He must be getting them from the same place you are getting the numbers that show how Fiji is going to stomp the competition.

 

[sontin]

Why not? GCN has manual VS-PS interpolation, it's unrolled by the complier. I'm already used conservative rasterization on PS4 GNM, and it is possible with the newer Mantle version, but I didn't try. It should be possible on D3D12 too, but many of the multiplatform devs don't care about it.

ROVs is also not hard. It's basically an UAV ordering solution with mutex-based approach. It needs a little shared memory on the whole GPU, so the hardware can use mutex to controll the work. For OIT this will lead to slightly higher performance and lower memory footprint..

So, we are back at this, again?
They dont have the hardware support. Everything else doesnt matter for DX12.

 

[frozentundra123456]

Lots of wishful thinking in this thread. NVidia brings out a very powerful card at a very competitive price and this is somehow a plus for AMD because it means they have some magical card waiting to be launched that caused nVidia to panic and bring out the 980Ti.
Okay........ if you say so. I dont always agree with RS, but in this case he is spot on. AMD needs a top to bottom linep of new and more efficient cards, brought out in a timely manner, not a bunch of rebrands of rebrands, along with one halo top end card. AMD's gpu strategy is eerily similar to their cpu strategy: keep selling old, inefficient technology at a cheaper price. We all know how that worked out in the cpu market.

As far as APUs replacing dgpus, that may be the ultimate strategy, but apparently that has been delayed at least until 2017. It still just blows my mind that after all the money spent to acquire ATI, and all the years of pushing "the future is fusion" the only zen product to come out in 2016 will have no igp.