No NGG for Current Gen Vega

[Despoiler]

https://www.mail-archive.com/amd-gfx@lists.freedesktop.org/msg24458.html

Re: Making a GDS Allocation for NGG
Jin, Jian-Rong Wed, 15 Aug 2018 06:24:02 -0700

GFX9 will not be supported.

Thanks,
Jian-Rong Jin
在 2018年8月15日，下午8:33，Jakub Okoński <ja...@okonski.org<mailto:ja...@okonski.org>>
写道：

Hey Jian-Rong,

Will gfx9 get support from KMD and LLPC by the time gfx10 releases, or is it
only going to be a gfx10-and-forward feature?

If not, is there something missing in gfx9 hardware? Are you allowed to say?

Thanks,
Jakub

On Wed, 15 Aug 2018 at 08:36 Jin, Jian-Rong
<jian-rong....@amd.com<mailto:jian-rong....@amd.com>> wrote:
Hi Jakub。
Both KMD and LLPC are not ready to support NGG for gfx9. We are going to
support NGG in amdvlk for next generation GPU.

Thanks
Jian-Rong JIN

-----Original Message-----
From: amd-gfx
<amd-gfx-boun...@lists.freedesktop.org<mailto:amd-gfx-boun...@lists.freedesktop.org>>
On Behalf Of Jakub Okoński
Sent: 2018年8月14日 21:31
To: amd-gfx@lists.freedesktop.org<mailto:amd-gfx@lists.freedesktop.org>
Subject: Making a GDS Allocation for NGG

Hi folks,

I was trying to play around with NGG on Vega using the open source AMDVLK stack.
After fetching all the code, I disabled an override in PAL, under
gfx9SettingsLoader.cpp that disables the NGG setting. I've put a new option in
my
amdPalSettings.cfg: `NggMode,7` and that was enough to enable it.

When I ran a small vulkan application, it crashed with a "Not implemented"
error in Pal::Linux:evice::AllocateGds. The comment there says:

// TODO: implement it once amdgpu is ready.

I couldn't find anything specific to GDS allocation in upstream kernel, so I
looked at agd5f/amd-staging-drm-next and found some code related to
AMDGPU_GEM_DOMAIN_GDS and allocating a bo object. This is in
amdgpu_gem_create_ioctl, so I think it's accessible to userspace, with a small
change in libdrm, probably.

Is this what is needed to make a GDS allocation that NGG would accept, or is it
unrelated? Do you know if there are any other blockers for NGG besides this?

Regards,
Jakub
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org<mailto:amd-gfx@lists.freedesktop.org>
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

This is an exchange between a guy trying to test NGG and having it fail. AMD contact says KMD and LLPC (drivers) are not ready and will NOT support the feature set. NGG coming in nextgen product only.

This vindicates those of us arguing that the features weren't implemented and thus the Vega chips could be faster. I also think this also adds to the leaning that Raja was let go for failure to execute. It's a huge mistake when you have a suite of features that speed things up, that could be the difference of beating or losing to competitors, that added to your cost, but they cannot be used. I'm a little disappointed to say the least.

 

[maddie]

https://www.mail-archive.com/amd-gfx@lists.freedesktop.org/msg24458.html



This is an exchange between a guy trying to test NGG and having it fail. AMD contact says KMD and LLPC (drivers) are not ready and will NOT support the feature set. NGG coming in nextgen product only.

This vindicates those of us arguing that the features weren't implemented and thus the Vega chips could be faster. I also think this also adds to the leaning that Raja was let go for failure to execute. It's a huge mistake when you have a suite of features that speed things up, that could be the difference of beating or losing to competitors, that added to your cost, but they cannot be used. I'm a little disappointed to say the least.

Yep. I was one of those arguing that Vega had more in store, and if we didn't see any improvement, it was borked. However, in my case, I'm a little optimistic, as my next full build will be 7nm, both CPU & GPU. Want freesync, so Nvidia is not an option.

 

[Despoiler]

Yep. I was one of those arguing that Vega had more in store, and if we didn't see any improvement, it was borked. However, in my case, I'm a little optimistic, as my next full build will be 7nm, both CPU & GPU. Want freesync, so Nvidia is not an option.

The good news is that this exchange confirms AMD will have this working for their next gen chip. I can't wait to see what they have cooking.

 

[kawi6rr]

I'll start a new build with a 2600x when these new cards drop

 

[del42sa]

This is an exchange between a guy trying to test NGG and having it fail. AMD contact says KMD and LLPC (drivers) are not ready and will NOT support the feature set. NGG coming in nextgen product only.

This vindicates those of us arguing that the features weren't implemented and thus the Vega chips could be faster. I also think this also adds to the leaning that Raja was let go for failure to execute. It's a huge mistake when you have a suite of features that speed things up, that could be the difference of beating or losing to competitors, that added to your cost, but they cannot be used. I'm a little disappointed to say the least.

not only NGG fast path also Primitive Shader. Both stays as "paper dragons" from AMD fancy presentation, they made for Vega architecture. Ground breaking Vega features, essentially only Fiji with RMP. What a joke .....

 

[Shmee]

Also, according to the front page article here, 7nm GPUs from AMD are coming 2018. So we will see the vega refresh this year from that timeline. 7nm Zen 2 soon after. I guess we will see if 7nm and any other tricks under the hood can help the vega refresh. Might make for an interesting competition come holidays against the RTX series.

 

[zinfamous]

Also, according to the front page article here, 7nm GPUs from AMD are coming 2018. So we will see the vega refresh this year from that timeline. 7nm Zen 2 soon after. I guess we will see if 7nm and any other tricks under the hood can help the vega refresh. Might make for an interesting competition come holidays against the RTX series.

Consumer Vega refresh this year is very likely? I can't recall seeing anything that convinced me of this. I'm in the market now for a new GPU, and I want to be able to push a new display that I purchased, @3440 x1440p ~60-100hz if possible. I am cool with finding a deal on a new or even better: used 580 and waiting until the end of the year to replace it with a refreshed Vega 56 if there is some appreciable benefit to that. Otherwise, I've convinced myself that I'm happy to stick with current Vega 56 for the next 2+ years.

Comparing benchmarks between the two, the gap at that resolution is noticeable enough where I lean towards Vega, if not for the price right now...which I know is basically at real MSRP, it's just such an expensive card for what it is, imo

Benchmarks say I should be fine at ~30fps on a ultra-wide 2k display with an RX 580, but I'm personally not fine with that. FWIW, I tend to go for value and mid-low-high end. My current 280X has been working well for me going on 5 years now....but I don't think that will be able to do the deed on the new display.

 

[maddie]

not only NGG fast path also Primitive Shader. Both stays as "paper dragons" from AMD fancy presentation, they made for Vega architecture. Ground breaking Vega features, essentially only Fiji with RMP. What a joke .....

Primitive shader is a subset of NGG, as is the (IWD), intelligent workload distributor, which I think is also a very big deal, as it's internal to the GPU (no game coding needed). If none of these are working, then Vega loses (20-30% ??) performance.

 

[del42sa]

...... as it's internal to the GPU (no game coding needed). If none of these are working, then Vega loses (20-30% ??) performance.

Its not true, without game developer it will never work.

https://techreport.com/news/33153/radeon-rx-vega-primitive-shaders-will-need-api-support

 

[maddie]

Its not true, without game developer it will never work.

https://techreport.com/news/33153/radeon-rx-vega-primitive-shaders-will-need-api-support

I was talking about the IWD, not Primitive Shaders.

From the linked article.

"The issue sparking the most controversy today is the status of the Next-Generation Geometry Engine, better known as "primitive shaders" in enthusiast shorthand. AMD emphasized that the Next-Generation Geometry path has several components, not just the more flexible programming model exposed through primitive shaders. One of those is the Intelligent Workgroup Distributor, a feature that ensures the GPU avoids performance-harming operations like context rolls (or context switches) and enjoys maximum occupancy of its execution units. AMD confirmed that the IWD is enabled, and that any performance benefits from that feature are already being demonstrated by Vega cards today."

I'm wrong in thinking that the IWD is not working, as AMD has stated it is working. However, I find this quite strange as their appears to be no improvement between Fiji and Vega per clock, which is exactly what you work see with greater utilization of the shaders.

 

[Headfoot]

Vega has turned out to be another immensely disappointing release. Since the 290 AMD just doesn't get it.

Tonga -> small front end tweaks that amount to equivalent performance to an existing older product.
Fij -> small front end tweaks that don't add much. new memory architecture that doesn't provide a competitive advantage over the competition but makes it really expensive. somewhat scaled up - but instead of scaling up the part that matters for current games, they rely on the small ineffectual tweaks and Devs to carry their water, which never materializes.
Polaris -> small tweaks that amount to equivalent performance as an older product. Again.
Vega -> small tweaks that amount to a decent clock speed increase.... which they completely squander by not scaling up the rest of the GPU. Rest of transistor budget wasted on features that require Devs to carry their water, which never materializes.

Compare this to the 290 and 290x. Scaled up version of all the good stuff from the 7970. Small useful tweaks. Better default clock speed. End result? Great performance, great price, one of the best GPUs AMD has released.

It doesn't take a rocket scientist. For the love of god, just scale up the GPU, keep your clockspeed magic, and deliver it in a timely manner. Cut the constant tricks and gotchas and just deliver a GPU that has more of the stuff that makes TODAYS games go fast, and higher clockspeed, and call it a day. That's exactly what nVidia did for Pascal which was wildly successful, and gave them years of breathing room to build the next architecture that tries all of the future looking stuff AMD has failed to do since the original GCN release.

 

[Despoiler]

Vega has turned out to be another immensely disappointing release. Since the 290 AMD just doesn't get it.

Tonga -> small front end tweaks that amount to equivalent performance to an existing older product.
Fij -> small front end tweaks that don't add much. new memory architecture that doesn't provide a competitive advantage over the competition but makes it really expensive. somewhat scaled up - but instead of scaling up the part that matters for current games, they rely on the small ineffectual tweaks and Devs to carry their water, which never materializes.
Polaris -> small tweaks that amount to equivalent performance as an older product. Again.
Vega -> small tweaks that amount to a decent clock speed increase.... which they completely squander by not scaling up the rest of the GPU. Rest of transistor budget wasted on features that require Devs to carry their water, which never materializes.

Compare this to the 290 and 290x. Scaled up version of all the good stuff from the 7970. Small useful tweaks. Better default clock speed. End result? Great performance, great price, one of the best GPUs AMD has released.

It doesn't take a rocket scientist. For the love of god, just scale up the GPU, keep your clockspeed magic, and deliver it in a timely manner. Cut the constant tricks and gotchas and just deliver a GPU that has more of the stuff that makes TODAYS games go fast, and higher clockspeed, and call it a day. That's exactly what nVidia did for Pascal which was wildly successful, and gave them years of breathing room to build the next architecture that tries all of the future looking stuff AMD has failed to do since the original GCN release.

I don't think it's that they don't get it. Vega was built as a professional workstation chip that they also sold as a gaming chip. AMD is getting it in that they are going where the big bucks are. Vega bested Quadro in many professional tasks as well as being much cheaper. Instinct is the same thing. AMD gives you a lot more perf for far less money than Nvidia. Prior to that Zen hadn't been released and AMD was just trying to stay alive. Not getting it is a stretch for engineers and fellows. Not having budgets for it was reality.


No discussion of Nvidia products is allowed in the AMD sub board.

AT Moderator ElFenix

 

[Muhammed]

Vega bested Quadro in many professional tasks as well as being much cheaper. Instinct is the same thing.

The only thing Vega was good at is mining, no body bought Vega for AI or compute or professional, they all bough Tesla and Quadros, because they were a lot faster and have a robust ecosystem around them. And now with RTX Quadro is more important than ever. The same thing with Tensor and Volta.

Also Vega didn't best squat, a TitanXp ran circles around Vega FE, and that wasn't even a Quadro variant which is significantly faster. WX9100 was worse than Vega FE even. Pascal Tesla is the benchmark of AI, even AMD avoided comparing Vega to it directly.

All of this and I still didn't mention Volta which stomped all over Vega in it's Tesla (AI) and Quadro variants. And it's going to be a slaughter house with Turing.

No discussion of Nvidia products is allowed in the AMD sub board.

AT Moderator ElFenix

 

[EXCellR8]

I think the RX Vega 64 is a pretty decent product... runs games @ 4K much easier than my previous Polaris and R9 series cards did so at the end of the day I was happy. Yep, NVIDIA sure does have more powerful cards--but they're also more expensive and obsoleted much faster. I'm also using a FreeSync monitor because it was cheaper than a comparable G-Sync model.

Everyone gloating about AI stuff but I haven't seen any actual breaking development news on the subject in well... ever.

 

[Despoiler]

The only thing Vega was good at is mining, no body bought Vega for AI or compute or professional, they all bough Tesla and Quadros, because they were a lot faster and have a robust ecosystem around them. And now with RTX Quadro is more important than ever. The same thing with Tensor and Volta.

Also Vega didn't best squat, a TitanXp ran circles around Vega FE, and that wasn't even a Quadro variant which is significantly faster. WX9100 was worse than Vega FE even. Pascal Tesla is the benchmark of AI, even AMD avoided comparing Vega to it directly.

All of this and I still didn't mention Volta which stomped all over Vega in it's Tesla (AI) and Quadro variants. And it's going to be a slaughter house with Turing.

Demonstratively false on all points. You've been proven wrong I think on almost everything you've said on this forum. Hey though...let's do it some more....

https://www.tomshardware.com/reviews/amd-radeon-vega-frontier-edition-16gb,5128-6.html This is toe to toe or beating a p6000 which is 4-5x more expensive. Ouch. You can get an Inventec AMD Instinct rack for slightly less than the price of single DGX-1 server. See how it says 3 petaflops? That's 3x perf/$$. http://blog.exxactcorp.com/digging-...instinct-gpus-miopen-gpu-accelerated-library/. The entire problem with your argument is that in order for it to be true you need to ignore cost which no one buying this gear does. http://blog.gpueater.com/en/2018/03/20/00006_tech_flops_benchmark_2/ and http://blog.gpueater.com/en/2018/04/23/00011_tech_cifar10_bench_on_tf13/ You also need to ignore the fact that it's utterly stupid to buy Nvidia GPUs for tensor ops when Google's TPUv2 walks all over them at better perf per watt. ie You buy AMD racks and then add Google racks for a better solution that costs less.


No discussion of Nvidia products is allowed in the AMD sub board.

AT Moderator ElFenix

 

[Despoiler]

I'm not sure if this muddies the waters again, but in the OP a user was trying to activate and program the features. From Beyond3d a dev proves that the driver and Vega are using the features.
https://forum.beyond3d.com/posts/2041569/ I re-upped the images to imgur for the time being. The captures are from AMD Profiler on RX470 vs Vega64. https://imgur.com/a/5td1JOt

 

[Muhammed]

https://www.tomshardware.com/reviews/amd-radeon-vega-frontier-edition-16gb,5128-6.html This is toe to toe or beating a p6000 which is 4-5x more expensive. Ouch. You can get an Inventec AMD Instinct rack for slightly less than the price of single DGX-1 server. See how it says 3 petaflops? That's 3x perf/$$. http://blog.exxactcorp.com/digging-...instinct-gpus-miopen-gpu-accelerated-library/. The entire problem with your argument is that in order for it to be true you need to ignore cost which no one buying this gear does. http://blog.gpueater.com/en/2018/03/20/00006_tech_flops_benchmark_2/ and http://blog.gpueater.com/en/2018/04/23/00011_tech_cifar10_bench_on_tf13/ You also need to ignore the fact that it's utterly stupid to buy Nvidia GPUs for tensor ops when Google's TPUv2 walks all over them at better perf per watt. ie You buy AMD racks and then add Google racks for a better solution that costs less.

LOL? What is this? some collection of some pathetic experimental pathways/tests? becasue that is at best what you have there! None of that crap you posted is actually true in any capacity, even Google uses NVIDIA's AI solutions, they never even bothered with AMD's instinct. NO ONE DOES.

The problem with your imaginary post is that you forgot this fact, NVIDIA has several cards way above the Vega 64/Vega FE:

1080Ti
Quadro 5000
TitanXP
Quadro 6000
Quadro GP100
TitanV
Quadro V100


The TitanXP and Quadro 6000 are both significantly faster than Vega FE and WX9100. Quadro GP100 absolutely crushes Vega FE in every metric possible. TitanV and Quadro V100 absolutely pummels the Vega FE to oblivion.


Also Tesla GP100 trounces Vega Instinct, it's not even a fair fight. And Tesla V100 kills them both with the Tensor cores.

Again the only thing Vega was successful as was mining, take that away no one even bothers with it, industry professtionals: they have Quadro Pascals, and Volta, and now Turing. AI: Only NVIDIA is the viable option there.


No discussion of Nvidia products is allowed in the AMD sub board.

AT Moderator ElFenix

 

[Headfoot]

I think you have a fair point that Vega's lackluster performance in consumer gaming may indicate it just wasn't really designed for that purpose. Some of the new features like the primitive shaders do seem to be gaming focused though, right?

 

[Head1985]

Vega has turned out to be another immensely disappointing release. Since the 290 AMD just doesn't get it.

Tonga -> small front end tweaks that amount to equivalent performance to an existing older product.
Fij -> small front end tweaks that don't add much. new memory architecture that doesn't provide a competitive advantage over the competition but makes it really expensive. somewhat scaled up - but instead of scaling up the part that matters for current games, they rely on the small ineffectual tweaks and Devs to carry their water, which never materializes.
Polaris -> small tweaks that amount to equivalent performance as an older product. Again.
Vega -> small tweaks that amount to a decent clock speed increase.... which they completely squander by not scaling up the rest of the GPU. Rest of transistor budget wasted on features that require Devs to carry their water, which never materializes.

Compare this to the 290 and 290x. Scaled up version of all the good stuff from the 7970. Small useful tweaks. Better default clock speed. End result? Great performance, great price, one of the best GPUs AMD has released.

It doesn't take a rocket scientist. For the love of god, just scale up the GPU, keep your clockspeed magic, and deliver it in a timely manner. Cut the constant tricks and gotchas and just deliver a GPU that has more of the stuff that makes TODAYS games go fast, and higher clockspeed, and call it a day. That's exactly what nVidia did for Pascal which was wildly successful, and gave them years of breathing room to build the next architecture that tries all of the future looking stuff AMD has failed to do since the original GCN release.

At this point it is too late for GCN.They should scale GCN to 6x shader engines/96rops after 290x launch, but they probably dont know how.Everyone from ATI whos designed GCN already left AMD when rory read was CEO.
Now they dont even know how make primitive shaders and draw stream binning rasterizer work...

I just hope Navi will be last GCN product and in 2020-2021 they will have new architecture.