From ATI to AMD back to ATI? A Journey in Futility @ [H]

[razor1]

Lot of words, little substance. The Vega reveal was clear enough. The changes were extensive and chip wide with added major functions new to GPUs. To call these changes 'trivial' implies what AMD said is mostly lies and propaganda. What calling these changes 'trivial' implies about the poster I'll leave unsaid.

There is a ton of substance in my post, I suggest you look up the async threads to figure out I'm not making things up. I don't have the patience to explain those things to someone who isn't bothered by reading......

You don't understand the different parts of a GPU, you don't know what Async compute is or how it functions at a GPU architectural level, you don't know how the polygon throughput improvements in Polaris show up outside of synthetics and when we do look at synthetics they only match a gtx 960.... which is 2x the geometry through put of Fiji.

You don't sit here and make crap up based on AMD marketing that has shown so many times they are incapable of keeping things realistic when you look at final "projected" best case numbers.

Everything AMD states has truth in it, how much truth is face value, don't extrapolate off that cause you will be disappointed. If they say something, its always best case, in most cases its lower, I can show you interviews about the changes in Vega about polygon through put and the TBR and primitive discard in conjunction with primitive shaders where best case is 11 polys discard which is the 4x polygon through put in best cases if not using primitive shaders 2.5x over Fiji, guess what, that is just a bit higher than Polaris, not too much though based on synthetics that we know for sure.

Red gaming tech on youtube has an interview with Scott Wasson about these things so be my guest and look it up if you like, AMD's own people aren't making the things up, they are being straight about it, yet you are believing in the highest possible performance increases without understanding the limitations of what they are saying are. And there are specific limitations.

What you are posting is apparently you know more then any other reviewer about Aysnc compute, any other programmer about it, or AMD empolyees about Vega?



Either you are making shit up because of what ever reason, or what I'm hearing here, is what I've been saying, you can't use the higher level features of Vega which are the numbers Vega had up in their word cloud, are going to need to use primitive shaders. Unlike all the polygon through put increases nV has done or their tiled rasterizer which are just automatic.

So pretty much Vega couldn't fixed the issues they had with its geometry pipeline, they are now telling dev's code it for their array. Which goes to show you anything to do with that will not work automagically. Cause AMD will need to release an SDK or API that include those extensions otherwise they can't be accessed. Ironically, games that are part of AMD's game dev program (one for sure Dice's engine) already do this in code lol. There has been culling examples in interviews and power point presentations by Dice which they mentioned this in conjunction with Mantle and later DX12 for AMD hardware for primitive discard!

 

[Shintai]

The latest official roadmap lists Navi for 2018 while leaked slides enumerates a Vega 20. Perhaps Navi was renamed. Perhaps Navi is engineered for a 7nm process that will be delayed and Vega 20 is a 14nm enhanced Vega. Insufficient data to know for certain at this time. I speculated based on the official released roadmap.

The information is known, so no need to make up random nonsenses. Vega 20 is 7nm and Navi is 2019.

 

[Tyns]

AMD communicated in detail the changes were non-trivial. There are investors related legal requirements. I would consider myself an ignorant dolt to not take AMD at it's word with an offical presentation.

Those 'deluded' AMD investor fanboys are making money hand over fist. What is your point exactly?

My point is you criticize other posters for lack of substance when you yourself bring nothing to the table. AMD has never misled their investors before, eh? Investor relations is just another form of marketing.

 

[{NG}Fidel]

To be fair both sides in here are making claims they cant possibly back up. Half the claims are based on feelings which are rooted in actions from the Past. Truth is we cannot be sure of what changes they have made until its released. Saying the changes are trivial without proof is baseless.

 

[razor1]

Does anyone remember these slides?

The problem was and is now with Vega some of the things are overlapping or were part of GCN 1.2 already.

Memory compression was already there in Tonga and Fiji.

Primitive discard was to give close to 350% improved performance was increasing performance with MSAA. Vega is to get 2.5X the performance increases from these things from Tonga, so add in MSAA on top of that well yeah kinda end up with where Polaris increases are.

4th generation core Next CU, NCU that is what they have called it in Vega, which stands for Next CU which was already mentioned in Polaris slide decks without the acronym. The main difference is the packed math for Vega which won't be used till Vega is released and games are being made for it (FP16), no real way around that and we have no clue what the performance implications are for such a mixed mode, the video even gets into that a bit where Scott wasn't comfortable answering that question, well because its going to be highly variable from the looks of it.

Now lets look at something else Improved load balancing, guess what that was mentioned with Polaris too! the new command processor, the slide for Vega mentions the IWD, that is part of the command processors.

Many of the changes in Polaris is what Vega slides have been talking about more in depth and a few other tweaks is what I see.

If we start going through the Vega word cloud which are the best numbers Vega will represent over Tonga

4x the power efficiency: It better be 4x but when you look at tonga and figure that all out, you end up with a chip that has 2x the performance of Polaris and 75% increased power consumption.

2x the through put, well yeah explained that up above, still around Polaris......

All of the performance, efficiency numbers are coming from a chip that is 2 or maybe even considered 3 generation old GCN, although we can say 2 as Fiji and Tonga have common roots.

All the other stuff, HBCC, DSBR, need Primitive shaders to get the most out of them, which the video above kinda tells us.

 

[dook43]

I dont dare even try otherwise my license to return stuff to Amazon might get revoked.

I had to sell my 290x cheap to get rid of it because it screwed with sound over DVI/HDMI.
Video devices kept being detected as DVI with no sound, even when it was HDMI.
DVI devices would give sound when detected as HDMI.
Problem was, it was random what they would be detected as on boot and even after boot they could change between HDMI and DVI.
Any device configured as DVI had no sound and they cannot be manually configured. (all were HDMI hardware on the output end)
Drove me mad.

Eventually gave up on it after 1.5 yrs, bought a GTX980 and now a 980ti. Not a problem on either.

I have dual R9 290s and SB Z, and the only sound-related annoyance I have is that Windows keeps wanting to install the HDMI Audio drivers and set it to primary sound device. Haven't wanted to touch sound over HDMI.

Regardless, if I needed that feature to work correctly, I *guess* I'd be annoyed enough to return it. Definitely not something I'd boycott a brand over.

However, implementing invisible tessellation to artificially gimp AMD in Batman games is something I'd boycott a brand over.

 

[Nenu]

I have dual R9 290s and SB Z, and the only sound-related annoyance I have is that Windows keeps wanting to install the HDMI Audio drivers and set it to primary sound device. Haven't wanted to touch sound over HDMI.

Regardless, if I needed that feature to work correctly, I *guess* I'd be annoyed enough to return it. Definitely not something I'd boycott a brand over.

However, implementing invisible tessellation to artificially gimp AMD in Batman games is something I'd boycott a brand over.

I gave them 1.5 years to sort it out. It was never fixed, I had enough.
I told their rep on this forum about it, he offered no help other than how to install the video driver.
There is no chance I will risk going through that again, it ruined my gaming and partially my AV experience during that period.

My hifi is extremely high definition, I bought it to get the best audio, not none!
I was hesitant to play a game for fear it would take an hour to accidentally get it working in surround.
I used to end up swearing so much by the time it would work that gaming wasnt enjoyable.
The convenience of using my PC as a movie server was gone.

ps boycott Batman games.

 

[Anarchist4000]

There is a ton of substance in my post, I suggest you look up the async threads to figure out I'm not making things up. I don't have the patience to explain those things to someone who isn't bothered by reading......

You don't understand the different parts of a GPU, you don't know what Async compute is or how it functions at a GPU architectural level, you don't know how the polygon throughput improvements in Polaris show up outside of synthetics and when we do look at synthetics they only match a gtx 960.... which is 2x the geometry through put of Fiji.

You mean the pages you went on about how async is absolutely worthless? Only to have every DX12/Vulkan title use the feature with devs seeing significant improvements. Every review site testing the feature and showing a difference. Devs having to disable the feature on Maxwell to avoid catastrophic performance.

Why is polygon throughput even an issue here? Polaris already seems more than competitive enough with geometry performance. Unless games transition towards colored triangles in favor of texturing meshes I'm not sure there is a huge need to increase the rate. Even Fiji had ample geometry performance, excluding scenarios where it was inundated with worthless invisible triangles.

All the other stuff, HBCC, DSBR, need Primitive shaders to get the most out of them, which the video above kinda tells us.

Why would a cache need a primitive shader? It's a hardware implementation of a memory manager that should be transparent to the kernel executing. No reason to think primitive shaders can't encapsulate the vertex and geometry shaders that is replaces either.

 

[razor1]

You mean the pages you went on about how async is absolutely worthless? Only to have every DX12/Vulkan title use the feature with devs seeing significant improvements. Every review site testing the feature and showing a difference. Devs having to disable the feature on Maxwell to avoid catastrophic performance.

Why is polygon throughput even an issue here? Polaris already seems more than competitive enough with geometry performance. Unless games transition towards colored triangles in favor of texturing meshes I'm not sure there is a huge need to increase the rate. Even Fiji had ample geometry performance, excluding scenarios where it was inundated with worthless invisible triangles.

Never stated it is worth less, just not as important as it was made out to be, and if you think of a max of 15% and average of 5% in most games miraculous improvements, that is just I don't know. I link back to the async threads where we discussed this and yeah what we see right now isn't what you thought async performance benefits to be and that was year and half ago where you were thinking 30%, 35% increases and we still haven't see that yet.

Polygon through put is less of an issue with Polaris, but it still needs more help in that front.

Fiji is geometry blocked this is why at some time we see Polaris in specific games catch up to Fiji. And this is also why Fiji and Tonga, and Hawaii (older GCN than 1.4) can't use its shader array effectively. So with chips that the shader array is bottlenecked by geometry which is all of GCN, Polaris will be in that boat too with next gen games where polycounts are doubling (working on a game right now that is using x4 the polycounts than most games today but not planning on releasing for another 4 years so in 2 gens yeah x4 the polycounts without tessellation looks good and we aren't even at pixel level triangle sizes even with base polygon counts at 40 million visible with tessellation at x4), async will actually help them more lol. Under-utilization getting help from more programming, should be the other way around.

Back to triangles and amount of them, since I know you don't know game asset production you don't understand why I would rather use morph targets and use procedural textures rather then different texture sets on lets say something like damaged hull on a ship. polygons are cheap on memory textures at 4k to 8k are expensive on vram.

Why would a cache need a primitive shader? It's a hardware implementation of a memory manager that should be transparent to the kernel executing. No reason to think primitive shaders can't encapsulate the vertex and geometry shaders that is replaces either.

I just lumped it together, so yeah the memory manager is nothing special, doesn't matter if its a High bandwidth controller or not, it will work with other types of graphics memory too. and actually most game engines already have streaming asset capabilities anyways, so its neither here nor there. And we did talk about this before, dev's still need access and control over memory management, or is all the talk about LLAPI and user control BS?

I don't care if they replace vertex and pixel shaders or GS units, and create a unified shader unit that works on all types of data types with good performance all the way around, that isn't the point. DX10 tried that with GS units and failed hence why fixed function hull shaders were added in DX11. This is why I don't see this a big step away from what they had already just made it more programmable by the use of their existing shader array with new functions. If AMD wants to keep backward compatibility with DX12, 11 tessellation pipeline and GS units they will need to still have those units, just tweak them so they have access to data that is available in shader pipeline.

AMD has a GPU that is much bigger than GP 104 even bigger than a GP102) that seems to have performance around the same as GP104 and at power consumption higher than GP 104 possibly even higher than GP102. Those are the three negatives AMD has been trying to get away from since they released GCN. All the while since GCN's first release promised just wait we have more to show you with newer games, API, features which will show up down the road like, yeah Async, Mantle, then DX12 and Vulkan. By the time Dx12 came out and was being used and Vulkan, well they got crushed by Pascal. A company can't rely on down the road because they aren't the biggest boy on the block, even if they were, they still can't rely on it because its feeding right into the hands of the competition when they are ready to release their products that can do better overall.

Give you a prefect example, VR and the TAM, guess what that feed into nV's Pascal launch. And now DX12 games, yeah it looked good for when Maxwell was around but Pascal came out in 6 months or so of DX12 games, and DX12 games quite a few of them were not in, well running performance circles around DX11 counter parts at launch sometimes worse to boot. So what are ya left with Vulkan we got one game Doom and it does run better on AMD cards but shit that is what they want to show us when they want to butter up a live demo that seems to be around a GP104 performance? Thats not butter that is salt. Cause the size of this chip with HBM2 it should be at GP102 performance if it doesn't reach that, and they need to price it down to 500 bucks or less to for market competitiveness RTG is in for a major overhaul of the way things needs to be done to get back on par with nV, that would mean 4 generations of GPU's AMD has slowly been loosing ground, you don't make that up in one generation leap without nV f'ing up. Just not possible.

Down the road isn't good enough when there is an aggressive competitor that is more capable then AMD is from a financial point of view.

 

[BrianB]

fyi:
Nvidia Call - Revenue Statement Assumes Discontinuation of Intel Licensing Agreement

 

[Jadawin]

"Now, turning to the outlook for the first quarter of fiscal 2018. We expect revenue to be $1.9 billion, plus or minus 2%. At the mid-point, this represents 46% growth over the prior year. We expect Datacenter to grow sequentially. Our GAAP and non-GAAP gross margins are expected to be 59.5% and 59.7%, respectively, plus or minus 50 basis points. This guidance assumes that our licensing agreement with Intel ends at March and does not renew. GAAP operating expenses are expected to be approximately $603 million. Non-GAAP operating expenses are expected to be approximately $520 million."

http://seekingalpha.com/article/4044678-nvidia-nvda-q4-2017-results-earnings-call-transcript?page=3

 

[TheLAWNoob]

Actually solidworks reportedly is perfect with quad cores.
But solidworks on raw iGPU... Ouch.

We not only model and simulate in Solidworks, but we also have to render the simulation for assignments.

It's painful rendering on the Xeon E3. I have a friend who lets people render on his mobile i7. I render on my i5 6600K. I'll probably get a 8 core Zen for rendering. But if it sucks then I'll go for 14+ core Xeon E5 engineering samples.

 

[CSI_PC]

We not only model and simulate in Solidworks, but we also have to render the simulation for assignments.

It's painful rendering on the Xeon E3. I have a friend who lets people render on his mobile i7. I render on my i5 6600K. I'll probably get a 8 core Zen for rendering. But if it sucks then I'll go for 14+ core Xeon E5 engineering samples.

You thought about laptops with latest mobile Pascal Quadro?
They should be coming out soon and the Pascal Quadro performance in Solidworks with the pro drivers is very good.

http://aecmag.com/technology-mainme...-quadro-p2000-and-quadro-p4000-for-cad-viz-vr
Yeah they are the dGPU, but should see the mobile variants available very soon in products: http://www.nextpowerup.com/news/33094/nvidia-officially-announces-mobile-quadro-pascal-based-gpus/
http://www.nvidia.com/object/quadro-for-mobile-workstations.html

Cheers

 

[noko]

Never stated it is worth less, just not as important as it was made out to be, and if you think of a max of 15% and average of 5% in most games miraculous improvements, that is just I don't know. I link back to the async threads where we discussed this and yeah what we see right now isn't what you thought async performance benefits to be and that was year and half ago where you were thinking 30%, 35% increases and we still haven't see that yet.

Polygon through put is less of an issue with Polaris, but it still needs more help in that front.

Fiji is geometry blocked this is why at some time we see Polaris in specific games catch up to Fiji. And this is also why Fiji and Tonga, and Hawaii (older GCN than 1.4) can't use its shader array effectively. So with chips that the shader array is bottlenecked by geometry which is all of GCN, Polaris will be in that boat too with next gen games where polycounts are doubling (working on a game right now that is using x4 the polycounts than most games today but not planning on releasing for another 4 years so in 2 gens yeah x4 the polycounts without tessellation looks good and we aren't even at pixel level triangle sizes even with base polygon counts at 40 million visible with tessellation at x4), async will actually help them more lol. Under-utilization getting help from more programming, should be the other way around.

Back to triangles and amount of them, since I know you don't know game asset production you don't understand why I would rather use morph targets and use procedural textures rather then different texture sets on lets say something like damaged hull on a ship. polygons are cheap on memory textures at 4k to 8k are expensive on vram.



I just lumped it together, so yeah the memory manager is nothing special, doesn't matter if its a High bandwidth controller or not, it will work with other types of graphics memory too. and actually most game engines already have streaming asset capabilities anyways, so its neither here nor there. And we did talk about this before, dev's still need access and control over memory management, or is all the talk about LLAPI and user control BS?

I don't care if they replace vertex and pixel shaders or GS units, and create a unified shader unit that works on all types of data types with good performance all the way around, that isn't the point. DX10 tried that with GS units and failed hence why fixed function hull shaders were added in DX11. This is why I don't see this a big step away from what they had already just made it more programmable by the use of their existing shader array with new functions. If AMD wants to keep backward compatibility with DX12, 11 tessellation pipeline and GS units they will need to still have those units, just tweak them so they have access to data that is available in shader pipeline.

AMD has a GPU that is much bigger than GP 104 even bigger than a GP102) that seems to have performance around the same as GP104 and at power consumption higher than GP 104 possibly even higher than GP102. Those are the three negatives AMD has been trying to get away from since they released GCN. All the while since GCN's first release promised just wait we have more to show you with newer games, API, features which will show up down the road like, yeah Async, Mantle, then DX12 and Vulkan. By the time Dx12 came out and was being used and Vulkan, well they got crushed by Pascal. A company can't rely on down the road because they aren't the biggest boy on the block, even if they were, they still can't rely on it because its feeding right into the hands of the competition when they are ready to release their products that can do better overall.

Give you a prefect example, VR and the TAM, guess what that feed into nV's Pascal launch. And now DX12 games, yeah it looked good for when Maxwell was around but Pascal came out in 6 months or so of DX12 games, and DX12 games quite a few of them were not in, well running performance circles around DX11 counter parts at launch sometimes worse to boot. So what are ya left with Vulkan we got one game Doom and it does run better on AMD cards but shit that is what they want to show us when they want to butter up a live demo that seems to be around a GP104 performance? Thats not butter that is salt. Cause the size of this chip with HBM2 it should be at GP102 performance if it doesn't reach that, and they need to price it down to 500 bucks or less to for market competitiveness RTG is in for a major overhaul of the way things needs to be done to get back on par with nV, that would mean 4 generations of GPU's AMD has slowly been loosing ground, you don't make that up in one generation leap without nV f'ing up. Just not possible.

Down the road isn't good enough when there is an aggressive competitor that is more capable then AMD is from a financial point of view.

Razor1,

Bolded sections of your rhetoric above, HardOCP did some real hands on, very hard work but really what stands out this time is the large Dataset that HardOCP did in recent Driver history overviews for AMD and Nvidia. The larger dataset nails the performance accurately. Here the FuryX is basically equal to the 980Ti from an experience standpoint. For the games tested there is no real geometry block and if there is then Nvidia must have some blocks as well since both perform about equal. I agree Nvidia geometry performance is better but it is not giving any clear convincing benefit from HardOCP game data on driver performance over time and current performance levels with the most recent drivers For Fiji and Maxwell. One could say AMD is giving what is needed for the time like what you say with Nvidia and DX 12. Now those with Fiji cards may well have better experience as time goes on then those with Maxwell cards making the Maxwell bottleneck more significant or hurting then then AMD geometry performance.

Now in VR where you have to process double the geometry, for each eye maybe first clear instance that is causing performance bottlenecks but I am uncertain that is the case, since some of the VR games with poor performance with AMD are not that geometry intensive - it maybe more due to latency of AMD geometry pipeline or lack of cache more then anything else - which Vega will have. I look forward to Vega VR performance testing here at HardOCP.

Plus the other aspect with geometry, it scales with clock speed. If Vega is significantly faster in clock speed, even with very little hardware improvements for geometry processing (which is not the case) it's geometry performance will scale with that.

 

[CSI_PC]

Scott Wasson did a great article on Fiji explaining the benefits and limitations.
In some ways it is much better than Hawaii and yet others equal, which is why the 390X was so close to the Fury X in some games.
http://techreport.com/review/28499/amd-radeon-fury-x-architecture-revealed

Now we have a pretty good sense of things. In certain respects, Fiji has grown by roughly half-again compared to Hawaii, including peak shader arithmetic, texture filtering capacity, and memory bandwidth. That 512 GB/s of memory bandwidth comes courtesy of HBM, Fiji's signature innovation, and puts the Fury X in a class by itself in at least one department.

In other respects, including peak triangle throughput for rasterization and pixel fill rates, Fiji is simply no more capable in theory than Hawaii. As a result, Fiji offers a very different mix of resources than its predecessor. There's tons more shader and computing power on tap, and the Fury X can access memory via its texturing units and HBM interfaces at much higher rates than the R9 290X.

In situations where a game's performance is limited primarily by shader effects processing, texturing, or memory bandwidth, the Fury X should easily outpace the 290X. On the other hand, if gaming performance is gated by any sort of ROP throughput—including raw pixel-pushing power, blending rates for multisampled anti-aliasing, or effects based on depth and stencil like shadowing—the Fury X has little to offer beyond the R9 290X. The same is true for geometry throughput.

The Fury X substantially outruns the GeForce GTX 980 Ti in terms of integer texture filtering, shader math rates, and memory bandwidth, too, since the 980 Ti more or less matches the 290X in those departments. But the Fury X has only about 70% of the ROP and triangle rasterization rates of the big GeForce.

With Fiji, AMD is offering a rather different vision of how GPUs ought to be used by game developers. That's one reason I'd expect to see continuing fights between the GPU vendors over what effects folks incorporate into PC games. Nvidia will likely emphasize geometric complexity and tessellation, and AMD will probably push for prettier pixels instead of more polygon
....
Once I had a sense that the Fury X might not deliver a resounding victory for AMD in overall performance, I asked AMD's Koduri about the situation. After all, HBM is some seriously impressive tech, and the Fury X has a massive advantage in terms of both memory bandwidth and shader processing power. Why doesn't it mop the floor with the competition?

Koduri answered by stating my question another way: why didn't AMD build a bigger engine? That's an astute way to view things, because a bigger GPU engine would have taken fuller advantage of HBM's considerable bandwidth.

It is a great article and the tables with the various graphics related factors is very interesting and highlights key aspects.
They could not build a bigger engine because of interposer-die limit, and of course back then there was also the 4 limit Shader Engine (indirectly a limitation).

Cheers

 

[TheLAWNoob]

You thought about laptops with latest mobile Pascal Quadro?
They should be coming out soon and the Pascal Quadro performance in Solidworks with the pro drivers is very good.

http://aecmag.com/technology-mainme...-quadro-p2000-and-quadro-p4000-for-cad-viz-vr
Yeah they are the dGPU, but should see the mobile variants available very soon in products: http://www.nextpowerup.com/news/33094/nvidia-officially-announces-mobile-quadro-pascal-based-gpus/
http://www.nvidia.com/object/quadro-for-mobile-workstations.html

Cheers

It was an interesting read. However I was more concerned about the time it takes to render a video of the simulations, not rendering a still photo.

For most of the regular solidworks, a mid range card should be enough most of the time. For simple scenes and simulations intel integrated could be good enough for 1080p 60fps. The professional card only features can be unlocked using RealHack or whatever that program is called. I'm in third year and don't make any off of my soildworks skills, so I'd rather not spend a lot of money on Quadros.

 

[razor1]

Razor1,

Bolded sections of your rhetoric above, HardOCP did some real hands on, very hard work but really what stands out this time is the large Dataset that HardOCP did in recent Driver history overviews for AMD and Nvidia. The larger dataset nails the performance accurately. Here the FuryX is basically equal to the 980Ti from an experience standpoint. For the games tested there is no real geometry block and if there is then Nvidia must have some blocks as well since both perform about equal. I agree Nvidia geometry performance is better but it is not giving any clear convincing benefit from HardOCP game data on driver performance over time and current performance levels with the most recent drivers For Fiji and Maxwell. One could say AMD is giving what is needed for the time like what you say with Nvidia and DX 12. Now those with Fiji cards may well have better experience as time goes on then those with Maxwell cards making the Maxwell bottleneck more significant or hurting then then AMD geometry performance.

Now in VR where you have to process double the geometry, for each eye maybe first clear instance that is causing performance bottlenecks but I am uncertain that is the case, since some of the VR games with poor performance with AMD are not that geometry intensive - it maybe more due to latency of AMD geometry pipeline or lack of cache more then anything else - which Vega will have. I look forward to Vega VR performance testing here at HardOCP.

Plus the other aspect with geometry, it scales with clock speed. If Vega is significantly faster in clock speed, even with very little hardware improvements for geometry processing (which is not the case) it's geometry performance will scale with that.

You don't seem understand why some games Polaris (and older gen cards) is getting close to Fury X. You also don't know why AMD has stated they have geometry through put limitations compared to nV cards, they have stated this number of times, when they talk about over tessellation for many generations now.

What essential happens is when AMD's hull shaders and geometry units get flooded, their shader array stalls, because its waiting for data from the geometry pipeline.

This is very little to do with frequency, as per clock has little do with things once the bottleneck is hit. And this is a very hard hit for AMD, 2xxx, 3xx, Fiji, all hit this around 20 million polys (visible + non visible).

We are looking at 40 million visible in next gen games, with tessellation, 10 million visible without. Counting both visible and non visible, 5 times that, now nV's hardware even their current hardware doesn't have problems with this. Now AMD on the other hard, their Polaris which is already 2 - 3.5x the performance for polygon throughput of Fiji has problems with this, Vega, which they stated 2.5 over tonga, looks to be its going to have problems with it too at least with a tessellation factor of 2 which is going to be the recommended settings at least for the game I'm working on.

And you aren't looking at how games are pushing more polys, this is why in some games depending on where the character is in doors or outdoors, nV or AMD might have a distinct advantage. Its the balance between fill rates, pixel counts, triangle through put, ROP's, etc.

Thats why I stated last page, we can't sit around and try to explain something that is so complex as Async compute with a single game (to the other guy), cause there are tons of things going on that we are just blatantly skipping.

 

[CSI_PC]

It was an interesting read. However I was more concerned about the time it takes to render a video of the simulations, not rendering a still photo.

For most of the regular solidworks, a mid range card should be enough most of the time. For simple scenes and simulations intel integrated could be good enough for 1080p 60fps. The professional card only features can be unlocked using RealHack or whatever that program is called. I'm in third year and don't make any off of my soildworks skills, so I'd rather not spend a lot of money on Quadros.

Yeah, unfortunately that is currently the only reviews of the new Quadro from a similation/rendering/CAD perspective.
But shows just how powerful the new Quadros are even for their lower dGPU compared to previous gen and they work very well with Solidworks/etc (albeit I agree CPU bottleneck is also a consideration), the P2000 is a pretty good price but not sure how much the P3000 mobile will be.
Cheers

 

[noko]

You don't seem understand why some games Polaris (and older gen cards) is getting close to Fury X. You also don't know why AMD has stated they have geometry through put limitations compared to nV cards, they have stated this number of times, when they talk about over tessellation for many generations now.

What essential happens is when AMD's hull shaders and geometry units get flooded, their shader array stalls, because its waiting for data from the geometry pipeline.

This is very little to do with frequency, as per clock has little do with things once the bottleneck is hit. And this is a very hard hit for AMD, 2xxx, 3xx, Fiji, all hit this around 20 million polys (visible + non visible).

We are looking at 40 million visible in next gen games, with tessellation, 10 million visible without. Counting both visible and non visible, 5 times that, now nV's hardware even their current hardware doesn't have problems with this. Now AMD on the other hard, their Polaris which is already 2 - 3.5x the performance for polygon throughput of Fiji has problems with this, Vega, which they stated 2.5 over tonga, looks to be its going to have problems with it too at least with a tessellation factor of 2 which is going to be the recommended settings at least for the game I'm working on.

And you aren't looking at how games are pushing more polys, this is why in some games depending on where the character is in doors or outdoors, nV or AMD might have a distinct advantage. Its the balance between fill rates, pixel counts, triangle through put, ROP's, etc.

Thats why I stated last page, we can't sit around and try to explain something that is so complex as Async compute with a single game (to the other guy), cause there are tons of things going on that we are just blatantly skipping.

So your saying future games will need future hardware - OK no problem there.
Nvidia older hardware falls flat on it's face such as Fermi and Kepler especially against GCN hardware for the same generations - Hence new hardware.
Nvidia Maxwell can have issue with DX 12 - Hence new Hardware to Pascal.
I agree with you Polaris may fall short but that is a mainstream card and not a high end card, except Nvidia is way more ready for geometry on their mainstream cards but may lack sufficient ram in the end. Hence new hardware maybe needed for both for future games.
Vega? It has new Next Generation Geometry Pipeline which most likely will be incorporated into DX due to the next XBox - Nvidia will need to support this otherwise they will be the ones behind with geometry processing. AMD future hardware will be ready for the future games with increase Geometry requirements.

As a note with Dues Ex Mankind Divided DX 12, the RX 480 beats the pants off of the GTX 1060, even with all that increase geometry processing needed. There is more then just geometry to worry about in other words.

 

[razor1]

So your saying future games will need future hardware - OK no problem there.
Nvidia older hardware falls flat on it's face such as Fermi and Kepler especially against GCN hardware for the same generations - Hence new hardware.
Nvidia Maxwell can have issue with DX 12 - Hence new Hardware to Pascal.
I agree with you Polaris may fall short but that is a mainstream card and not a high end card, except Nvidia is way more ready for geometry on their mainstream cards but may lack sufficient ram in the end. Hence new hardware maybe needed for both for future games.
Vega? It has new Next Generation Geometry Pipeline which most likely will be incorporated into DX due to the next XBox - Nvidia will need to support this otherwise they will be the ones behind with geometry processing. AMD future hardware will be ready for the future games with increase Geometry requirements.

As a note with Dues Ex Mankind Divided DX 12, the RX 480 beats the pants off of the GTX 1060, even with all that increase geometry processing needed. There is more then just geometry to worry about in other words.

What do you think is holding Dues Ex back on nV hardware? I say nV hardware because its all their hardware even their high end, so where is the problem? It ain't ram, it ain't shader performance, it ain't geometry, so where does that leave the problem? The code.

Even with Vega's "new" or "modified Geometry pipeline, it only catches up to nV's current implementation albeit might give some flexibility as in adaptive tessellation of different tessellation levels on different objects.

 

[tangoseal]

Go read post no 1 by Kyle a year ago. Now consider the nuclear weapon called Zen that is about to be dropped on the market haha. AMD did something right I am speculating here.

 

[FrgMstr]

Go read post no 1 by Kyle a year ago. Now consider the nuclear weapon called Zen that is about to be dropped on the market haha. AMD did something right I am speculating here.

I did go back and reread it, and I am sure I never mentioned anything regarding CPU development in the first post or the editorial.

 

[FrgMstr]

Oh Kyle I was commenting on them as a whole. Not just the graphics division. What I am saying is that many people were speculating whether or not they would even be around still. Yet they are about to drop a potentially really good processor line which will give them a much needed boost. All guessing. its just funny how things can change over time.

And people are saying that Ryzen is going to fail ..... lets see what happens. Yes I know my comments are not directly related nor inline with your original editorial or post.

All good brother....I want AMD to succeed probably more than anyone else here.

 

[bezant]

Whatever happened with the intel amd license deal?

 

[SighTurtle]

If this deal was intended for Apple, is it possible for both Intel and AMD to be mum on it until Apple announces it first?

Edit: From a financial reporting standpoint. AMD has said nothing about any potential deal recently, nor Intel to my knowledge.

 

[Deleted member 93354]

Whatever happened with the intel amd license deal?

The rumor is Intel is going to create an i series chip which does not have onboard graphics and will offload the graphics onto an external chip with AMD designs. Intels focus of late really hasn't been to improve the speed or IPC of their chips, but increase the transistor count for 3D graphics and video, which to date has had abysmal performance. By removing all those transistors, intel can focus more power and silicon real estate to bringing the processor part up in performance.

I think Intel is finally realizing that anybody who needs 3D graphics is going to go 3rd party (NVIDIA/AMD) and that Intel has to actually offer improvement in speed now that AMD is offering viable alternatives CPU wise. After all a $100 card can outdo what Intel offers.

 

[Shintai]

The rumor is Intel is going to create an i series chip which does not have onboard graphics and will offload the graphics onto an external chip with AMD designs. Intels focus of late really hasn't been to improve the speed or IPC of their chips, but increase the transistor count for 3D graphics and video, which to date has had abysmal performance. By removing all those transistors, intel can focus more power and silicon real estate to bringing the processor part up in performance.

I think Intel is finally realizing that anybody who needs 3D graphics is going to go 3rd party (NVIDIA/AMD) and that Intel has to actually offer improvement in speed now that AMD is offering viable alternatives CPU wise. After all a $100 card can outdo what Intel offers.

You are not getting any Intel mainstream CPUs without IGP. Its simply not worth making a new mask.

Secondly, bandwidth. Are you going to pair very expensive HBM2 memory with lowend GPU to make up for it? Or add external GDDR5 etc? Ye that's not happening.

 

[Deleted member 93354]

You are not getting any Intel mainstream CPUs without IGP. Its simply not worth making a new mask.

Secondly, bandwidth. Are you going to pair very expensive HBM2 memory with lowend GPU to make up for it? Or add external GDDR5 etc? Ye that's not happening.

What are you smoking? Seriously. Look at the performance charts. A RX460 can beat & DUST a intel enabled graphics chip.
http://gpu.userbenchmark.com/Compare/Intel-HD-530-Desktop-Skylake-vs-AMD-RX-460/m33102vs3641

 

[Shintai]

What are you smoking? Seriously. Look at the performance charts. A RX460 can beat & DUST a intel enabled graphics chip.
http://gpu.userbenchmark.com/Compare/Intel-HD-530-Desktop-Skylake-vs-AMD-RX-460/m33102vs3641

You only prove my case. That RX460 uses 112GB/sec memory and 75W. And sales wise? People prefer RX480 that sells over twice as many cards.

Do you see the issue?

 

[Deleted member 93354]

You only prove my case. That RX460 uses 112GB/sec memory and 75W. And sales wise? People prefer RX480 that sells over twice as many cards.

Do you see the issue?

And you are missing my point. $100 gets you infinitely better performance. (well at least 3x's) intel graphics just aren't good for anything beyond the most basic 3D routines and 2D. Killing that worthless side of the chip will give them much more flexibility in the power envelope space.

 

[ManofGod]

You only prove my case. That RX460 uses 112GB/sec memory and 75W. And sales wise? People prefer RX480 that sells over twice as many cards.

Do you see the issue?

I did not realize just how peaceful my AMD computing days have become since I have seen none of your comments, thanks for the reminder. I would say later but, I certainly hope not.

 

[lolfail9001]

Killing that worthless side of the chip will give them much more flexibility in the power envelope space.

Not really.

 

[OrangeKhrush]

It is actually a good question, Nvidia agreement is expired and if Intel don't have any deal with AMD they could be in for a world of hurt if they so dream of putting anything that resembles graphics processing units on their platform.

The Apple rumour makes not sense if AMD can build their own APU's to sell to Apple, why let Intel in on that pie.

 

[Deleted member 93354]

Not really.

Yeah and you don't know jack schitt about processors and power envelopes. Anybody knows when you start maxing out one side of the processor, the other side throttles to keep it in the heat/power spec.

Intel used to have the GMA and Extreme Graphics series chips that were nothing short of horrendous. Putting it directly on the die did improve latency issues and improved memory access times. But too much of the chips budget was for the CPU. So Intel's hands were really hamstrung for creating enough of a powerful GPU side. And even Iris-Pro couldn't solve the bandwidth issues. Now Intel could do away with the GPU side and make it a straight CPU. But that's a lot of real estate they just chopped off. You could add more cores, but Intel is loathe to do this for the previously stated reasons.

 

[lolfail9001]

Anybody knows when you start maxing out one side of the processor, the other side throttles to keep it in the heat/power spec.

Precisely, and when that spec is 15W, adding a dGPU that has memory alone consuming that much power is the worst idea ever.

 

[Kamber]

Good read and reporting.

 

[whateverer]

Yeah, it's been half a year since this "sure thing" fresh shit was first aired, and I still see nothing but dried-up shit.

The only reason Intel licensed Nvidia's patents was part of case settlement payout, and I'm sure that settlement included no more suits from Nvidia (or they would have continued the license deal).

If Intel doesn't fear other GPU vendor's patent portfolios, then I don't see them wasting the money. Everyone assumes they're in for a "world of hurt," but how many GPU companies aside from Nvidia went on the patent warpath recently?

And feeding something significantly more powerful than Intel's IGP would raise the costs.

The entire purpose of keeping integrated graphics low-cost is reusing the memory bandwidth they already provide for the CPU. You don't save any die space by integrating the CPU and GPU onto the same die, and if you need the extra bandwidth of dedicated vram, you might as well just make it an external GPU.

 

[Shintai]

Apple is also developing its own GPU, yet no patent licensing.

 

[SighTurtle]

Yeah until Intel or AMD say something, it can't be confirmed.

 

[razor1]

yeah for the most part that article reads as if they are using AMD's patents instead of nV's and that does nothing for Intel lol. Intel never needed nV's patent to begin with, they were forced to pay nV.