AMD Radeon RX Vega 64 and 56 Reviews [*UPDATED* Aug 28]

[antihelten]

There is an element of not biting the hand that feeds you, that will keep a lot of retailers quiet, and that apparently only select retailers were offered the rebate deal. I bet OCUK won't be in that select list next time. I do remember one other saying they were offered the rebate deal and passed, presumably because they could make more money just selling the card at what the market would bear.

I would be extremely surprised if OCUK gets put on any kind of blacklist by AMD, given their size in the UK market, but other than I agree that smaller retailers might not want to rock the boat.

Also there has been no other retailers who has said publicly that they were offered the rebate, however OC3D mentioned in his video that Scan.co.uk said to him that they were offered the rebate. However I have a bit of a hard time with these second hand accounts, since the whole issue to me seems to based on misunderstandings and semantics, and with those kind of issues such second hand accounts can often tend to distort and exaggerate things quite a bit. Especially when the second hand accounts are at their core made for clicks.

In the end it's a small part of the mess that is Vega.

Vega's ultimate problem is that it is a poor technical solution, that is 15 months late to market. All the mistakes flow from that. The bad marketing to try and put lipstick on a pig, the Bundles to try to boost weak margins a expensive to produce cards, the alleged rebates to paint a temporary appearance of value...

A new video on the high cost of HBM and why AMD was stuck using it:
https://www.youtube.com/watch?v=p9ih5vmcDEk&t=914s

You will find no disagreement from me on these points.

 

[tential]

Can you expand on this?

AMD has claimed differently and I hope you at least have some justification in your statement except (this is what I believe, so it must be this way). It's one thing to be negative, but at least contribute to the discussion as to why, otherwise it's just threadcrapping.

Are you not doing the same by not providing any justification as to why primiative shaders will deliver large boosters? Otehrwise you're doing exactly what you're accusing him of. Practice what you preach.
Or are you suggesting that we should take what companies say as gospel?

 

[Kenmitch]

Are you not doing the same by not providing any justification as to why primiative shaders will deliver large boosters? Otherwise you're doing exactly what you're accusing him of. Practice what you preach.
Or are you suggesting that we should take what companies say as gospel?

How is it the same? He at least stated a source.

I'd take AMD s word over what looks to be a AMD haters....Not to the gospel level.

 

[Paul98]

http://radeon.com/_downloads/vega-whitepaper-11.6.17.pdf

"The “Vega” 10 GPU includes four geometry engines which would normally be limited to a maximum throughput of four primitives per clock, but this limit increases to more than 17 primitives per clock when primitive shaders are employed.⁷"

"Primitive shaders have many potential uses beyond high-performance geometry culling. Shadow-map rendering is another ubiquitous process in modern engines that could benefit greatly from the reduced processing overhead of primitive shaders. We can envision even more uses for this technology in the future, including deferred vertex attribute computation, multi-view/multi-resolution rendering, depth pre-passes, particle systems, and full-scene graph processing and traversal on the GPU. Primitive shaders will coexist with the standard hardware geometry pipeline rather than replacing it. In keeping with “Vega’s” new cache hierarchy, the geometry engine can now use the on-chip L2 cache to store vertex parameter data."

As you can see when we start getting primitive shaders enabled there will be a lot of uses which will increase performance

 

[Janooo]

http://radeon.com/_downloads/vega-whitepaper-11.6.17.pdf

"The “Vega” 10 GPU includes four geometry engines which would normally be limited to a maximum throughput of four primitives per clock, but this limit increases to more than 17 primitives per clock when primitive shaders are employed.⁷"

"Primitive shaders have many potential uses beyond high-performance geometry culling. Shadow-map rendering is another ubiquitous process in modern engines that could benefit greatly from the reduced processing overhead of primitive shaders. We can envision even more uses for this technology in the future, including deferred vertex attribute computation, multi-view/multi-resolution rendering, depth pre-passes, particle systems, and full-scene graph processing and traversal on the GPU. Primitive shaders will coexist with the standard hardware geometry pipeline rather than replacing it. In keeping with “Vega’s” new cache hierarchy, the geometry engine can now use the on-chip L2 cache to store vertex parameter data."

As you can see when we start getting primitive shaders enabled there will be a lot of uses which will increase performance

It appears to me this is the reason why Vega will be regarded as a revolutionary GPU when all the dust settles.
It takes time for the SW to follow HW.

 

[Muhammed]

Primitive Shaders are not revolutionary at all, NVIDIA already uses similar shaders which accomplish the same function but with different name. If you think that 17 primitives per clock is something special, then you're wrong, it is not.

 

[Konan]

No one is going to do extra work unless AMD are going to incentive them for something, which isn't happening.

The new programmable geometry pipeline on Vega will offer up to 2x the peak throughput per clock compared to previous generations by utilizing a new primitive shader. This new shader combines the functions of vertex and geometry shader and, as AMD told it to me, with the right knowledge you can discard game based primitives at an incredible rate. This right knowledge though is the crucial component, it is something that has to be coded for directly and isn't something that AMD or Vega will be able to do behind the scenes.

 

[maddie]

Are you not doing the same by not providing any justification as to why primiative shaders will deliver large boosters? Otehrwise you're doing exactly what you're accusing him of. Practice what you preach.
Or are you suggesting that we should take what companies say as gospel?

I'm asking for some justification to his claim since AMD states otherwise. I never said that they are better, although I personally believe it to be true, just there is no reasoning as to why he states they are negligible.

If I accuse a company of making false claims, the burden is on me to justify the belief.

See the difference?

 

[maddie]

Primitive Shaders are not revolutionary at all, NVIDIA already uses similar shaders which accomplish the same function but with different name. If you think that 17 primitives per clock is something special, then you're wrong, it is not.

Again, an outright statement with no reasons or explanations. Are we to view you as a god, passing down the gospel truths?

 

[itsmydamnation]

Primitive Shaders are not revolutionary at all, NVIDIA already uses similar shaders which accomplish the same function but with different name. If you think that 17 primitives per clock is something special, then you're wrong, it is not.

Care to detail what NV do then?

There is no doupt that right now NV is extracting more game performance per flop then AMD compared to what Compute workloads look like. No one knows exactly how much of Geometry setup is compute based or FF based on either GPU.

So simple question how many primitives can Pascal(say GP104 or 102) process a clock?

Again, an outright statement with no reasons or explanations. Are we to view you as a god, passing down the gospel truths?

I would assume a fish out of water......

 

[krumme]

Primitive Shaders are not revolutionary at all, NVIDIA already uses similar shaders which accomplish the same function but with different name. If you think that 17 primitives per clock is something special, then you're wrong, it is not.

1-2% you know

 

[Konan]

http://radeon.com/_downloads/vega-whitepaper-11.6.17.pdf

"The “Vega” 10 GPU includes four geometry engines which would normally be limited to a maximum throughput of four primitives per clock, but this limit increases to more than 17 primitives per clock when primitive shaders are employed.⁷"

"Primitive shaders have many potential uses beyond high-performance geometry culling. Shadow-map rendering is another ubiquitous process in modern engines that could benefit greatly from the reduced processing overhead of primitive shaders. We can envision even more uses for this technology in the future, including deferred vertex attribute computation, multi-view/multi-resolution rendering, depth pre-passes, particle systems, and full-scene graph processing and traversal on the GPU. Primitive shaders will coexist with the standard hardware geometry pipeline rather than replacing it. In keeping with “Vega’s” new cache hierarchy, the geometry engine can now use the on-chip L2 cache to store vertex parameter data."

As you can see when we start getting primitive shaders enabled there will be a lot of uses which will increase performance

⁷ Peak discard rate: Data based on AMD Internal testing of a prototype RX Vega sample with a prototype branch of driver 17.320. Results may vary for nal product, and performance may vary based on use of latest available drivers

Whats going on here?

This whitepaper contradicts earlier information.
http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2

And while AMD's presentation and comments itself don't go into detail on how they achieved this increase in throughput, buried in the footnote for AMD's slide deck is this nugget: "Vega is designed to handle up to 11 polygons per clock with 4 geometry engines."

This is from the old locked thread -
http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...ega-64-unboxing.2508928/page-93#post-39030271
@JDG1980

That "11 triangles per clock" figure is basically a lie from RTG's notoriously dishonest marketing department. It appeared only in the old architecture preview slides, and was part of a footnote claiming "over 2x peak throughput per clock" comparing to Fiji, which had no Primitive Discard Accelerator (introduced in Polaris). Note the careful use of the term throughput, not rendering. What they really mean is that it can render 4 triangles/clock and discard another 7 via the Primitive Discard Accelerator. There is NO WAY that Vega can actually render more than 4 triangles/clock no matter what the programmer does.

Note that this tendentious claim does NOT appear on any of the release day slide decks. Here is the release day architecture overview slides - please try to find anything backing up the 11 polygons/clock claim. It's not there, because that claim was always dishonest marketing BS. The emphasis in Vega slides is trying to reduce the amount of geometry the card actually has to render. When these tricks run out, Vega is bottlenecked. It WILL NEVER and CAN NEVER match GP102 in geometry-intensive games, because GP102 can actually render 6 triangles/clock and Vega can only actually render 4.

Either way, as @Ryan Smith said"

AMD isn’t offering any real detail here in how the primitive shader operates, and as a result I’m curious here whether this is something that AMD’s shader compiler can automatically add, or if it requires developers to specifically call it (like they would vertex and geometry shaders).

Looks like the "are employed" relates to additional developer work, and not a "magical switch" type update.

 

[Paul98]

Whats going on here?

This whitepaper contradicts earlier information.
http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2


This is from the old locked thread -
http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...ega-64-unboxing.2508928/page-93#post-39030271
@JDG1980


Either way, as @Ryan Smith said"


Looks like the "are employed" relates to additional developer work, and not a "magical switch" type update.

Yeah that is newer information than what was released back then. The prim shaders are handled in the driver, and don't require extra developer work. But that currently isn't active in the drivers. There is still a good bit of driver work to do to get everything up and running it seems.

 

[Konan]

The prim shaders are handled in the driver, and don't require extra developer work. But that currently isn't active in the drivers.

The new programmable geometry pipeline on Vega will offer up to 2x the peak throughput per clock compared to previous generations by utilizing a new primitive shader. This new shader combines the functions of vertex and geometry shader and, as AMD told it to me, with the right knowledge you can discard game based primitives at an incredible rate. This right knowledge though is the crucial component, it is something that has to be coded for directly and isn't something that AMD or Vega will be able to do behind the scenes.

Programmable defers to additional dev work. Needs to be coded for, therefore, no magical driver will make magic happen around Prim Shaders unless someone puts time and effort into it outside of AMD. It isn't something behind the scenes that can be flicked on for immediate effort unless the work goes into it.

 

[Muhammed]

Ryan Smith:

AMD is still trying to figure out how to expose the feature to developers in a sensible way. Even more so than DX12, I get the impression that it's very guru-y. One of AMD's engineers compared it to doing inline assembly. You have to be able to outsmart the driver (and the driver needs to be taking a less than highly efficient path) to gain anything from manual control.

So unless the developer codes for it, gains and even activation methods of that feature are freaking hard.

That 17 primitives per clock is nothing to sneeze at, it's a max value, under very specific conditions. For example: NV's GP102 can do 15 primitives clock, as each SM/PolyMorph engine can do 0.5 primitive clock since the days of GTX 680. TitanXp has 30 SMs, so a max peak is 15 primitives per clock. But in practice it never reaches that number, It's typically 6 triangles per clock. Vega/Fiji are at 4 triangles per clock.

Primitive Shaders are just that: programmable shaders just like any other shader, they are programmable because they combine Vertex Shader with Geometry Shader. Any GPU can do these shaders, because any GPU is programmable. That's what a Unified shader core means.

A shader becomes hardware when the GPU is aware of it's inputs and outputs, when the GPU knows when to connect a data source to the right stage, and how to connect a data "result" to the next stage. In other words, the GPU is hardwired to move the data of that shader around and balance it in a specific way. Examples are shaders for Vertex, Pixel, Hull and Geometry.

A shader is software when the developer tells the GPU exactly what do, the GPU is not hardwired in this case, but is softwired, the developer controls the data flow, telling the GPU how to connect the data and how to load-balance it, and the GPU obeys, both methods achieve the same goal. One is easier to program "automatized" and one is not, but both reach the same end.

So Primitive Shaders are just that, some software shaders that reconfigure already present hardware shaders (Vertex and Geometry), nothing revelatory about that, they can be done on any modern GPU with unified shaders. Vega even has another type of software shader called "Surface Shader" which combines Vertex shader with Hull shader.

Any vendor whether AMD or NVIDIA can do similar types of software shaders to control geometry data, how do you think NVIDIA process Tessellation very efficiently? How do you think they cull triangles and discard them at very high rates? They are using programmable shaders to do that, AMD just came in late to the party, made one and called it a new name, and put it on marketing slides to populate their white paper with new stuff. If primitive shaders were so unique and earth shattering, AMD would have got it out of the door before anything else, they would have made it priority one, after all they are in a desperate need for a clean win with Vega. At least they would have delayed Vega until it was ready. But no, they had Vega for too much time to release it without something seemingly important as primitive shaders.

Conclusion: Prmitive shaders are useful, but they are not revolutionary and will not change current status quo, at driver level they are very hard to activate, and their usefulness is questionable. If games developers worked with these shaders and optimized for it we might see some gains, but it requires effort and that remains a big if.

 

[Arachnotronic]

How is it the same? He at least stated a source.

I'd take AMD s word over what looks to be a AMD haters....Not to the gospel level.

AMD's word would have us believing that Vega is a huge perf/watt increase over Polaris (it's not), and that Vega compares favorably to 1080 Ti/Titan Xp (which it doesn't).

 

[ub4ty]

Ryan Smith:

So unless the developer codes for it, gains and even activation methods of that feature are freaking hard.

That 17 primitives per clock is nothing to sneeze at, it's a max value, under very specific conditions. For example: NV's GP102 can do 15 primitives clock, as each SM/PolyMorph engine can do 0.5 primitive clock since the days of GTX 680. TitanXp has 30 SMs, so a max peak is 15 primitives per clock. But in practice it never reaches that number, It's typically 6 triangles per clock. Vega/Fiji are at 4 triangles per clock.

Primitive Shaders are just that: programmable shaders just like any other shader, they are programmable because they combine Vertex Shader with Geometry Shader. Any GPU can do these shaders, because any GPU is programmable. That's what a Unified shader core means.

A shader becomes hardware when the GPU is aware of it's inputs and outputs, when the GPU knows when to connect a data source to the right stage, and how to connect a data "result" to the next stage. In other words, the GPU is hardwired to move the data of that shader around and balance it in a specific way. Examples are shaders for Vertex, Pixel, Hull and Geometry.

A shader is software when the developer tells the GPU exactly what do, the GPU is not hardwired in this case, but is softwired, the developer controls the data flow, telling the GPU how to connect the data and how to load-balance it, and the GPU obeys, both methods achieve the same goal. One is easier to program "automatized" and one is not, but both reach the same end.

So Primitive Shaders are just that, some software shaders that reconfigure already present hardware shaders (Vertex and Geometry), nothing revelatory about that, they can be done on any modern GPU with unified shaders. Vega even has another type of software shader called "Surface Shader" which combines Vertex shader with Hull shader.

Any vendor whether AMD or NVIDIA can do similar types of software shaders to control geometry data, how do you think NVIDIA process Tessellation very efficiently? How do you think they cull triangles and discard them at very high rates? They are using programmable shaders to do that, AMD just came in late to the party, made one and called it a new name, and put it on marketing slides to populate their white paper with new stuff. If primitive shaders were so unique and earth shattering, AMD would have got it out of the door before anything else, they would have made it priority one, after all they are in a desperate need for a clean win with Vega. At least they would have delayed Vega until it was ready. But no, they had Vega for too much time to release it without something seemingly important as primitive shaders.

Conclusion: Prmitive shaders are useful, but they are not revolutionary and will not change current status quo, at driver level they are very hard to activate, and their usefulness is questionable. If games developers worked with these shaders and optimized for it we might see some gains, but it requires effort and that remains a big if.

Wow.. what an eye opener some of these pages have been. I just got done explaining in another thread how HBCC is similar overhyped marketing rewording of dynamic memory paging that all modern GPUs have. A feature for which Nvidia dominates Radeon on w.r.t to latency. Yet, here Radeon is again pulling the same crap w/ another feature. Sigh... You can't keep pulling this crap on your customers (potential customer) in my case.

The funny thing is, they seem to have their pro market potentially under control and are actually delivering on features there? Are RX vega consumer cards literally throw away pro architectural components? That would explain why Vega uses so much energy yet lacks performance and features. It's like they took a pro level card and disabled/took random things off, slapped a traditional GPU pipeline from fiji on and went from there.

 

[Elixer]

The funny thing is, they seem to have their pro market potentially under control and are actually delivering on features there? Are RX vega consumer cards literally throw away pro architectural components?

AMD's entire Vega line is the same chip, only thing that changes is BIOS & drivers.

 

[ub4ty]

AMD's entire Vega line is the same chip, only thing that changes is BIOS & drivers.

Welp, that explains it all. They had low sales in consumer GPUs thus went after the pro/workstation compute market hard with a completely revised architecture. Later they slapped a fiji like traditional GPU pipeline and went from the there. This explains why the marketing fell off and why they never formally did a big launch of the card. Instead they treated it as a bastard child and let it just roll out the gate. They went after the pro market completely at siggraph and most of the demos were on workstation/server grade GPU cards. They showed off vega as a footnote..

Explains the power inefficiencies and lackluster GPU performance.
Explains why they have all of these new features but most are disabled or they haven't figured out to integrate them properly with drivers. What really is a head scratcher is FP16 which is also another half working/disabled feature. FP16 is more prominent in compute and in order to get 2x FP32 performance you need a considerable amount more pipeline hardware which would lead to more power usage. What's a head scratcher is that gaming is going towards more detail and higher precision thus where does FP16 fit in gaming. It's why you don't see this being a prominent feature on Nvidia consumer gaming cards. But here comes Radeon reusing a pro-architecture feature for their gaming architecture and therein lies the insane power vs. performance dilemma. It's why FE is even worse. They're literally writing hand me down drivers and formulating the traditional GPU pipeline as they go. They're also trying to cleverly gimp the compute/pro features that are on the die.

HBCC is for Radeon SSG that has flash storage on the GPU. You need this for huge datasets and you prefer it to be on the GPU to avoid the insane costs of PCIE communication. For Vega RX, they stripped the Flash storage on the GPU but still have this huge HBCC cache controller sucking down gobs of power w/ nothing to cache. They then decide to do the cache protocol over the PCIE bus to the main memory they were trying to avoid and call it a feature and market it.

What's going to be seemingly more ridiculous is that its more than likely that they're going to gimp the hell out of FP16 on the consumer cards because they don't want it being available for compute and compete against their pro cards. It's why they've been so quiet about this feature as well as others. It's why they can do demos but you can't use it on your retail cards. It works they likely are just trying to work out how to gimp it or make it viable for the traditional GPU pipeline. It's more than likely that they're literally trying to come up w/ ways to gimp/disable certain aspects/features in drivers and software so as to hose down the compute capability that is sucking up soo much power.

All together, I can see the ugliness of this architecture beyond what it was meant for : Compute.
Compare that to Nvidia that has different die spins per use and you see where the power/performance gap is filled vs. 1070/1080/1080ti.

Now, I can perfectly understand that Radeon is nowhere near being on Nvidia's budget thus why they can't spin different dies but at least come out w/ it man. They likely spun some good pro/workstation cards. So, why not level and be honest w/ your die hard GPU consumers? Why the bait and switch and b.s? People are going to find out eventually and now you set yourself up for disappointing expectations.

I'm holding out for 24 more hours to investigate the likelihood of them gimping anything that is of value on this card. Based on what I find, I'm not even going to bother evaluating this card for compute.

 

[Elixer]

What's going to be seemingly more ridiculous is that its more than likely that they're going to gimp the hell out of FP16 on the consumer cards because they don't want it being available for compute and compete against their pro cards. It's why they've been so quiet about this feature as well as others.

That is what nvidia has been doing...so, I guess you won't be buying any of their cards.

When they finally ship the Rocm drivers, it should be fully enabled (as it is in Vulkan/DX12), I don't see AMD doing what nvidia has been doing in compute, they want as many people as possible to use Vega for compute.

 

[Bouowmx]

For HBC on the Pro SSG, isn't connection between GPU and on-board storage still PCI Express? But that the GPU, not the CPU, handles the interconnect instead.

 

[ub4ty]

That is what nvidia has been doing...so, I guess you won't be buying any of their cards.

When they finally ship the Rocm drivers, it should be fully enabled (as it is in Vulkan/DX12), I don't see AMD doing what nvidia has been doing in compute, they want as many people as possible to use Vega for compute.

Indeed it has been and what AMD does for the same features.
Thus, I negate either of them when it comes to this metric and instead look at others.

I don't see them enabling it for general purpose compute which is for OpenCL not vulkan/DX12. AMD already disables a slew of features on their consumer cards just like Nvidia thus there's no reason to believe this will change. I have a good feeling that HBCC and FP16 will be restricted to the gaming pipeline so as to ensure it doesn't compete w/ their workstation cards because they do this very thing right now for other features. Another thing is the beta state of their compute stack and Rocm. They targetted compute with this architecture clearly which is why they are heavily pushing their workstation/pro/server cards and not vega consumer. Vega consumer was an afterthought of a compute architecture. Now the issue because, how much stuff are they going to disable/gimp for compute on vega RX as they dont want it competing with their other cards. As they have made no mention of it (which they would if they hoped to sell more cards), it's likely that they will be gimping it and what wiggle room for the technical gimping they plan to implement..

So, FP16 .. yeah amazing feature for gaming and x,y,z gaming....
Q : Will it be available for general purpose compute?
A : No.. But if you want to use compute it is slightly enabled on Vega FE and fully on pro/workstation

 

[ub4ty]

For HBC on the Pro SSG, isn't connection between GPU and on-board storage still PCI Express? But that the GPU, not the CPU, handles the interconnect instead.

The flash is directly attached to the board and its likely connected via a lower latency mechanism through HBCC than just straight out PCI-E. Infinity fabric is moldable in this way.. They use general Phy and can run various protocols over it. So, they can run PCI-E or some other proprietary protocol. So, to answer your question : the latency is likely much lower than piping across PCI-E to the CPU domain. They haven't released a white paper going into detail but its clear that HBCC is a catch-all engine for multiple cache levels the main being onboard the GPU w/ the least being PCI-E to sys memory.

So, imagine them taking this from the Pro cards, not putting flash storage on the GPU board, disabling the interconnects but keeping this yuge HBCC to access sys mem. This is where the huge die comes from and the insane power usage... To cut costs, they're using a huge die meant for their workstation/pro cards including subsystems for features on those cards and they're only gimping those subsystems. Thus you get huge power util. and avg. performance as this wasn't a purpose built die for gaming. This is the same way you see FE being cut but w/ a less gimped compute die vs RX vega.

Not enough white papers out and no detailing of what's under the shroud of their workstation/pro cards but it's literally the same die w/ actual functional modules.

They could literally be taking the underperform-ant or error'd pro dies and using them in RX vega which explains the shortage. Same concept as CPUs and what GPUs do but stretched across a far wider spectrum of cards.

It's be like Nvidia taking a big arse p100 and using it in consumer graphics cards but w/ all the major pro features gimped/disabled :

See power util :
https://images.nvidia.com/content/tesla/pdf/nvidia-tesla-p100-PCIe-datasheet.pdf
250Watts

So, this is literally what's going on here. In this case, as i've stated, the only question is how hard they're going to gimp/disable certain features and which are actually just hardware disabled due to defective die regions.

Edit : Validation source sauce :

The issue being that this is a hand me down feature not fully exploited on vega.. Same with FP16. It's not even fully implemented for GPU land and its questionable as to whether you will have full access to CPU land. Reasoning : Otherwise, you'd have a pro level card. Both AMD/Nvidia pull this crap on varying levels but at least Nvidia has the budget to cut new dies without any of this stuff so as to have better power util. All of this is just on your RX vega which is why they suck down so much power.

 

[iBoMbY]

Primitive Shaders are just that: programmable shaders just like any other shader, they are programmable because they combine Vertex Shader with Geometry Shader. Any GPU can do these shaders, because any GPU is programmable. That's what a Unified shader core means.

You are missing the point. Yes, you have all these shaders on previous GPUs. The difference is, they are all tied to a specific pipeline stage, while the Primitive Shader can handle all stages in one shader, and can cull primitives earlier, to reduce the workload of following steps.

 

[Olikan]

You are missing the point. Yes, you have all these shaders on previous GPUs. The difference is, they are all tied to a specific pipeline stage, while the Primitive Shader can handle all stages in one shader, and can cull primitives earlier, to reduce the workload of following steps.

Found an interesting quote, at beyond 3d forums, it seems that PS4 pro, have similar feature....
http://www.gamasutra.com/view/feature/191007/inside_the_playstation_4_with_mark_.php?page=3

There are a broad variety of techniques we've come up with to reduce the vertex bottlenecks, in some cases they are enhancements to the hardware," said Cerny. "The most interesting of those is that you can use compute as a frontend for your graphics."

This technique, he said, is "a mix of hardware, firmware inside of the GPU, and compiler technology. What happens is you take your vertex shader, and you compile it twice, once as a compute shader, once as a vertex shader. The compute shader does a triangle sieve -- it just does the position computations from the original vertex shader and sees if the triangle is backfaced, or the like. And it's generating, on the fly, a reduced set of triangles for the vertex shader to use. This compute shader and the vertex shader are very, very tightly linked inside of the hardware

Looking at the diagram, it seems indeed that primitive shaders "compile twice" the geometry