• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

AMD trademarks Infinity Cache for RDNA2, confirming its existence

LordOfChaos

Member

Alright, basic gist, makes the new shared L1 in RDNA more effective (higher hit rates) by dynamically clustering CUs together, to pool their cache lines, to minimize duplication of cache use (+effective capacity) and increase hit rate. This reduces pressure on LLC/L2.

This makes the L2 more effective, thereby reducing pressure on memory bandwidth. This kind of design is enabled by the change in RDNA 1, with a 128KB shared L1 for the CUs within a shader array.

I still think the 128KB is too small for so many CUs, so one of the key changes I expect to see in RDNA 2 is at least 2x L1 size. Combine with this dynamic L1 changes, would be much more potent cache system, both L1 & L2 effective capacity & hit rate up.






Even more tl;dr, moar cache = moar better, dynamic sharing possibly even betterer
 

theddub

Banned
XSX RNDA2 is more likely than PS5 to have Infinity Cache than PS5......because they are apparently on different "revisions" of RDNA2 with XSX on a more advanced revision of RDNA2. They both could have, they both could not. The information is out there, since removed but obvious with the info we DO have when Sony talks primitive shaders(which is an older name) and geometry engine and since have advanced and been consolidated with into "mesh shaders" in the newer revisions of RDNA2. And the current term in RDNA2(point).whatever they are on INTERNALLY. (2.2, 2.3, etc... who knows the internal number). Though with the smaller amount of CUs on console, it's probably not needed for efficient parallel use of the CU count on either console.
 
Last edited:

Tripolygon

Banned
Would be an advantage to XSX, PC RDNA2 gpu's, and ps5, compared to RDNA 1. We gotta see how it works first.


Edit: SantaC SantaC answered below.
Over...RDNA 1? GCN? Yeah. But Xbox is also using RDNA 2, it'll be a shared advantage.

Edit: If they get it. Not really sure yet if this applies to under Big Navi.
XSX RNDA2 is more likely than PS5 to have Infinity Cache than PS5......because they are apparently on different "revisions" of RDNA2 with XSX on a more advanced revision of RDNA2. They both could have, they both could not. The information is out there, since removed but obvious when Sony talks primitive shaders(which is an older name) have advanced and been consolidated with geometry engine into "mesh shaders" in the newer revisions of RDNA2.
Series X has a 5MB L2 cache. In total across the entire SoC, they stated there was 76MB cache.
202008180219321.jpg

Xbox-Series-x-CPU-and-GPU-breakdown.jpg
 
Last edited:

Faithless83

Banned

Alright, basic gist, makes the new shared L1 in RDNA more effective (higher hit rates) by dynamically clustering CUs together, to pool their cache lines, to minimize duplication of cache use (+effective capacity) and increase hit rate. This reduces pressure on LLC/L2.

This makes the L2 more effective, thereby reducing pressure on memory bandwidth. This kind of design is enabled by the change in RDNA 1, with a 128KB shared L1 for the CUs within a shader array.

I still think the 128KB is too small for so many CUs, so one of the key changes I expect to see in RDNA 2 is at least 2x L1 size. Combine with this dynamic L1 changes, would be much more potent cache system, both L1 & L2 effective capacity & hit rate up.






Even more tl;dr, moar cache = moar better, dynamic sharing possibly even betterer

Wouldn't they need faster I/O for such small L1 cache? I mean, wouldn't it make some CU's idle or "starving" for information?

I think this explanation is easy to understand:

 
XSX RNDA2 is more likely than PS5 to have Infinity Cache than PS5......because they are apparently on different "revisions" of RDNA2 with XSX on a more advanced revision of RDNA2. They both could have, they both could not. The information is out there, since removed but obvious with the info we DO have when Sony talks primitive shaders(which is an older name) and geometry engine and since have advanced and been consolidated with into "mesh shaders" in the newer revisions of RDNA2. And the current term in RDNA2(point).whatever they are on INTERNALLY. (2.2, 2.3, etc... who knows the internal number). Though with the smaller amount of CUs on console, it's probably not needed for efficient parallel use of the CU count on either console.

"More advanced version". Stop the bullshit.

Unless you are Mark Cerny, you cant know how advanced ps5 is.

It is not official yet.
 

DeepEnigma

Gold Member
XSX RNDA2 is more likely than PS5 to have Infinity Cache than PS5......because they are apparently on different "revisions" of RDNA2 with XSX on a more advanced revision of RDNA2. They both could have, they both could not. The information is out there, since removed but obvious with the info we DO have when Sony talks primitive shaders(which is an older name) and geometry engine and since have advanced and been consolidated with into "mesh shaders" in the newer revisions of RDNA2. And the current term in RDNA2(point).whatever they are on INTERNALLY. (2.2, 2.3, etc... who knows the internal number). Though with the smaller amount of CUs on console, it's probably not needed for efficient parallel use of the CU count on either console.

You all try wayyyy toooooo hard.
 

LordOfChaos

Member
Wouldn't they need faster I/O for such small L1 cache? I mean, wouldn't it make some CU's idle or "starving" for information?

I think this explanation is easy to understand:



They note 128KB L1 for up to 80CUs does seem small, and this is another area that may be doubled on big navi. The L2 being used more effectively in this way should also alleviate the IO bottleneck.

PS5 has cache scrubbers, not infinity cache.

i dont think either console has this.

I don't think either has it either but there's no reason cache scrubbers and infinity cache are mutually exclusive. Better cache dumping vs more effective CU level sharing.
 
Last edited:

Tripolygon

Banned
If IF is to come with doubled L2, and the 5700 generation had 4MB, I would take this as evidence the next gen consoles do not have it?
I don't know. But I would guess both consoles have a similar setup in terms of GPU cache but PS5 has extra cache for their IO setup. I would guess PS5 in total has more cache in the SoC than Series X because of that.
 
Last edited:
XSX RNDA2 is more likely than PS5 to have Infinity Cache than PS5......because they are apparently on different "revisions" of RDNA2 with XSX on a more advanced revision of RDNA2. They both could have, they both could not. The information is out there, since removed but obvious with the info we DO have when Sony talks primitive shaders(which is an older name) and geometry engine and since have advanced and been consolidated with into "mesh shaders" in the newer revisions of RDNA2. And the current term in RDNA2(point).whatever they are on INTERNALLY. (2.2, 2.3, etc... who knows the internal number). Though with the smaller amount of CUs on console, it's probably not needed for efficient parallel use of the CU count on either console.

UnnaturalGratefulAntelope-max-14mb.gif
 
The assumption of One X having a later revision of RDNA2 is somewhat at odds with its relatively low clock-speed. The key thing about RDNA2 is that it facilitates much higher clocking.
Larger GPU's have always been clocked lower than smaller GPU's. Xsx has a beefier GPU than ps5. Clockspeed isn't everything, just look at 2080 Ti, vs 20xx series.
 

Piku_Ringo

Banned
XSX RNDA2 is more likely than PS5 to have Infinity Cache than PS5......because they are apparently on different "revisions" of RDNA2 with XSX on a more advanced revision of RDNA2. They both could have, they both could not. The information is out there, since removed but obvious with the info we DO have when Sony talks primitive shaders(which is an older name) and geometry engine and since have advanced and been consolidated with into "mesh shaders" in the newer revisions of RDNA2. And the current term in RDNA2(point).whatever they are on INTERNALLY. (2.2, 2.3, etc... who knows the internal number). Though with the smaller amount of CUs on console, it's probably not needed for efficient parallel use of the CU count on either console.
If true, Lisa really did Cerny dirty on their deal. :pie_roffles:
 

LordOfChaos

Member
The assumption of One X having a later revision of RDNA2 is somewhat at odds with its relatively low clock-speed. The key thing about RDNA2 is that it facilitates much higher clocking.

Eh. It's bigger, which would take more power to clock higher. I don't buy the blueiswhatever PS5 is RDNA1 conspiracy theories, but a lower clock speed on a 44% larger GPU is not unusual to suggest anything about the SeX either.

Also higher clocking memory also taking more power, Sony having to engineer the whole variable clock power anticipation thing to clock this high, without exploding the power budget, etc.
 
Last edited:

j^aws

Member
t6T0a9i.jpg


I've labelled the XSX die shot of its APU from Hotchips above. The CPU has 8MB of L3 cache, and the GPU has 5 MB of L2 cache. Infinity Cache would show up as a large amount here if it wasn't disclosed, but can't be seen. Unless 5MB of L2 in the GPU is classed as Infinite Cache, which doesn't sound like much.

The XSX has a wide bus at 320bits, and wider than reported Navi21, and since Infinite Cache is to mitigate narrow buses like 256bit, it doesn't make sense for XSX.

On the other hand, the PS5 has a narrow 256bit bus, and we've yet to see its APU die, or more details. But Infinity Cache would make the die substantially larger and hotter at high clocks. Need more details.

EDIT: GPU cache label updated to be L2, not L3.
 
Last edited:

theddub

Banned
The PS5 is more likely to have Infinity Cache than the XSX. It's well known how close of a relationship that Sony has had in developing features for RDNA2. It's probably a more advanced version of RDNA2 than what the XSX has.

Yes, this is just an opinion like theddub's opinion above claiming the opposite. Using terms to give weight to an argument when the terminology of "mesh shaders" were well known before Cerny's talk, and information cannot be out there, and removed at the same time. It's illogical.

So much for claiming the PS5 was RDNA1, now people are claiming it's a lesser version of RDNA2. :rolleyes:
Mesh shaders are in the "latest" RDNA2 version, you'll hear more about it....again, on Oct 25th. Let me ask you, does PS5 have "mesh shaders" or the RDNA1 defined "primitive shaders"? They are both part RNDA2 ,XSX uses mesh shaders from RDNA2, ps5 doesn't.
 
Last edited:

wolffy71

Banned
Im getting numb to all this guessing, hoping, assuming, debunking, and figuring. Just tell us about the damn things already. Show the games because regardless of the shit in them, how they run is what matters.
 

Elog

Member
Mesh shaders are in the "latest" RDNA2 version, you'll hear more about it....again, on Oct 25th. Let me ask you, does PS5 have "mesh shaders" or the RDNA1 defined "primitive shaders"? They are both part RNDA2 ,XSX uses mesh shaders from RDNA2, ps5 doesn't.

It is hard to have these discussions due to the lack of knowledge.

Do you know what these <insert buzz word> shaders do? I bet you do not.

Shaders are not hard to do. It is very straight forward. The problem is that shaders are computationally heavy. This is the reason why these <insert buzz word> shaders have been introduced, and they do primarily one thing: they cull what to shade or in other words they decide through an algorithm what to shade and what not to shade. The concept is to not shade things that does not give a clear visual advantage or to put numbers on it (as an example): Decrease visual quality by 3% due to no shaders for a lot of areas of the screen and reducing the amount of computations with 50%. This is done by not shading objects that are obscured, objects that are not in the centre of the visual field, objects that are far away etc.

Or in other words: Primitives shade as well as mesh shaders, VRS etc. It is how to get to what not to shade that differs.

The true next-generational way of doing this is by letting the geometry be the determinant instead of the shader - or in other words that the geometry engine creates all priorities in itself. This is what Nanite in UE5 is all about. Listen to the UE5 demo video to understand more about Nanite.

This is also what the PS5 is doing. Cerny described how the geometry engine is the driver behind the shader pipeline as well. And when you have pixel sized geometry you need pixel level shaders or in other words basically primitives. Listen to the UE5 demo to understand more about this - they do not get into all the nitty gritty but listen to the specific words they use when talking about Nanite at 2:10 onwards.

Your point that the PS5 is somehow a generation back here could not be more wrong. It is the other way around. This way of handling the geometry and shader functions is what people have called 'RDNA3' features in the PS5. I do not like that either since it a meaningless buzz word unless you know what is meant by it.
 
Last edited:

Loxus

Member
Mesh shaders are in the "latest" RDNA2 version, you'll hear more about it....again, on Oct 25th. Let me ask you, does PS5 have "mesh shaders" or the RDNA1 defined "primitive shaders"? They are both part RNDA2 ,XSX uses mesh shaders from RDNA2, ps5 doesn't.
Geometry Engine
“A new block that will give developers more control over triangles, primitives and geometry culling.”
“Removing back faced or off-screen vertices and triangles.”
“More complex usage involves something called Primitive Shaders, which allow the game to synthesize geometry on the fly as it’s being rendered.”
Mark Cerny, Lead Architect.
Source:

Mesh Shaders meaning: a new type of shader that combines vertex and primitive processing.
Geometry Engine sounds a lot like the PS5 version of Mesh Shaders to me.
 

theddub

Banned
It is hard to have these discussions due to the lack of knowledge.

Do you know what these <insert buzz word> shaders do? I bet you do not.

Shaders are not hard to do. It is very straight forward. The problem is that shaders are computationally heavy. This is the reason why these <insert buzz word> shaders have been introduced, and they do primarily one thing: they cull what to shade or in other words they decide through an algorithm what to shade and what not to shade. The concept is to not shade things that does not give a clear visual advantage or to put numbers on it (as an example): Decrease visual quality by 3% due to no shaders for a lot of areas of the screen and reducing the amount of computations with 50%. This is done by not shading objects that are obscured, objects that are not in the centre of the visual field, objects that are far away etc.

Or in other words: Primitives shade as well as mesh shaders, VRS etc. It is how to get to what not to shade that differs.

The true next-generational way of doing this is by letting the geometry be the determinant instead of the shader - or in other words that the geometry engine creates all priorities in itself. This is what Nanite in UE5 is all about. Listen to the UE5 demo video to understand more about Nanite.

This is also what the PS5 is doing. Cerny described how the geometry engine is the driver behind the shader pipeline as well. And when you have pixel sized geometry you need pixel level shaders or in other words basically primitives. Listen to the UE5 demo to understand more about this - they do not get into all the nitty gritty but listen to the specific words they use when talking about Nanite at 2:10 onwards.

Your point that the PS5 is somehow a generation back here could not be more wrong. It is the other way around. This way of handling the geometry and shader functions is what people have called 'RDNA3' features in the PS5. I do not like that either since it a meaningless buzz word unless you know what is meant by it.
Yes, I understand that they provide similar functionality....that was my point. Primitive shader is RDNA1 tech, technically Vega, AMD PREVIOUS pipeline's way of achieving similar results, whereas mesh shaders are used in a revised consolidation of the rendering pipeline in RDNA2:

"RDNA(1) also introduces working primitive shaders. While the feature was present in the hardware of the Vega architecture, it was difficult to get a real-world performance boost from and thus AMD never enabled it. Primitive shaders in RDNA(1) are compiler-controlled.[8] "

https://en.wikipedia.org/wiki/RDNA_(microarchitecture) (RDNA 1)




"A Mesh shader is a new type of shader that combines vertex and primitive processing. VS, HS, DS, and GS shader stages are replaced with Amplification Shader and Mesh Shader. Roughly, Mesh shaders replace VS+GS or DS+GS shaders and Amplification shaders replace VS+HS. "

Explained for directX 12

In RDNA2 AMD has, like Nvidia in Turing, also implemented mesh shaders and consolidated and simplified the pipeline, replacing RDNA1 primitive shaders.

I don't even have to claim that the newer mesh shaders are superior than the former pipeline using primitive shaders(though it is), my point is that primitive shaders == RDNA1 method and mesh shaders (with geometry engine consolidation see slide above in post#10 "Mesh Shading Geometry Engine plus MS's description, changes in pipeline) == RNDA2.....ie....it's NEWER

From AMD's community:


"Prior to mesh shader, the GPU geometry pipeline hid the parallel nature of GPU hardware execution behind a simplified programming abstraction which only gave developers access to seemingly linear shader functions… Mesh shaders change this by making geometry processing behave more like compute shaders. Rather than a single function that shades one vertex or one primitive, mesh shaders operate across an entire compute thread group, with access to group shared memory and advanced compute features such as cross-lane wave intrinsics that provide even more fine grained control over actual hardware execution."

"This should remove substantial overhead from geometry processing and allow for the creation of more detailed objects and scenes. Microsoft’s blog post has additional details on how they expect this advance to simplify GPU programming. Geometry culling should also be more efficient — mesh shaders basically add a mid-step cull between the coarse (per-mesh) and fine (per-triangle) culling performed in the existing geometry pipeline. The one catch? Implementing mesh shaders would probably require a game to be written for DX12U hardware and above, meaning we won’t see those kinds of titles until the standard is established as the expected baseline.

These are features and capabilities we’ll see rolled out over time, but with ray tracing debuting on consoles and AMD GPUs for the PC market, we should start to see increased uptake of that feature, at least, in 2021 – 2022. If you don’t have a DX12U card, there’s no reason to worry — you’ll be fully compatible with games for quite some time yet."

 
Last edited:

Elog

Member
Golden video;



Fantastic video. Just shows how wrong it is to evaluate silicon based on theoretical peak numbers when so much is left on the table with the current utilisation numbers of CUs. His numbers are close to crazy in terms of output increase due to better work distribution and cache management.
 

Kumomeme

Member
Alright, basic gist, makes the new shared L1 in RDNA more effective (higher hit rates) by dynamically clustering CUs together, to pool their cache lines, to minimize duplication of cache use (+effective capacity) and increase hit rate. This reduces pressure on LLC/L2.

Hit Rate + 100
 

Elog

Member
[/URL]

In RDNA2 AMD has, like Nvidia in Turing, also implemented mesh shaders and consolidated and simplified the pipeline, replacing RDNA1 primitive shaders.

I don't even have to claim that the newer mesh shaders are superior than the former pipeline using primitive shaders(though it is), my point is that primitive shaders == RDNA1 method and mesh shaders (with geometry engine consolidation see slide above in post#10 "Mesh Shading Geometry Engine plus MS's description, changes in pipeline) == RNDA2.....ie....it's NEWER

From AMD's community:


"Prior to mesh shader, the GPU geometry pipeline hid the parallel nature of GPU hardware execution behind a simplified programming abstraction which only gave developers access to seemingly linear shader functions… Mesh shaders change this by making geometry processing behave more like compute shaders. Rather than a single function that shades one vertex or one primitive, mesh shaders operate across an entire compute thread group, with access to group shared memory and advanced compute features such as cross-lane wave intrinsics that provide even more fine grained control over actual hardware execution."

"This should remove substantial overhead from geometry processing and allow for the creation of more detailed objects and scenes. Microsoft’s blog post has additional details on how they expect this advance to simplify GPU programming. Geometry culling should also be more efficient — mesh shaders basically add a mid-step cull between the coarse (per-mesh) and fine (per-triangle) culling performed in the existing geometry pipeline. The one catch? Implementing mesh shaders would probably require a game to be written for DX12U hardware and above, meaning we won’t see those kinds of titles until the standard is established as the expected baseline.

These are features and capabilities we’ll see rolled out over time, but with ray tracing debuting on consoles and AMD GPUs for the PC market, we should start to see increased uptake of that feature, at least, in 2021 – 2022. If you don’t have a DX12U card, there’s no reason to worry — you’ll be fully compatible with games for quite some time yet."


You completely missed the point: (New geometry engine with priorities down to the pixel level + primitive function shaders) IS NOT EQUAL TO (RDNA1 primitive shaders)

You mix up a buzz word with a basic word used when performing geometry calculations: https://en.wikipedia.org/wiki/Geometric_primitive
 

Elog

Member
This rdna 3 talk popping up again. Smh. You can immediately tell these people get their info from YouTube. This type of misinformation is rampant on video game YouTube channels.

You do realise that I completely agree with your statement? Read what I wrote again. The use of RDNA3 is meaningless in itself.
 

theddub

Banned
You completely missed the point: (New geometry engine with priorities down to the pixel level + primitive function shaders) IS NOT EQUAL TO (RDNA1 primitive shaders)

You mix up a buzz word with a basic word used when performing geometry calculations: https://en.wikipedia.org/wiki/Geometric_primitive
Sigh... you completely missed my point, to which you responded to multiple times. The implication isnt that it's RDNA1 equivalent, it's that it's based on RDNA1 and modified just like you just implied.
 
Last edited:
I dont think infinty cache is in the consoles. This is a big navi PC feature.

TBF MS did say their system(s) are full RDNA2. So there's a possibility this is at least present in those systems, maybe?

I would be kinda surprised if it wasn't in PS5's GPU, too. MS and Sony'd know of Big Navi for a long time, why pass up adding a feature like this in their systems?

EDIT:

Elog Elog : You do realize that Series X has both Primitive and Mesh Shaders, as well as Geometry Engine...right?

T theddub K Kerlurk Look there's no reason to turn every tech thread into a Series X vs. PS5 debate. Why can't people just be satisfied at speculating if both next-gen consoles have this feature and focus more on the benefits of the feature in question in a platform-agnostic fashion?
 
Last edited:

oldergamer

Member
I wouldn't be surprised if xbox had this. Could it be related to shaders capability to daisy chain and pass work to other shaders without hitting cache?
 

Elog

Member
Elog Elog : You do realize that Series X has both Primitive and Mesh Shaders, as well as Geometry Engine...right?

Hardware geometry engines has been around for quite some time - I know. The existence is not the key point. The key point is how geometry engine is used in the graphical pipeline and what the output is from the hardware GE. This is what is new with the GE in PS5. Culling and other type of priorities are denominated and applied at the GE level which means that prioritisation calculation/determination does not need to be done in downstream steps of the pipeline.

And as you know almost all exotic names we know for GPUs is about how individual steps in the graphical pipeline calculates, determines and applies priorities in each step. The default model in the PS5 is that this is done at the GE level which results in significant efficiency gains in later steps. It is very similar to how the UE5 is described by Epic.

The caveat about this is that a software engine that is built around the old paradigm might perform worse. This will of course change over the next two years as the PS5 API is utilised in full but I can see a BC or cross-get title hurt on the PS5 because of this - we will see. However, letting the GE do this work seem like the clear path forward given that both PS5 and UE are thinking the same about this.

Sigh... you completely missed my point, to which you responded to multiple times. The implication isnt that it's RDNA1 equivalent, it's that it's based on RDNA1 and modified just like you just implied.

It is not based on RDNA1 - it is a meaningless statement for the masses to try to imply that the PS5 is using old technology despite being false. Both of the new consoles uses RDNA2.
 
Last edited:
Top Bottom