AMD FidelityFX on Consoles

did Microsoft talk anything about this custom instructions ?
AFAIK there's only this slide that was mentioned in the previous page:

202008180228381.jpg
 
„Microsoft engineers talked further about the custom hardware support that they developed working with AMD to support machine learning applications.”


"We knew that many inference algorithms need only 8-bit and 4-bit integer positions for weights and the math operations involving those weights comprise the bulk of the performance overhead for those algorithms," says Andrew Goossen. "So we added special hardware support for this specific scenario. The result is that Series X offers 49 TOPS for 8-bit integer operations and 97 TOPS for 4-bit integer operations. Note that the weights are integers, so those are TOPS and not TFLOPs. The net result is that Series X offers unparalleled intelligence for machine learning."

https://www.resetera.com/threads/xb...led-intelligence-for-machine-learning.176829/
 
I assumed them (Microsoft) saying "we added special hardware support" would mean they actually added special hardware over the generic RDNA1 capability of running mixed precision dot4 INT8 / dot8 INT4.
 
I assumed them (Microsoft) saying "we added special hardware support" would mean they actually added special hardware over the generic RDNA1 capability of running mixed precision dot4 INT8 / dot8 INT4.
nah, just pr, they are very openly celebreting all xsx/rdna2 features so if they have some other special souce for ml they would talked about it for months now
 
Ok so MS is lying?
Did they lie when they said they would be using the power of the cloud special sauce to make the XBOne's CPU 4x as powerful?


To be honest, I don't think they lied. They're probably talking about special instructions for ML inference loads. How much of it is useful for games and how much is useful for predicted Azure loads is up to anyone's guess, though.

Or maybe they indeed "added" dot4 INT8 / dot8 INT4 support over what was present in the One and OneX. That way you could call it "adding".
 
they have int8/int4 support so "they add" it read used rdna2 cheap with all features ;) wouldn't called it lying

I am not surprised this comment comes from you. Always very quick to dismiss everything XSX/MS related in defend to ps5. TF/vrs now ml. And yet you always quote Cerny as he is the only source of truth.
 
I am not surprised this comment comes from you. Always very quick to dismiss everything XSX/MS related in defend to ps5. TF/vrs now ml. And yet you always quote Cerny as he is the only source of truth.
when did I quoted Cerny lmao, its your opinion I'm dismissive towards microsoft, I'm here more dismissive towards ps5 int8/4 capabilities simple because I don't think sony added it to console
 
I now believe that AMD custom is a lot more detailed than first thought.
Always thought it was mainly big blocks, media blocks, rops, front end, cache sizes, then specific customer customizations.

Now I think it's down to features also.
VRS, int 4/8 etc.

So it's not about removing features from RDNA2 it's about paying to have them included from list of available features.

So then you have to consider Sony and MS possible different reasoning.

Example reduced precision:
Sony
  • What use will it have in games
  • Cost of inclusion
  • Spend that money on other features (cache scrubbers etc)
  • Can be done without reduced precision for the odd situation where would be nice to have.

MS
  • Hardly any die footprint
  • Light Azure ML workloads when not running xcloud

Back when these was being designed DLSS wasn't a thing, etc.

Much like Nvidia it's up to MS to now make use of features that may be more unique.
ML texture de/compression, ML upscaling for example.
These are all things than can be done on all platforms but may have a decent performance benefit on xbox. And can shared across internal studios and also be included in playfab. Same for AVX libraries.
 
Show me where on the patent it says PS5.
Its a patent that shows tech that may or may not be in any peice of tech, today or in 10 years.
You have made the jump to say that its proof the PS 5 has VRS, which it in no way is. Every other single peice of information says the PS5 doesn't have it.
Sometimes you just have to admit you are wrong.

Again, happy for you to show me where Sony has said the PS 5 has hardware VRS. Lets see what you come.up with. Hopefully its better than the patent you showed.

There's no Sony patent showing unified L3 cache in the CPU.
https://patents.google.com/patent/US9892024B2/en

Here is the patent which had people claiming the PE5 has unified L3 Cache on the CPU.

You can google "patent shows PS5 unified L3 cache" to see where fanboys were running with it, and as I said it was quashed by the die shots.


You mean you couldn't find a single official statement from Nvidia proving your "We know DLSS uses INT4 and INT8" statement, which is why you linked to a page with zero mention of DLSS, upscaling or supersampling.
You are just being ridiculous at this point.
As if Nvidia is going to make an official statement about using Int8 and Int4 instructions.
They have int8 and int4 on their cores. They write technical articles like the one I showed you about how using lower precision Int8 and Int4 gives them a performance boost when doing inferance ML work, and you claim there is no evidence they use their own tech to do what their technical articles talks about.
SMFH.
I can't find an official statement from Rockstar that they used the 2nd CPU core on the PS4 Pro for RDR 2, so they must not have done so.

Yeah sure, just for "texture sampling" and "nothing to do with compute".
First paragraph of page 14 in the RDNA1 whitepaper:
When a series of GPUs all dont contain a particular feature, then it isn't a base feature. Not all RDNA1 cards had it. It isn't on all RDNA 1 cards. Not sure what part of that you don't understand.

Changing goalposts from "Microsoft changed the GPU to get INT8/4 on RDNA2" to "it was on RDNA1 but not all of them".
Here's an official statement from Microsoft.

"So we added special hardware support for this specific scenario. The result is that Series X offers 49 TOPS for 8-bit integer operations and 97 TOPS for 4-bit integer operations. Note that the weights are integers, so those are TOPS and not TFLOPs. The net result is that Series X offers unparalleled intelligence for machine learning."

So Microsoft ADDED the hardware for Int8 and Int4. That would rule out it being a base RDNA 2 feature or they wouldn't have said that. They never said "we added compute units to the GPU", or "we added Ray tracing hardware" did they?

Here it is, the trolling and flamebaiting.
Thank you for confirming my suspicion that you didn't create this thread to be clarified, but to post concern trolling on the PS5.
Bye then. This thread is useless.
Your input to this thread was useless. You challenged something and were shown to be wrong. Rather than say that was the case you just added more and more silly things to confound the issue.
To show how disingenuous you are, you claim that there is no proof Nvidia uses Int8 and int4 in DLSS because there is no official statement, all the while pushing the point that the PS 5 must have VRS and int8 and int4 abilities when Sony has not mentioned it nor confirmed it. Not only have they not confirmed it, but their own employess and devs have said it doesn't.

But I agree, let's leave it at that.
 
I am very curious about this. DLSS requires Machine learning.
FidelityFX is the equivalent of DLSS and supposedly coming to both platforms. But we know the PS5 doesnt have ML. So how is it supposed to have it?
Maybe? Isn't there a version of Control that had DLSS that was Cuda based, in between the DLSS1.0 and 2.0 versions. And it looked better than 1.0 and worse than 2.0. But nVidia said they have no plans to release it for non RTX cards, they've moved on to 2.0 and above.
 
Or maybe they indeed "added" dot4 INT8 / dot8 INT4 support over what was present in the One and OneX. That way you could call it "adding".

This is how I would take it. The context is comparing to the previous console...

Much the same way I took Cerny's statement about the geometry engine to be in comparison to the previous console, whereas many people seem to have assumed he meant some special sauce over AMD's existing geometry engine designs.
 
I assumed them (Microsoft) saying "we added special hardware support" would mean they actually added special hardware over the generic RDNA1 capability of running mixed precision dot4 INT8 / dot8 INT4.

RDNA 1 has dot product hardware? How do u use it? There is no mention of it in the RDNA 1 isa from what I can tell and the RDNA 2 isa lists added dot product instructions as a notable feature change for RDNA2.

https://developer.amd.com/wp-content/resources/RDNA2_Shader_ISA_November2020.pdf


• Dot product ALU operations added accelerate inferencing and deep-learning:
◦ V_DOT2_F32_F16 / V_DOT2C_F32_F16 ◦ V_DOT2_I32_I16 / V_DOT2_U32_U16
◦ V_DOT4_I32_I8 / V_DOT4C_I32_I8
◦ V_DOT4_U32_U8
◦ V_DOT8_I32_I4
◦ V_DOT8_U32_U4
 
Last edited:
So Microsoft ADDED the hardware for Int8 and Int4. That would rule out it being a base RDNA 2 feature or they wouldn't have said that. They never said "we added compute units to the GPU", or "we added Ray tracing hardware" did they?
Do we know how granular the availability of different blocks for a semi-custom ic is?

For instance, if we rephrase as: "We knew that many burger meals are served with fries and that salt comprise the bulk of the dehydration overhead for those meals," says Andrew Goossen. "So we added a soft drink for this specific scenario."

This does not necessarily imply that MS invented soft drinks for this specific purpose. Rather that it was useful and available, so they ordered it off the menu.
 
Last edited:
  • Like
Reactions: snc
Is there an ETA on this tech's implementation?

Nothing in the first post. I'd rather not search through 9 pages (including 3 pages of console warring) to find it.
 
No date, but if you want us to restart the war starting here we can. That way you don't have to read all of the old posts.
 
  • Like
Reactions: JPT
Back
Top