Xbox Machine Learning & DirectML EXPLAINED - No, Xbox Series X Doesn't Use Tensor Cores

Xbox Machine Learning & DirectML EXPLAINED - No, Xbox Series X Doesn't Use Tensor Cores

Next-generation xbox games will make heavy usage of Machine Leaning through a combination of Microsoft's DirectML API and the Xbox Series X / Xbox Series S ability to run ML instructions on the GPU with lower precision operations.

These 4 and 8 bit operations are part and parcel of the RDNA 2 spec, which Microsoft have used to great effect inside the Xbox Series X.

There is however, a misconception that the Xbox uses Tensor Cores, dedicated ML accelerators on the consoles GPU, when this isn't true. Instead, these operations are ran on the Xbox Series X general Compute Units, which are general purpose and designed to run game code in general.

The Xbox Series X also focuses on inference for its ML operations, which is running trained code.

SOURCES

AMAZON AFFILIATE LINKS

GreenManGaming Affiliate link -

SOCIAL MEDIA

[mehdi]

Tensor core is just a naming from Nvidia, but technically they are units that are suitable for doing operations on int4 and int8 (4 and 8 stands for the number of bits so they mean it of 4 bits or 8bits). So both Xbox series x|s and Nvidia units are similar in this, but the advantage of tensor cores is that these units are designed for doing 4x4 matrix multiplication and therefore they are faster to run Ai model in inference mode. For the training part . Ai models are training floating points weights then they get converted to int4 or int8 to be executed fast.

[InternetSupervillain]

"Ladies and Gentleman welcome". The sheer Britishness.

[mrstraiban]

Can't wait to see a successful implementation of a technique similar to DLSS on the new Xboxes.

[NizarElZarif]

Using lower precision operation is not a problem. For example, running an 8-bit operation on 32-bit core is usually just as fast running at a 32-bit operation. Running a 64-bit operation on a 32-bit core is also possible, but requires multiple intermediate steps, which can be about 10 times slower than 32-bit or hardware 64 bit.
However, modern cores can fuse multiple operations into a single operation. FOr example, you can run two 16-bit ops or four 8-bit ops at the same time on the same core, effectively doubling or quadrupling your performance. However, both your hardware implementation and the software used must be correctly implemented for this work. It seems that Microsoft has done for their GPU hardware, you just need software to be written to use that features. Nvidia can fuse two of their 32-bit CUDA cores to run a 64 ops, but they disable this feature on their Gaming GPU, but it is on their professional GPU. They used to disable the acceleration from fusing two 16-bit ops on pre-Turning gaming GPU, but they enabled this feature on Turing.

[BrotherO]

No one ever said Xbox has a nvidia tech chip in the Xbox. The question people had was if the Xbox had dedicated hardware like how nvidia does it or not.

[DEVILTAZ]

Nobody said Series X had Tensor cores that is Nvidia GPU specific. It was only stated it had dedicated hardware for ML.
i still can't see it getting widely used for at least a year or two though. XBox is only just releasing the latest version of the GDK prior to the show this month. The great news is EPIC games is already testing it. Anything to improve Unreal Engine on console can't be bad.

[x]

I honestly can't believe you actually had to point this out. Props to you for having the patience to explain things to stupid people. God knows how many tech discussion get totally derailed by the followers of Mister X Media and and that Blue Something Something guy, whose ability of tech analysis boils down to constructing conspiracy theories from marketing material minutiae.

[photonboy]

*RGT idea for video...*
SSD or SSD/HDD?
How about a discussion with a game developer about the challenges of designing a game that will benefit from a fast SSD (for asset streaming) while also being able to run on an HDD?
Can you do both in the same engine for the PC version? When you start creating the game with SSD and HDD in mind how early do you plan for that? What does planning for an HDD limit you to even if you plan to fork the code for a fast SSD version? Does UE5 Lumen/Nanite solve any issues with this? Will there be increased flexibility in a game engine to automatically exploit Video/System memory pools to compensate for slower storage? And so on...
If you did this already I apologize (if you see it).

[mayen]

This is one area (compute) where TFLOPS are a perfect gage of performance. However the SeriesS may muddy the waters because that will have to be catered for when doing ML calculations.

[adrianhosein]

I'm confused now, does this mean the Series consoles have dedicated hardware just for ML or not? From what he's saying I gather he's trying to say they don't and that the only thing that's been confirmed for Series consoles that's not known on PS5 is that they have the 4 and 8 INT for this task available for use and that these will be run on the CU's also. @RedGamingTech is this a correct summary?

[Coloradocaliber]

When people say mimic that doesn’t mean it is that exact same thing. That means they mimic or copy that thing. The Xbox can also use ML for many other things. This is just like those times before I told you, you needed to retract your RDNA3 and full RDNA 2 vids. Infinity cache. You are incorrect about the acceleration too. Xbox features that too.

[KageMarug]

We have an engineer at Sony that already confirmed the PS5 doesn't have any ML support in hardware. This would include the 4 and 8 int support.

[benwoolley]

Great video Paul but let’s be honest if Sony had machine learning they would already be shouting about it like Microsoft are. The reason Microsoft are shouting about it so much as they know once Sony come out and finally say it, it will be a bigger issue as more people would of heard of it IMO

[theemulationportal]

Any more info on 3070ti Paul? need that 16gb vram for my modded skyrim 😅😂

[glenn]

Does a game right now on the series use M.L. ? Not yet right?

[AlboManny]

Great video with a great explanation. This generation console fanboys (mainly Xbox fanboys) have a big interest in hardware, /software/feature/tool sets like ML, teraflops and all those lovely things, but they don't really seem to understand it. I think they just take what Xbox centric channels say at face value, which sometimes has been incorrect information as we have seen time and time again and just run with it.

[m_sedziwoj]

4:25 for real, you take this schematic and use as source for size? Do you see RT cores, this big thing which in reality is only 2% of die?

[UltraCasualPenguin]

PS5 can probably run 32bit and 64bit. They just have to borrow some power from CPU.
Just lower CPU speed to somewhere around 500Mhz/1Ghz it should have enough electrons flowing in GPU's/Tempest's/Geometry's direction for 64bit calculations. I even heard that Nvidia has designed Tempest and Geometry to fully utilize RT cores and up to 128bit ML.

[werksfeerii]

I'd like to know what this implies for the series s.

Can it do 720p 60fps / 1080 60fps / 1440p 60fps?

[MrPhyxsyus]

So your sources are an 1 year old article about Xbox and and nvidia PDF while talking about amd. That's seems about rigth.