Revenge of Cell, the DPU Rises *spawn*

[Shifty Geezer]

Again, my interpretation of the above blog, 'control domain' is CPU. GPU doesn't get a mention, and rightly so because Tensilica is in competition with compute. But there are domains where GPU still isn't ideal where some other processing element makes sense. If CPU deals with branch heavy code, and GPU deals with latency-tolerant wide data processing, fast serial data processing on more efficient cores than monolithic CPUs can justify a third processor type. Something like heterogeneous cores, mixing some power efficient throughput cores with CPU and GPU, certainly makes sense in power conscious devices.

Not for VR reprojection though.

 

[liolio]

But what's the control domain in an soc, cpu, gpu? This one is going over my head.

The Anandtech article I linked dkes actually a good job at describing what TENSILIA DPU may be thought the scope of their desings may be greater as they do not focus on image processing alone.

 

[Scott_Arm]

I guess the issue becomes, the more "general purpose" the DSP becomes, the more you question whether it wouldn't be better to just put that silicon into the CPU/GPU. If you have very specific uses in mind, like some kind of reprojection, you have to be confident your custom silicon is going to be a very good solution, and not something devs will ignore if they come up with better approaches on the gpu (as they tend to do).

Edit:
They already have audio dsps in xbox one and PS4 for all of the voice codecs, and that's a good fit. Audio in general is a good place for DSPs. I suppose with the rise of game clips and screenshots, there's a lot of room for DSPs to apply video filters and more sophisticated video editing. But maybe not, because you probably wouldn't be gaming and editing video at the same time, so you have all of that gpu available.

 

[onQ]

I don't even understand what data plane and control plane would mean outside the context of telecom.

Again, my interpretation of the above blog, 'control domain' is CPU. GPU doesn't get a mention, and rightly so because Tensilica is in competition with compute. But there are domains where GPU still isn't ideal where some other processing element makes sense. If CPU deals with branch heavy code, and GPU deals with latency-tolerant wide data processing, fast serial data processing on more efficient cores than monolithic CPUs can justify a third processor type. Something like heterogeneous cores, mixing some power efficient throughput cores with CPU and GPU, certainly makes sense in power conscious devices.

Not for VR reprojection though.

That's 3 processor types I remember reading somewhere that it's best to have at least 4 different types of processors on your HSA SoC.

 

[Shifty Geezer]

From the article:

In this article I would like to pick up on the processor and multiprocessor taxonomy themes that Robert Cravotta introduced in his article, and his two blog posts (first and second). Robert divided the multiprocessing world into four categories: “channel-based, aggregate-based, multi-domain, and feedback architectures”.

That suggests four types of processor. But I'm not seeing the need for a fourth architecture myself. The three seems capable of covering all workloads, at least in a console.

 

[turkey]

That's 3 processor types I remember reading somewhere that it's best to have at least 4 different types of processors on your HSA SoC.

Given how Microsoft seem to go all in with their tech toys you can probably count on them adding a HPU?

 

[mrcorbo]

Wouldn't this also make sense of how Microsoft was able to update the Xbox One for 10-bit HEVC?

Not necessarily. AMD partnered with Strongene to develop an OpenCL decoder which, "supports the OpenCL devices like AMD HD 5000 and above discrete GPUs, and AMD APUs (like Richland and Kaveri)." They could also be using a hybrid decoder like Intel did to add HEVC decode to Haswell, using a combination of the fixed function blocks and GPU compute. I'm not aware of anyone having done power usage measurements while playing back HEVC content on the XBOne, but that could give some indication of which approach they took.

 

[onQ]

From the article:
That suggests four types of processor. But I'm not seeing the need for a fourth architecture myself. The three seems capable of covering all workloads, at least in a console.

Imagination & Siliconarts have RTUs I wouldn't mind a RayTracing chip getting tossed in the mix of Next Gen consoles.

http://www.extremetech.com/computin...-3d-and-ray-tracing-hardware-in-a-single-chip

http://www.siliconarts.com/#!raycore-gpu-ip/ca0a

 

[onQ]

Not necessarily. AMD partnered with Strongene to develop an OpenCL decoder which, "supports the OpenCL devices like AMD HD 5000 and above discrete GPUs, and AMD APUs (like Richland and Kaveri)." They could also be using a hybrid decoder like Intel did to add HEVC decode to Haswell, using a combination of the fixed function blocks and GPU compute. I'm not aware of anyone having done power usage measurements while playing back HEVC content on the XBOne, but that could give some indication of which approach they took.

If that was the case wouldn't it be just Netflix & other apps updating their software instead of Microsoft updating their firmware? unless they are using some of the compute that was being used for Kinect on the OS side.

 

[HTupolev]

Imagination & Siliconarts have RTUs I wouldn't mind a RayTracing chip getting tossed in the mix of Next Gen consoles.

If you mean as a replacement for the traditional GPU, then we're still at 3.

If you mean alongside the traditional GPU, where we have two versatile independent graphics processors, we have a bloated, inelegant design with a bunch of really awkward redundancy.

If you mean integrate a ray-trace accelerator into the main GPU, then we're still at 3, and we're having a typical discussion about GPU featureset.

 

[Scott_Arm]

@onQ Microsoft provides the codec in the OS, which the app calls through an api to playback video. The api is basically a black box from the developer perspective. Most likely a combination of CPU and GPU decoding.

 

[steveOrino]

Has there been any shipping software or demos to show off the benefit of these RTUs? I can see RTUs being a good fit for mobile SoCs where current battery tech is a huge wall to climb but I am not seeing how its a benefit to the current trajectory non-mobile GPUs and software development.

 

[Scott_Arm]

The other issue with having custom silicon is that it's probably not the favourite option for multi-platform devs, which is really what drives the industry now.

 

[onQ]

The other issue with having custom silicon is that it's probably not the favourite option for multi-platform devs, which is really what drives the industry now.

What if it's transparent to the game devs?

Edit:

AMD new GPU's will have AR / VR acceleration I'm guessing that will be part of the DPU.

 

[Scott_Arm]

@onQ That image you posted literally shows VR/AR acceleration as part of the GPU. It makes even less sense to make custom silicon outside the GPU, if GPUs going forward are going to have it.

 

[onQ]

@onQ That image you posted literally shows VR/AR acceleration as part of the GPU. It makes even less sense to make custom silicon outside the GPU, if GPUs going forward are going to have it.

The DPUs are inside of the GPUs lol. what you think is doing the voice/face recognition & all the video recording & streaming on these consoles? I guess it's all going over your head.

AMD's TrueAudio and Unified Video Decoder are ASICs based on Xtensa.

http://ip.cadence.com/ipportfolio/tensilica-ip/xtensa-customizable

http://ip.cadence.com/ipportfolio/tensilica-ip/audio

http://ip.cadence.com/ipportfolio/tensilica-ip/image-vision-processing

 

[Shifty Geezer]

The DPUs are inside of the GPUs lol. what you think is doing the voice/face recognition & all the video recording & streaming on these consoles?

If the video en/decoder isn't programmable, it's not a processor but a fixed function unit.

 

[Scott_Arm]

@onQ

The DPUs are inside of the GPUs lol. what you think is doing the voice/face recognition & all the video recording & streaming on these consoles? I guess it's all going over your head.

So what you're saying is, GPUs already include Tensilica "DPU"s now, and that will continue going forward. So, absolutely nothing will change. We'll have a CPU and a GPU, and the GPU may contain some new blocks for processing video and possibly interpolation/reprojection. What's the significance of the DPU here?

This thread started with you claiming parts of the GPU would be moved into their own "DPU". That doesn't appear to be the case.

 

[onQ]

If the video en/decoder isn't programmable, it's not a processor but a fixed function unit.

It's part of the DPU like APC (TrueAudio) is part of the DPU & the DPU is the GNB SoC that I told y'all about years ago Starsha

PS4

• 1. New Starsha GNB 28nm TSMC
     Milos
     Southern Islands
     
     DX11
     SM 5.0
     Open CL 1.0
     Quad Pixel pipes 4
     SIMD’s 5
     Texture Units 5TCP/2TCC
     Render back ends 2
     Scalar ALU’s 320
     
     EDIT: Some of those were crossed, may be they were updated/changed at a later date, I have no idea.
     Quote:
     Couple of more updates
     
     Graphic North Bridge(GNB) Highlights
     Fusion 1.9 support
     DCE 7.0
     UVD 4.0
     VCE
     IOMMU
     ACP
     5x8 GPP PCIE cores
     SCLK 800MHz/LCLK 800MHz

https://forum.beyond3d.com/threads/graphic-north-bridge-playstation-4s-custom-chip.53986/

AKA High Performance Bus & Memory Controller.

 

[BRiT]

@onQ


So what you're saying is, GPUs already include Tensilica "DPU"s now, and that will continue going forward. So, absolutely nothing will change. We'll have a CPU and a GPU, and the GPU may contain some new blocks for processing video and possibly interpolation/reprojection. What's the significance of the DPU here?

This thread started with you claiming parts of the GPU would be moved into their own "DPU". That doesn't appear to be the case.

What, you're not used to onQ especially changing the definitions and moving the goal posts by now? Thats why nearly all posts from him and the follow ups start off as a royal mess of a topic and only go down hill from there. Its practically pure noise and no signal.