AMD CEO talks of long-term turnaround

[jeff_rigby]

Ok, I reenter this discussion.

In this interview the ESRAM decision is explained by costs and the requirement to be able to do GPGPU stuff.
http://www.eurogamer.net/articles/digitalfoundry-the-complete-xbox-one-interview

I agree that the ESRAM choice was for GPU NOT GPGPU of which 15 blocks are primarily ARM and Xtensa Stream processors. Xtensa stream processors have their own memory and access main memory or other blocks on the ARM AXI buss as NOC not needing the bandwidth the GPU needs.

PS4 Additional hardware: GPU-like Compute module, some resources reserved by the OS
"However, there's a fair amount of "secret sauce" in Orbis and we can disclose details on one of the more interesting additions. Paired up with the eight AMD cores, we find a bespoke GPU-like "Compute" module, designed to ease the burden on certain operations. We're assured that this is bespoke hardware that is not a part of the main graphics pipeline but we remain rather mystified by its standalone inclusion, bearing in mind Compute functions could be run off the main graphics cores and that devs could have the option to utilise that power for additional graphical grunt, if they so chose."

Durango additional graphics hardware - "rumours have circulated for quite some time that it is some way behind Orbis, but equally there has been the suggestion that the GPU itself is supplemented by additional task-specific hardware. We could not confirm this, but an ex-Microsoft staffer with a prior relationship with the Xbox team says that two of these modules are graphics-related."

For the PS4 the GPU compute blocks mentioned above are in Southbridge and for the XB1 are the ARM block in the XB1 APU. Both are using Xtensa processors which can emulate a GPU or provide GPU compute or as an accelerator for codecs just as a GPGPU can do except 10-100X more efficiently. "Two of those blocks are graphics related" are the codec encoder and decoder Xtensa accelerators.

For costs I think you misread.

Nick Baker: Yeah, I think that's right. In terms of getting the best possible combination of performance, memory size, power, the GDDR5 takes you into a little bit of an uncomfortable place. Having ESRAM costs very little power and has the opportunity to give you very high bandwidth. You can reduce the bandwidth on external memory - that saves a lot of power consumption as well and the commodity memory is cheaper as well so you can afford more. That's really a driving force behind that. You're right, if you want a high memory capacity, relatively low power and a lot of bandwidth there are not too many ways of solving that.

Power use is stressed multiple times but price only once. Add that the ESRAM block costs more to implement than the difference between DDR3 and GDDR5...at least that was part of the explanation for the XB1 APU having fewer CUs than the PS4 which is said to cost less to produce.

If your design requires the ARM block to be inside the APU like current and future AMD designs so the work on the XB1 can be used with future AMD APUs and you don't have access to HBM, the only choice is to have ESRAM and DDR3.

Why did the PS4 move it's ARM block out of the APU to Southbridge with it's own DDR3 memory? Power use. That was more of a break with AMD APU designs while the XB1 is more future AMD design but without the "25X20 power efficiencies" that are coming by 2020.

The ARM blocks are also the DRM security blocks; Codec, Encryption, Keys, Player, output buffer and more. They are key to media streaming both in the home and from the cloud. Access to those blocks is limited to APIs and both Sony and Microsoft are still writing the OS that uses those blocks. Playready ND is still a work in progress. Playready ND is the 4K version of DTCP-IP WMDRM10 which content owners trust only for 1080i and lower media.

 

[Dresdenboy]

I agree that the ESRAM choice was for GPU NOT GPGPU of which 15 blocks are primarily ARM and Xtensa Stream processors. Xtensa stream processors have their own memory and access main memory or other blocks on the ARM AXI buss as NOC not needing the bandwidth the GPU needs.

For the PS4 the GPU compute blocks mentioned above are in Southbridge and for the XB1 are the ARM block in the XB1 APU. Both are using Xtensa processors which can emulate a GPU or provide GPU compute or as an accelerator for codecs just as a GPGPU can do except 10-100X more efficiently. "Two of those blocks are graphics related" are the codec encoder and decoder Xtensa accelerators.

Xtensa are 10-100fold more efficient than GPU? This likely depends on the given tasks.

To fill this discussion with some plain architecture overview, I think this one should help:

In this configuration ESRAM would still be no help for doing VISC like stuff.

For costs I think you misread.

Power use is stressed multiple times but price only once. Add that the ESRAM block costs more to implement than the difference between DDR3 and GDDR5...at least that was part of the explanation for the XB1 APU having fewer CUs than the PS4 which is said to cost less to produce.

OK, you're better informed about the XB1. What are the ESRAM costs and what's the cost delta between DDR3 and GDDR5 incl. "hidden" costs like power adapter and cooling requirements? I read one estimation of $30 for the ESRAM, while GDDR5 would have costed $18 more according to some. So something is not fitting here.

And that part of your quote should also be bold:

You can reduce the bandwidth on external memory - that saves a lot of power consumption as well and the commodity memory is cheaper as well so you can afford more.

Either he lies (according to Dr. House he does), or he should know what he's talking about.

Implementing a memory solution does not only cost ordering the parts.

 

[jeff_rigby]

Xtensa are 10-100fold more efficient than GPU? This likely depends on the given tasks.

To fill this discussion with some plain architecture overview, I think this one should help:

In this configuration ESRAM would still be no help for doing VISC like stuff.


OK, you're better informed about the XB1. What are the ESRAM costs and what's the cost delta between DDR3 and GDDR5 incl. "hidden" costs like power adapter and cooling requirements? I read one estimation of $30 for the ESRAM, while GDDR5 would have costed $18 more according to some. So something is not fitting here.

And that part of your quote should also be bold:

Either he lies (according to Dr. House he does), or he should know what he's talking about.

Implementing a memory solution does not only cost ordering the parts.

Notice the thin black line (above block diagram) going from ESRAM to the CPU coherent cache which I think represents the custom Jaguar to ESRAM connection you can physically see in die shots.

Compare the XB1 Jaguar to PS4 Jaguar. In the XB1 upper Jaguar upper right and lower Jaguar lower right are two tiny blocks not seen on the PS4 Jaguar. Power gating is at the center line of the Jaguar block for both the PS4 and XB1. ARM block for the XB1 is upper right to the left of the SRAM block.

Notice the PS4 has no ARM block...that is hard to see as the IO blocks on the PS4 right side could be mistaken for ARM blocks. If you notice they have a courser design and are split up on that right side while the XB1 ARM block is packed as closely as possible in the XB1 because they are all on the same bus.

I am better informed on only a few features of the consoles....I am rapidly out of my depth here. Of one thing I am sure and that is the decision to use DDR3 was because the Low power modes run on the ARM block need as close to 500mw as possible for Network standby. Both consoles have the ARM blocks running with DDR3.

 

[Dresdenboy]

Notice the thin black line (above block diagram) going from ESRAM to the CPU coherent cache which I think represents the custom Jaguar to ESRAM connection you can physically see in die shots.

Compare the XB1 Jaguar to PS4 Jaguar. In the XB1 upper Jaguar upper right and lower Jaguar lower right are two tiny blocks not seen on the PS4 Jaguar. Power gating is at the center line of the Jaguar block for both the PS4 and XB1. ARM block for the XB1 is upper right to the left of the SRAM block.

I don't think, that we could see such connections that clearly, especially in these images. Could it be some tag bus?

As far as the image quality permits, these 2 blocks look like the square block sitting between the 2 PLLs left from the Jaguar module:

Still no signs of VISC.

 

[Dresdenboy]

One more thing ...

 

[buletaja]

that arm block is for PSP
not related to MCU or DSP, or Tensilica
not all is tensilica based
some FPGA etc
you see the eSRAM has lighter color vs the surrounded
because it is stacked

the reason chipwork not doing cross sectional cut
when they do to pS4

same thing to techinsight doing cross sectional Xray to X1 Nandflash which is 3D IC
but refuse to showed the Xray or cross sectional cut of main SOC

it is like Nandflash is more important ..... go figure

 

[jeff_rigby]

One more thing ...

Good find. So the two blocks in the XB1 Jaguar are identical to a block in the PS4 memory controller. That makes sense as it's supposed to support coherent memory access to ESRAM.

 

[jeff_rigby]

Back to features that could result in a AMD turnaround.

DSTAC WG3 Report

DSTAC Mission of identifying “a not unduly burdensome, uniform, and technology- and platform-neutral software-based downloadable security system”.

Most of this paper is discussing the implementation of a security system both the hardware and software requirements.

The two proposals for a downloadable security scheme, both likely use Sony's Passage, are:

1) HTML5 <video> EME MSE (Used with TV Tuners then direct IPTV from cable modems) from the browser or an APP. An APP may be required for special features not supported by the browser. The example given was gesture control but Gesture and Voice control are or will be part of some browsers.
2) Vidipath (Used with Cable TV Gateways, first DVRs and then direct from a cable modem with DLNA)

Using HTML5 allows for competitive UIs while Vidipath does not.

PCs with TEE either AMD APUs or modern dGPUs, XB1, PS4 can be a DVR or STB for Cable TV.
A Media Hub for ATSC 2.0
A 4K blu-ray player with digital bridge and stream blu-ray over the home network

This will sell more PCs and video cards and explains the thrust for G series APUs http://www.engadget.com/2014/01/07/amds-impossibly-thin-nano-pc/

 

[buletaja]

this is just only several work resume that hinted ....

the test chip is 3D start with 32nm, the final one is 22nm (fabbed at CPA)
the Jag SOC is 28nm TSMC (later on)

 

[jeff_rigby]

this is just only several work resume that hinted ....

the test chip is 3D start with 32nm, the final one is 22nm (fabbed at CPA)
the Jag SOC is 28nm TSMC (later on)

These might confirm that Yukon got to FAB testing before it was abandoned in favor of Durango without a PowerPC on interposer for BC. Then Jaguar's direct ESRAM was designed in and a higher clock Jaguar than PS4 for BC emulation.

 

[mrmt]

Back to features that could result in a AMD turnaround.

The thing missing at AMD, jeff, isn't technical features, these are already plenty at the company, but the lack of BUSINESS features. AMD is in a no-man's land today, they are not a bleeding edge company in the market they operate (that honor falls to AMD's competitors) and they are not the leader on costs, as AMD competitors are much more focused on this.

The net result on the GPU market is that AMD is a nobody in the fast growing mobile market, and they are having bad margins on premium segments with their premature HBM debut, while on the CPU the situation is catastrophic, with Beema and Mullins sporting bigger die than current Skylake 2C, and I'm not going to even talk about the latest CMT crap they fielded on the market.

Until AMD refocus its BUSINESS, e.g., developing things someone actually wants to buy with a descent cost structure, you'll see plenty of hot air coming from the company but no profits at all.

 

[buletaja]

These might confirm that Yukon got to FAB testing before it was abandoned in favor of Durango without a PowerPC on interposer for BC. Then Jaguar's direct ESRAM was designed in and a higher clock Jaguar than PS4 for BC emulation.

you reading wrong Jeff
always downplay abit on X1 news
but add more sauce to the PS4
when there is nill fact of tensilica on PS4 for example


they not abandoned OBAN
OBAN is eSRAM 3D IC
How they can support BC
they have HPAPU-->
If you have 2 SOC the story told in media wil be
1st they said 1st SOC --> OBAN or whatever
then the last before launch it will be 2nd SOC (the Jag one)


you can clear check the XDK why
eSRAM on GPU has 2 memory controller
esram is 2048 bit wide (256 byte per cycle)

because the eSRAM you see on news
is the scratch one not OBAN eSRAM one
the scracth one is 4x256 bit 4xMC

or why ADK only has DX 11.0
when XDK is basically DX 11.x / DX 12.x
11.x is designated of always uptodate feature level
means it is fully programmable

 

[buletaja]

More hint ....
===========

Then VS DX11 GPU
pay attention to stream out

 

[jeff_rigby]

you reading wrong Jeff
always downplay abit on X1 news
but add more sauce to the PS4
when there is nill fact of tensilica on PS4 for example

It's been confirmed the PS4 has Tensilica Audio DSPs which are Xtensa processors with DSP co-processors on the Xtensa controlled package. If you use Tensilica IP for audio you MUST use Tensilica IP for the video or you can't use the Tensilica software libraries for Codecs etc.

Sony moved all ARM hardware including Tensilica IP out of the AMD APU into the ARM SoC as Southbridge because they wanted to use GDDR5 for the APU. Microsoft chose to keep the ARM hardware including multiple Xtensa processors in the APU so they had to use DDR3 so that they could have low power modes. Here is where you are wrong; I have repeatedly said that the Xtensa processors in the APU help level the playing field between the XB1 and PS4. How much and what features the Xtensa processor can support in the XB1 that the PS4 can't take advantage is unknown but it's potentially there.

How this matters for this thread is also interesting. All AMD APUs without expensive SDRAM using conventional memory are memory starved and can't use GDDDR5 memory because GDDR5 can't support REQUIRED by Energy Star low power modes. This is why they have small GPUs. IF they didn't support Energy Star power levels they can't be sold to the US government.

HBM can support low power modes and Performance modes and with HBM larger AMD APUs are coming. That plus Windows 10 could support a comeback for AMD.

they not abandoned OBAN OBAN is eSRAM 3D IC How they can support BC they have HPAPU--> If you have 2 SOC the story told in media wil be 1st they said 1st SOC --> OBAN or whatever then the last before launch it will be 2nd SOC (the Jag one)

Just posting because I find it interesting. The Oban term is being used for IBM Power and ARM, the common theme is about switching/networking inside the chip for SERVERs. Oban is a city in Scotland where the first transatlantic telephone cable, (multi channel telephone switching) considered a marvel of the age, lands.

Cadence-Tensilica-Xtensa processors use Networking on Chip and they are on a ARM AXI buss which can support the same. So inside the Tensilica package and outside the package NOC is used. In the XB1 inside the ARM SoC inside the AMD APU NOC is being used but I believe the 4 memory controllers which are also used for the AMD APU are also Cadence IP so Oban features may be in the XB1 APU for GPU and Jaguar support. I can't confirm that and in any case Oban SERVER CPUs I Think support NOC externally which I think we know the XB1 does not do.

The Sony ARM SoC as Southbridge in the PS4 does support the same but is primarily a ARM TEE DRM secure media processor and AOAC network standby processor.

 

[jeff_rigby]

This is speculation trying to determine what the "Beyond Gaming" AMD semi-custom win is all about. From past AMD quotes these "wins" are min 300 million in revenue. That volume just about only supports Gaming SoC/APUs but it's beyond gaming???

Oban was a descriptor for a NOC Fabric used by both a ARMV8 design and IBM 32nm PPC design for servers. It was coined from the Europe landing site (Oban Scotland) for the first transatlantic cable able to carry media (voice) using an addressing scheme. It was considered a marvel.

The ARM AXI buss and internally Xtensa processors use NOC.

Yufei Li

-Power Server System Enablement at IBM

Oban ( 32nm XBox Chip )

January 2011 &#8211; December 2011

Acting as pervasive verification engineer in Oban project, which is PowerPC based multi-core gaming processor (MS XBox). In this project, he is responsible for pervasive JTAG logic verification

The 32nm Xbox 360 was the Xbox 361 mentioned in the Xbox 720 leaked powerpoint which was supposed to be released December 2012 to support the FCC Vidipath mandate which was delayed twice by TiVo suits to June 2015.

The design differences between the Xbox 360 and Xbox 361 were Low power modes, 1080P and S3D support as well as concurrently running a browser and other apps at the same time. The Xbox Mini rumor from 2013 is for a Xbox 360 that supports the above as well as game streaming to the XB1 and PCs.

A Xbox 360 Oban, if I have the naming scheme correct, would have a AXI bus or some form of NOC internal addressing. I'd guess AXI buss and Xtensa processors for codec and low power media modes as well as voice and gesture control.

Global Foundries FD-SOI @ 22nm is capable of supporting 3.2 Ghz PowerPC or Cell and ARM IP has also been "ported" to FD-SOI.

So what does beyond gaming refer to? NOC Fabric HSA IOMMU tied through the USB 3 or 10GB/sec LAN with distributed processing on the local home network. A refreshed Xbox 360 and PS3 cheaper and supporting this? Xbox TV as a semi-custom Mullins designed to emulate a Xbox 360?