Technical investigation into PS4 and XB1 audio solutions *spawn

Bilikan can you post a list of functions that shape does
Looking at the tensilica product briefs they are mainly about hardware support for many different codecs and accelerating many audio streams at once plus voice processing (echo cancellation ect)
the only references to more realistic game audio I can find is the core can put many audio streams correctly into 3d space. I dont see anything about wave tracing or being able to send level geometry or level materials (carpet, wood, marble ect) to the core ?

 

You can see the fixed functions of the SHAPE block here, as well as descriptions of the other blocks in the audio processor.

http://www.vgleaks.com/durango-sound-of-tomorrow/

 

hmm but wont developers want to keep cpu parity this time around since both machines use the same processor? that wouldn't leave much room for ps4 version of a game to do audio on the cpu while the xbox one version is doing it on shape. wouldnt that like cause troubles? it seems theyd HAVE to use the cus for audio.

 

Or you limit XB1 to whatever the PS4 can do on its CPU...

 

The good news is that one of the two major audio middleware tools (FMOD) already has support for SHAPE. I would assume that in order to stay competitive on the XBO platform, the other (WWISE?) will as well.

If the middleware tools have native support for SHAPE, does that mean its easy for developers to get the hardware benefits of SHAPE using their regular workflows or does it still require additional effort?

 

Shape is not a tensilica core. It is designed from the ground up by MS. The four tensilica cores are general purpose DSPs with added instructions also designed by MS. A general purpose DSP can be used for any audio processing task, including the examples you listed, although I don't know if game developers can offload any of their DSP effects to the tensilica cores. The shape hardware supports dma in and out of main RAM during pipeline processing, so any custom DSP work could be done on the CPU or GPU while still making use of all the other shape acceleration.

 

I believe you heavily underestimate the power of the best sound cards. You can get audio cards with 1-12 fully programmable ADSP-21369 onboard. A single 21369 can perform 4100+ biquad filters (40bit float) at 48KHz.

According to Vgleaks - http://www.vgleaks.com/durango-sound-of-tomorrow/ - the Shape hardware block have 6 fixed functions. I can't find any information about shared resources, so let us assume that SHAPE can run all 6 functions in parallel at full speed and all numbers are based on an operation rate of 48KHz.

* EQ/CMP = 512 * (3 * biquad + hard-knee compressor/expander) = 48% of an ADSP-21369.
* FLT/VOL = 2500 * (MUL + SVF) = 90% of an ADSP-21369. Here we assume that the filter outputs all 4 types (LP, HP, BP and notch) in parallel- if not we can optimize the 21369 algorithm further.
* Mix buffers = 128 * clip/mix/metering = less than 2% of an ADSP 21369.
* SRC : Unable to estimate because of missing information. How many orders are the polyphase filters? The 21369 also have SRC blocks
* XMA : Unable to estimate because of missing information. Audio decoders are available in the 21369.

You can't compare CU's, CPU's and SHAPE like that. The primary functions of SHAPE are simple IIR filters based on recursive structures which CU's are extremely bad at emulating. On the other hand CU's are extremely good at different transfomations, convolution etc. which is not available in SHAPE.

Forza could use SHAPE for the filter functions and therefore free up some CPU power, but since Forza use convolution, they use 1+ CPU core for the processing of the convolution. The same audio engine would look different on PS4, since you would use the CU's for convolution to free up the CPU to process all the recursive filters.

 

Bkilian helped design it. No offence, but I think he knows what he is talking about..

 

He seems like he might be right when comparing to Soundblaster. That chip you're talking about looks like it's used in incredibly expensive products that cost thousands of dollars. Is it used in any soundcards that a gamer might buy?

 

I don't think he did design it. He was working on the audio portion of Kinect.

Not that he needs me to speak on his behalf.

 

I'm quite sure he would respond if he (or any one else with DSP programmming experience) finds any wrong facts in my post.

He would be right if he meant Soundblaster or equal game sound card, but I was responding to 'best sound card'. The chip is used in sound cards (although not a sound card that a gamer would buy) to around $249, so I don't think that's incredible expensive.

 

Sorry, I should have said "consumer sound cards". The X1 is not a professional mixing station, it's a game console.
And the mix buffers have 128 physical buffers, but can be used with over 4000 virtual buffers per audio frame. Think of them as registers that can hold an entire audio frame. The 21369 has 32, much smaller ones. The SRC can process 512 channels per audio frame, and the XMA decoder can decode 512 channels per audio frame.

The clock speed of the audio block is twice that of a 21369, and the fixed function blocks were calculated, per the hotchips presentation, to be 18 GOPS equivalent. The 21369 is 2.4 GFLOPS. If you assume the scalar tensilica cores are about the same power per clock of a 21369, and use the 15.4 GFLOPS value for the two vector cores, you're talking 23 21369s equivalent for the whole audio block. How much did that 12 core sound card cost again? I found an 8 core one for something like $1500. Let's change my statement to "the Xbox one audio block is far more powerful than any sound card you can buy for less than or equal to the price of an entire Xbox one."

I believe you heavily underestimate the power of the X1 audio block.

 

Well then this is getting interesting .

 

He was quoting the price for a card with a single core on it. It's not even in the same ballpark.

 

I did not help design it except in the sense that I helped verify that the design could do what was expected of it.

 

GFLOPS on this board is heavily inflated. Are you arguing that the X1 audio block can perfrom 96K+ biquad filters at 48KHz? Or 990K 1024 tap FFT (Radix 4) per second?

I might, but then please enlighten me.

How many X tap convolution (FFT or brute)/second can be performed by the X1 audio block in total? Or biquads? FDN's? Anything actually. Then I have the possibility to measure against x86 and ADSP and we can praise the power of X1 together. Believe me when I say we would rather sell systems to cheap X1's instead of expensive DSP cards or PC systems.

 

So bilikan on the question of enviromental audio modeling (or eax5 if you like as this is my benchmark)
it seems your saying
can shape do eax5 - no
can the xb1 as a whole do eax5 - yes

am I right

 

Using the two vector cores, I don't know. If it uses the same floating point ops as what they're using to define a flop on your DSP, then the vector cores can perform 6x the number of operations than the Sharc DSP. and for some reason I stupidly added the vector cores twice, so only 17x the ops of a Sharc, in total, not 23. But not all general purpose ops, so hard to compare directly.

 

Does your definition of xb1 include the cpu? The cpu can easily be used for environmental simulation in eax5 quality.

 

I'm not using the term FLOPS since they don't represent any workload in real life scenarios. It's would be highly appreciated if you could post any audio DSP related measurements - it could be x taps of convolution (favors wide SIMD), recursive structures (Allpass, biquad, fdn etc.) or anything else related to audio processing. This way we could very easily compare between platforms, SHAPE, x86, ADSP, CU etc