AMD HSA ARCHITECTURE : Could we be looking at the PlayStation 4 SoC & not even know it? ( Fixed Function Accelerator & DPS )

You all might remember a interview with Sony's CTO last year where he talked about the future of PlayStation & he described the PS4 Chip as being a chip with most of the processing still being done by the CPU & GPU but also being helped out by DSP's & Programmable Logic  

“It took five years before we saw games that used the full power of Cell, so we are used to looking ahead and having capacity,” Tsuruta says. “We are looking at an architecture where the bulk of processing will still sit on the main board, with CPU and graphics added to by more digital signal processing and some configurable logic.”  

Fast forward  to today & we have a SoC pictured in AMD's Next Generation APU (Kaveri ) & HSA presentation.

This SoC shown in the documents fit the description having a DSP & 'Fixed Function Accelerator'  which could be the Vector Processor\ Compute Modular  that some people have reported to have seen but we haven't heard much about it.


From EuroGamer

 "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 - physics calculations are a good example of traditional CPU work that are often hived off to GPU cores. 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."



From a trusted member on Ars Technica forums

So, a couple of random things I've learned: -It's not stock x86; there are eight very wide vector engines and some other changes. It's not going to be completely trivial to retarget to it, but it should shut up the morons who were hyperventilating at "OMG! 1.6 JIGGAHURTZ!". -The memory structure is unified, but weird; it's not like the GPU can just grab arbitrary memory like some people were thinking (rather, it can, but it's slow). They're incorporating another type of shader that can basically read from a ring buffer (supplied in a streaming fashion by the CPU) and write to an output buffer. I don't have all the details, but it seems interesting. -As near as I'm aware, there's no OpenGL or GLES support on it at all; it's a lower-level library at present. I expect (no proof) this will change because I expect that they'll be trying to make a play for indie games, much as I'm pretty sure Microsoft will be, and trying to get indie developers to go turbo-nerd on low-level GPU programming does not strike me as a winner.

From this person ? 

Originally Posted by JonathanCarabalo

The PS4's world-famous Vector co-processor 

A truly innovative feature - the Playstation 4's world-famous vector co-processor. It's connected to a CPU core. It doesn't look like much on paper:

• 100 million transistors
• 300 mhz
• $25 to produce
• 25 nm production process
• 16-way SIMD floating-point multiply-add instructions (512-bit)
• 1024-bit load-store instructions
• 512 Kbytes local memory
• L2 cache of the X86 CPU
• Write-to-memory feature - bypassing the CPU
• Has a connection to the instruction dispatch of a regular x86 Jaguar core

It doesn't look like much, except for the 16-way multiply-adder! Besides that though - it only costs $25.00. "How can $25.00 worth of CPU logic improve the graphics pipeline?" Thus - top-secret alien technology. The world-famous vector co-processor. With it - developers can pull off pretty amazing effects, in real-time. I've already seen:

Full-screen per-pixel lighting! Pretty amazing - I haven't seen any games do that.
Extremely realistic fire effects
Extremely realistic ice simulation
Massive degrees of normal/displacement mapping
And this is just the start. 


_______________________________________________________________________

Originally Posted by JonathanCarabalo

SIMD Vector Processor Comparison

Playstation 3 - Cell

• Developed by IBM, Sony, and Toshiba - mostly IBM, at Austin Texas
• Based on the Power4 architecture, the PowerPC 970 main core
• 7, 128-bit Floating point multiply-adders
• 256 Kbyte local memory for each SPU
• 1 PowerPC 970-based main core
• 512 Kbytes L2 cache connected to the 7 SPUs via a Token-Ring bus - which they used to inter-communicate as well
• 200 GB/s EIB L2 cache bandwidth
• 3.2 ghz clockrate
• 230 million transistors
• $400 million to develop, $100 to produce
• 4-way SIMD

Playstation 4 - Vector Co-Processor

• Developed by AMD and me (just a joke)
• Based on the X86 architecture
• 1, 512-bit floating point multiply-adder
• 1.6 ghz clock rate
• 512 Kbyte local buffer + the CPU core's main buffer
• Connected to the bus directly via a bus-interface unit, and to a CPU core
• About 300 million transistors
• $0 to develop, $80 to produce
• 307.2 GB/s buffer to execution unit
• 16-way SIMD
• 8 X86-based cores, each with 128-bit FP unit, L2 cache, and shared local memory between pairs of CPUs and between all 8 

Doesn't seem as powerful as Cell does it? In fact - it's way, way more powerful. I just found this an interesting comparison. This design is more powerful - as the CPUs do not require direct developer intervention to get them to perform, as opposed to Cell, which only worked if the developers specifically coded for it.

16-way SIMD makes it 4 times faster, per clock. I.e. to do 16 FP multiply-adds, it takes 8 cycles. 4 load-stores, followed by 4 multiply adds - the Cell. But for the Vector Co-Processor to do 16 FP mulitply-adds, it takes 1 cycles. The Out-of-Order, 512-bit load-stores load the data, while the Vector Co-Processor is performing the data. Thus - the Vector Co-Processor alone is equal to 8 SPUs. 






& from this person 

The PS4 specter vector? I'm not a 100% sure what it is at this stage... I'm 45% leaning towards physics offloading or helping out within that department. The other 55% is screaming a modified component for helping PS3 games work within the G/Cloud environment.

And as you can see The SoC in the picture above is not Kaveri.