Mark Cerny "The Road to PS4"

As said in the other thread 8 ACES with 8 queues each,mem volatile tag and onion + were Sony impulsed enhancements for its APU,and of these, AMD will use whatever they want in its desktop products(the Sony ACEs for example will be used in Sea Islands but was a Sony custom thing first).
Actually, 3dcgi said the following regarding the ACEs:
There's gray area in terms of custom features. Some features wouldn't exist without a customer pushing for them and others can be influenced via collaboration during development. The compute features would certainly have evolved on their own, without console customer feedback.
I interprete this as it was in the works anyway. Sony may have had some influence so some detail may look different than it would have been otherwise, but in principle the extended compute features would have been there also without the console deals.
Did some googling. You're right, it's already been discussed on here. Apparently the Volatile tag is already in GCN? http://forum.beyond3d.com/showpost.php?p=1729818&postcount=1467.
I have to correct myself here. When I wrote this, I thought I had seen this before, but Temash/Kabini appear to be the first APU/GPU to support that volatile flag and the corresponding instructions. So it is not supported by GCN1.0, only the updated "GCN1.1" version used for Temash/Kabini and apparently the PS4 (not sure about XB1).
 
Sony's attempt at playing at having leading-edge semiconductor prowess was one of the largest overreaches of the PS3's generation.
There was the massive expenditure for a fab for all the products Cell was supposed to find its way into that never came to pass.

No, you're off by a generation. It was during the PS2. PS3/Cell was just a repeat of the PS2. Nor would I necessarily call it an overreach. It was solid effort that lead to Sony being the first to fab and ship an MPU on 90nm, and it was certainly plausible to continue. Executive management didn't see that as the direction they wanted to take the company. Partially because they didn't know what markets they wanted to leverage their capability and capacity in, and partially because they didn't know how to be a foundry or desire to become one.

Looking back now, in my opinion, it seems much of what made Cell was exotic stemmed not from good vision, but a deficiency in the expertise of two of the three partners in STI when it came to creating hardware as complex as a modern and robust computing solution.

There was no lack of skill, just a different philosophy of design.

IBM wanted a homogenous solution that played to its expertise in multiprocessor coherent systems, one which we see in hindsight won decisively.
In retrospect, the SPE was what you design if you have limited experience with complex memory pipelines and processor cores, and an lack of experience in self-scheduling hardware linking them together. It was then topped off with some strange internal politics and Toshiba's desire for a split-off media processor it could barely find a use for. The harsh lesson in relying on exotica and backwards tools seems to fit with the acceptance at the time for cheap, primitive consumer electronics with limited functionality and closed, proprietary software tools that try to band-aid over glass jaws.

For IBM, Cell was simply an experiment. A collaboration on homogenous computing with a couple of partners footing a large chunk of the bill. Not terribly different from Intel with Terra Scale and Larrabee, only without the partners to share the costs. Then again Intel has the benefit of large volume and relatively high margins. For Toshiba, a relatively low cost way of acquiring new IP for it's semiconductor business (although they ended up being very good at marketing the product). For Sony, a continuation of what started with Emotion Engine, however with theoretically less presumed risk (ultimately that didn't work out).

The SPE isn't what you get from limited experience. It's a conscious design philosophy. Kutaragi was a believer of maximizing computational throughput with the fewest gates possible. That the hardware engineer should be providing as much computation as possible within a transistor budget. That it's the responsibility of the software engineer to exploit it and utilize it. It's his hardware engineer mentality (I kind of have it too). The problem is that ultimately you're not producing something that's meant to solve big computational problems, but to provide a platform for content factories.

I don't think Sony could maintain the illusion that it really was playing in the same league anymore.

I don't think it was so much of maintaining an illusion as it was just simply a lack of interest in that industry.
 
Actually, 3dcgi said the following regarding the ACEs:
I interprete this as it was in the works anyway. Sony may have had some influence so some detail may look different than it would have been otherwise, but in principle the extended compute features would have been there also without the console deals.
I have to correct myself here. When I wrote this, I thought I had seen this before, but Temash/Kabini appear to be the first APU/GPU to support that volatile flag and the corresponding instructions. So it is not supported by GCN1.0, only the updated "GCN1.1" version used for Temash/Kabini and apparently the PS4 (not sure about XB1).

Onion + is also present in Kabini?.
 
There seems to be something special about operating systems, CPU and GPU which are dominated by western companies, while eastern companies are best at all other semi conductors in general, integration and low-level hardware, including discreet semi conductors and passive components.

Honestly there's VERY FEW companies these days that develop and ship their own operating system (Western or not), particularly on the desktop. There's nothing special about Western countries, they're just at a different phase (post manufacturing), where a lot of technical industry is focussed on IP instead of manufacturing prowess.

I can't speak for other east Asian countries, but in Japan there's never been a large domestic demand for desktop CPUs, GPUs, or operating systems to facilitate enough growth to develop into a pillar industry that it could export. Most processor and software development has focused around embedded electronics and software (e.g. consumer electronics, appliances, automotive, industrial electronics), particularly in the mobile space. Also, for a long there's there's a been a nascent supercomputer industry as well (w/Hitachi, NEC, and Fujitsu)...

FWIW, Sony used to have it's own BSD variant (Sony NEWS) in the 80's and 90's to power it's line of m68k and MIPS UNIX workstations—the later being the original development workstations for the Playstation. Sony's Computer Science Labs also developed the RTOS that powered the AIBO and Qrio robots (was used on a few set top boxes).

Also, not only is the PS4 BSD based, the PS3 is/was as well...
 
Onion + is also present in Kabini?.
In principle, the possibility to bypass the GPU caches (which is said to be the difference to Onion without a plus) exists in all GCN GPUs. It is basically an additional bit in the instruction field for memory instructions which controls this. There are two different bits to set the caching behaviour: GLC (globally coherent) and SLC (system level coherent). A set GLC bit means to bypass/miss the L1 and causes to read from the L2 cache or write through to the L2 cache (a writethrough is actually always performed, but the respective lines don't persist in the L1 for the next wavefront if the bit is set). That is important if you want to ensure coherency between different CUs in the GPU. But the L2 cache still works as a write back cache, so the CPU wouldn't be able to see the changes (as the CPU can't probe GPU caches, it only works the other way around) without a flush. The SLC bit resolves that and really bypasses all GPU caches and reads/writes directly to/from system memory with an APU (with a discrete GPU it would be the VRAM).
 
Last edited by a moderator:
No, you're off by a generation. It was during the PS2. PS3/Cell was just a repeat of the PS2. Nor would I necessarily call it an overreach. It was solid effort that lead to Sony being the first to fab and ship an MPU on 90nm, and it was certainly plausible to continue. Executive management didn't see that as the direction they wanted to take the company. Partially because they didn't know what markets they wanted to leverage their capability and capacity in, and partially because they didn't know how to be a foundry or desire to become one.
They invested heavily in a fab for Cell and all the Cell derivatives that never happened. They recouped some of the cost by selling it to Toshiba, and then later bought it back for camera sensors.
The PS2 made money. Cell and its massive physical investment put Sony in a multibillion dollar hole.

There was no lack of skill, just a different philosophy of design.
As opposed to the other Sony in-house high-performance compute designs?

For IBM, Cell was simply an experiment. A collaboration on homogenous computing with a couple of partners footing a large chunk of the bill.
It was a way to get someone to pay for the circuit techniques that went into POWER6, a design that was proof that IBM wasn't immune to architectural missteps.
If Cell were homogenous, Toshiba would have walked.

Not terribly different from Intel with Terra Scale and Larrabee, only without the partners to share the costs.
Those were not given the same kind of major project status as Cell. Intel didn't built a billion dollar fab for either.

The SPE isn't what you get from limited experience. It's a conscious design philosophy.
There were multiple instances in the years leading up to that point that showed how outdated that view had become. Perhaps he felt this could be avoided, or perhaps he didn't weight them the same because Sony managed to skate by on what had become unacceptable since the PS2 dominated the market.

I don't think it was so much of maintaining an illusion as it was just simply a lack of interest in that industry.
The hundreds of millions (possibly a billion?) dollars in cash poured into the 65nm Nagasaki fab is not what I characterize as a lack of interest in trying to use Cell to drive itself to a leadership position in consumer electronics and beyond. The design was ludicrously over-engineered if the company was content to plug away at low-functionality devices.
 
Last edited by a moderator:
ah, thanks. I can imagine there a huge amount of work involved for the entire system. It makes me curious about how much they had to touch the scheduler, the network stack, and the i/o hooks, considering there's a CPU in the south bridge.

That is really interesting but unfortunately we will probably never know.

Also, the PS4 OS is probably a direct sequel to the one in Vita.
 
The hundreds of millions (possibly a billion?) dollars in cash poured into the 65nm Nagasaki fab is not what I characterize as a lack of interest in trying to use Cell to drive itself to a leadership position in consumer electronics and beyond. The design was ludicrously over-engineered if the company was content to plug away at low-functionality devices.

It would be interesting to find out how much Sony/SCE spent on fabs. That has to be a lot of money down the drain.
 
It should be pointed out, Mark Cerny was invited to talk about his life and work, and to receive an award, like Peter Molyneux before him, hence his presentation was unavoidably self-centered ("...i began my career..." etc).

Welcome to the Forum :)
 
It would be interesting to find out how much Sony/SCE spent on fabs. That has to be a lot of money down the drain.

IIRC Nagasaki Fab 1 and 2 were built from the ground up for the PS2 (1999 and 2001 respectively), and cumulatively I recall it about $2-2.8 billion being spent. Of course that was just the GS (and EE+GS) fabrication (not including the initial batch of GS from Kokubu). The standalone EE was at Oita TS (as was the RSX).

Cumulatively, I have no clue how much was spent of fabs. Of you want to go back to the start, Kokubu was the first in 1973, followed by Oita in 1984, then Nagasaki in 1987 (acquired from Fairchild/Fujitsu before Sony turned it into the beast it is today), and finally Kumamoto in 2004 (I think). Start your calculator... :)
 
It should be pointed out, Mark Cerny was invited to talk about his life and work, and to receive an award, like Peter Molyneux before him, hence his presentation was unavoidably self-centered ("...i began my career..." etc).

Do you by any chance know if he's going to talk about the same thing in the Develop keynote ? (i.e., Time to triangle)
 
Back
Top