Did they collab with Jim Kahle?
STI leadership (especially Kutaragi) collabed with Kahle to hash out CELL's particulars. AMD wasn't involved.
Thanks for posting the Su interview/CELL commentary in the thread. There’s a really old interview (originally posted on B3D, posted on GAF a couple days later) with Su and Kahle that you may not have seen. It discusses the CELL project from their perspective. I'll repost it below, make it more orderly and bold some parts I found interesting...
A Cell of a chip
Key players weigh in
Lisa Su, vice president of technology development and alliances, STG, and Jim Kahle, IBM Fellow and lead architect for Cell, STG, have been involved with the Cell project from day one. Here are their thoughts on some significant aspects of this announcement.
Project history
Lisa Su: We announced the Austin design center with Sony and Toshiba in March of 2001, and before that we spent about a year defining the architecture. It was unique in that we had the ability to start with a clean sheet of paper, under the direction of both Ken Kutaragi of Sony and John Kelly of IBM. ISSCC marks the first time we are talking about the technical details of the project.
Jim Kahle: With the clean sheet, we stepped back and looked at a number of aspects of the Cell, such as chip design, system design and power management. The partnership allowed us to think outside of the box. And the consumer electronics link gave us a new level of human interchange, and a new way of designing computers.
Key details
JK: The magnitude of the multi-core design is very important. Each cell chip has nine processors on board, and with the central Power core having two threads, it's more like a 10-way processor model. We optimized the whole system architecture around multi-core, taking the chip to the next level of performance.
LS: Cell represents a confluence of lots of technologies. For instance, virtualization allows multiple operating systems to run at the same time, including open operating systems such as Linux. In designing Cell, we employed techniques that were developed over many years for use in our high-end Power servers, but with the advances of semiconductor technology, we are able to bring these techniques into consumer devices. This will enable a whole new set of applications that have yet to be developed.
Prospective uses
JK: Cell's synergistic process elements, or SPEs, are able to handle large amounts of graphics, improving media processing capability. In addition, the chip will be able to manage compute-intensive type applications.
LS: Cell will enable a wide range of applications, such as next-generation game machines, home servers and digital media servers. A smaller version of Cell could conceivably be used for HDTV, and a larger version for content-creation workstations for the entertainment industry or even scientific computing applications.
Chip speed
JK: We've seen speeds of Cell greater than 4 GHz in the lab, and speeds of isolated SPE as high as 5.6 GHz. We approached the design with the idea of attaining high-performance frequency.
LS: You might want to think of Cell as the world's "most powerful" chip. And although frequency is important, even more important is the parallelism and being able to run on nine cores allows the ability to do a large number of operations simultaneously.
Power Architecture
LS: It was important for Sony and IBM to start with a mature architecture for the general purpose microprocessor core so that we could build on a stable programming base. And since IBM's strategy is to have an open Power Architecture, this is an excellent example of how the Power Architecture can be leveraged into new markets such as the gaming and entertainment industry.
JK: Power Architecture is a great base for conventional processing and we extended its capabilities for extreme media processing and real-time behavior for human interactions. It adds a whole new dimension on how to program.
Production
LS: Cell will be a 90 nanometer SOI chip with more than 230 million transistors, and will be produced both in the 300mm fab in East Fishkill and Sony's 300mm fab in Nagasaki. We'll start production in 2005.
JK: As far as OEM is concerned, we'll approach the industry through IBM Engineering & Technology Services for third-party services and applications, and we have already received significant interest.
Growth opportunities
JK: Cell technology will allow IBM to go into new and emerging areas, and show leadership in the gaming area. This is a tremendous opportunity for all three companies.
LS: Cell uses the strength of all three partners and it starts with IBM's leadership silicon-on-insulator technology and microprocessor design expertise, combined with Sony's market and system insight and Toshiba's semiconductor experience and to apply this technology in new markets, such as digital media, entertainment and consumer electronics. It is a great example of the "sum is greater than the individual parts." Sony is one of our deepest technology partners and we've spent a combined $400 million in joint microprocessor development with Sony and Toshiba, Sony and Toshiba are partners in our SOI process development alliance, and Sony has invested $325 million in the fab in East Fishkill.
Confluence of technology
LS: We often refer to the confluence of technology in Cell. That includes features like high frequency, parallel multi-thread, efficient architecture and state-of-the-art virtualization. When you think about it, it's pretty amazing to have this type of chip and this type of processing power going into a game system.
JK: And we've also added autonomic power management, real-time resource management for human interaction and smart memory flow controllers.
source
Any idea what prevented Cell from having more die shrinks?
Sony will eventually need to retire those old PS3 motherboards on PS Now/PS+ Premium:
Do they still manufacture Cell/RSX chips in 2024? I strongly doubt it, but maybe I'm wrong.
CELL is probably already at 14nm, with a possible path to 7nm. Here comes another detailed spiel...
2012 - 2014 - IBM supplies SIE with 22nm SOI CELLs. SIE builds PS3 server boards with them to prepare for PS Now's launch in 2014.
2014 - IBM
sells its semiconductor business to GlobalFoundries. GF becomes IBM's supplier of 22nm SOI
POWER8 chips with the goal of 14nm and 10nm chips down the road.
2017 - GF and IBM
jointly develop a 14nm FinFET/SOI (14HP) process. IBM uses it for
POWER9 and z14 CPUs
2017 - GF skips 10nm,
announces 7nm (7LP) FinFET which is said to have a high level of overlap with the 14HP process
2018 - GF halts further 7nm development indefinitely to focus on 14nm/12nm essential and specialty chips
2020 -
Kenichi Nakano (General Manager of Sony Semiconductor Solutions managing GNSS products,
unique SoC products, and LTE related products) chooses GF's 22nm FDX process for Sony's GNSS chips. Selecting GF (now owner of IBM's SOI process) gives SIE the ability to extend Kutaragi's CELL
roadmap to 14nm (14HP) FinFET SOI and 7nm (7LP) FinFET bulk, which is just waiting for an investment in what's left of its development to create a 7nm FinFET SOI (7HP) process.
At this stage in its life, CELL is a specialty chip (only used for the PS Now platform) that's always been a unique
SoC. In '13, Masayasu Ito said that SIE could
“freely manufacture CELL if the decision is made that it is needed”, but support components
"will become difficult to obtain since 7 years is already considered to be long in the IT industry". Surely, some of those components must be getting nigh on impossible to obtain now. I think component scarcity, PS Now's ageing fleet of 22nm PS3 servers and a desire to have more CELLs per wafer drove SIE to quietly redesign and migrate CELL (and likely the RSX) to GF/IBM’s 14HP process at some point in ’21. SIE probably has stockpiles of 14nm PS3 replacement servers ready to deploy.
Imagine a dirt cheap single-die PS3 at $99 or even less... would it ever be possible? (considering the fact PS3 had 2 separate memory buses/technologies: XDR + GDDR3, unlike XBOX360 having a unified GDDR3 memory)
There was a time when I thought SIE would make a $99 - $149 PS3 attachment, but no longer. Not because of any real or imagined technical challenges. I've just come to the conclusion that SIE has moved on from PS3 hardware, save for PS Now.
The only PS3s SIE cares to make are the ones fit for PS Now server boards.
God of War Ascension looks even better than GoW 3:
I mean, take a close look at Kratos' textures.
I can't believe they achieved such a texture quality with only 256MB of VRAM...
More like 512MB of DRAM + VRAM. Contrary to popular belief, PS3’s memory is unified. It’s just physically split between XDR and GDDR3. A
whitepaper by two SCEA researchers explained:
"The system's unified memory architecture allows the Cell/B.E. and GPU to exchange data through shared textures"...
"This system combines the Cell/B.E. with a state of the art GPU in a unified memory architecture. In this architecture both devices share access to system memory and to graphics memory. As a result they can share data and processing tasks". The researchers went on to describe PS3's architecture as a
"hybrid real time rendering system".
The best devs exploited PS3's hybrid real-time rendering system to their utmost. Santa Monica Studio was certainly one of the best…
Also, CELL’s
heterogeneity and simultaneous access to XDR and GDDR3 memories made PS3’s hybrid real-time rendering system a HSA with hUMA long before AMD made the terms
'HSA' and 'hUMA' marketable. PS3 was a marvel of engineering.
Do you guys think that the PS6 could perhaps have a beefy Zen 4/5 8-core CPU with full AVX-512 support @ 3.5 GHz?
RPCS3 runs a lot smoother with AVX-512:
I wouldn't rule it out, but geez. Wouldn’t that be a ton of heat to dissipate and exhaust from a console form-factor? That PS6 would probably need liquid metal and dual fans whistling/blowing like a
KLM at takeoff power to keep cool.
I doubt SIE can do much better than the inconsistent RPCS3, despite everything they know about CELL. Jacob Stine (of PCSX2 fame, former SIE, now Implicit Creations) filed for an emulator that uses a JIT compiler to remap the functions of a target (e.g. a SPU core) to a host CPU (e.g., a Zen 2 core).
Entry
[0043] reads:
The Target System may include processor units which are incompatible with the Host CPU 400, for Example Vector Processing units or Synergistic Processing Units which will have their functions remapped for the architecture of the Host CPU core 400 by the JIT compiler during emulation of the target system.
SIE’s been sitting on this for 7 years (there's that number again), yet we haven't heard peep nor squeal from them about PS3 backwards compatibility. I think SIE is mum because Zen 2 can’t emulate CELL to a satisfactory degree, whatever SIE deems satisfactory.
That’s why I gave up hope of a PS3 emulator for PS5 a long time ago. Zen 4/5 doesn't change that for me.
I think a redesigned, integrated 14 or 7nm CELL that doubles as an AI accelerator or other special use processor is the surest way to PS3 backwards compatibility on PS5 Pro and PS6, should SIE want them to have the feature.