The Engineers Who Created Cell

[one]

Awesome read. 8)

I knew Sony and Toshiba were more involved than just saying to IBM: "Build us something."

It's interesting the tensions there clearly must have been. I wonder how 'pure' versions of each of the visions would have turned out - and I wonder how the seeming hybrid Cell compares to those 'pure-plays.' But, I have a lot of respect for all the engineers who took part in this.

Yes, apparently there's even a tension between Sony and Toshiba, over the EE project, so it's very reasonable that Kutaragi decided to call in veteran IBM to ensure the success of the CELL project.

http://www.careerjournal.com/myc/survive/20000810-guth.html
The story of Mitsuo Saito

At Toshiba, some engineers had never had faith in Mr. Saito. A common view, according to one engineer, was that "if Saito is involved, in some sense, a project is under a bad star." Mr. Saito doesn't dispute that he had this reputation among some colleagues.

As the months dragged on, Mr. Kutaragi at Sony also began to have doubts. His meetings with Mr. Saito turned into tongue-lashings over delays. In the midst of one of these sessions in mid-1998, recalls Mr. Saito, Mr. Kutaragi raised the memory of their mutual friend to express his displeasure: "Ohashi-san was great, wasn't he? He gave it everything he had and never was late."

"I'm sorry, but he is gone," Mr. Saito says he replied. Mr. Kutaragi says he doesn't recall the exchange.

Despite the unpleasantness, progress did slowly come. Mr. Kutaragi made some important technical concessions, such as increasing the power supply the PlayStation 2 would offer the Emotion Engine. The concessions allowed Toshiba to deliver the chip just in time for a March 1999 advance unveiling of the PlayStation 2 in Tokyo.

It seems it's one of requirements for good project managers to constantly cope with bitching by related parties even in a hard dilemma

 

[Carl B]

I'm really enjoying these things you keep digging up One!

I find the history behind what decisions are made behind the scenes leading up to what technologies are included in the consoles nearly as fascinating as the technologies themselves.

 

[Shifty Geezer]

Many thanks for this, one. This is what B3D is supposed to be about - technology and insights, and not inane bickering about trivial matters.

I'd love to know how the parties, Toshiba and IBM, think of Cell's current design. Do they think the compromise was the best solution, or do they feel their ideas were right and compromising only weakened the design?

Anyway, if you have time I'll be very happy to read more of the articles!
Many thanks again.

 

[ChryZ]

I got to second the praise. Excellent job "one". Ã￾t's most appreciated. Please keep it up.

 

[-tkf-]

Excellent stuff! thanks.

 

[Mikage]

Machine translation:

Part 1

Aim at 1TFLOPS.

At first ..the start-up of the project.., offices of the person in charge of each company had divided into Japan and the United States. There was a limit only by E-mailing and reporting the progress report by telephone each other to unite intentions each other. Then, they opened the meeting of three companies by the pace more once a month in Tokyo, San Francisco, and Austin. They aimed to deepen the discussion by being prepare the answer of "Homework" that each company took home by the last meeting by next time, and exchanging opinions again based on it in a short term. It was one of the important roles to hold a meeting to warm friendship drinking sake together at night after an ardent discussion in daytime. Yamazakigou that attended every time as a member of the member of SCE was known as a gourmet. In the favor that the person in charge of IBM ardently introduced the restaurant to him, he became considerably detailed from an Italian restaurant to the seafood restaurant about famous restaurants in San Francisco.
The value named 1TFLOPS was decided several times while the meeting was repeated as the floating point arithmetic performance at which it aimed with next generation's microprocessor. It is a figure that hits about 160 times 6.2 GFLOPS that Emotion Engine has. It was clear to have to adopt the multi core composition in which two or more CPU cores were accumulated to achieve this as anyone can see. Toshiba ..under the assumption.. and IBM presented each idea.
It is architecture that the development formation of Toshiba proposed that is called "Force System" that had been percussed to SCE before cooperation with IBM starts. A large amount of CPU cores as simple as "R3000" that U.S. MIPS Computer Systems Co. developed in 1988 that fell on at the dawn of RISC were accumulated, and the point to integrate this in CPU core of the main was a feature.
On the other hand, IBM insisted on the structure to accumulate two or more microprocessors "POWER 4" for the efficient server machine that had been ending the design at that time. It is Jim Kahle of IBM Fellow, Broadband Processor Technology, and Microelectronics Division who was the person in charge of development of POWER4 for SCE and Toshiba to explain the idea.
On the other hand, it is a simple CPU core at RISC chip level before ten years or more. The other is state-of-the-art POWER4 that has CPU core where the issue of five instructions simultaneously is possible. The same multi core composition had the difference such as water and oil in the idea of IBM and Toshiba.

Collision of belief

The event happened while examining the idea of the two companies closely. The cause was only the idea that a member of Toshiba started in a mouth without what mind in a meeting opened in Austin. "There can be a composition in which CPU core of the main is not accumulated either. " The complexion of gentle Jim Kahle usually changed as soon as he heard the words.
"We do not permit such a thing! "
And the conference room quieted down at the sudden change of Jim. Engineers of SCE and Toshiba were just bewildered. The gloomy atmosphere was not able to be wiped to the last minute though someone started arbitrating and the discussion restarted.
If CPU core of the main is not accumulated, an important place of the Power architecture is lost. Jim was not able to admit this at all. The Power architecture was his child for Jim that had spent the life on the design of the microprocessor in IBM. If the Power architecture that is the basis of the microprocessor business is not adopted, the meaning that IBM participates in the joint development is almost lost.
Is a simple core arranged or is complex cores combined?
The engineers' discussions over the proposal of the two companies were offered belief
each other afterwards and continued In the E-mail exchanged every day, in the meeting that attends together..

 

[Fafalada]

If the Power architecture that is the basis of the microprocessor business is not adopted, the meaning that IBM participates in the joint development is almost lost.

So as a concession they decided to dump VMX (and its bastardized versions) on us...? oh happy joy

 

[one]

Part 3 excerpt

+ The basic architecture of Cell shaped up in the fall of 2000. It's unprecedent that APU has no cache. Many in the development team doubted the usefulness of the tiny 128k dedicated memory called Local Store. But Takeshi Yamazaki of SCE insisted that realtime response is essential for games while cache interferes with it, as a result LS was adopted for APU. Then the APU ISA was discussed between generalized VLIW and object-code-efficient SIMD, at a hotel in NY. Peter Hofstee of IBM succeeded to persuade Toshiba engineers into SIMD.

+ Masakazu Suzuoki of SCE proposed about 200 instructions for APU based on the experience of EE VU at a meeting at Austin. In the room, among engineers in their 30-40s, one old man had been writing something on a paper. He was Marty Hopkins, one of the architect of IBM 801 RISC machine. What he wrote on a paper was a sample program written in machine language and he argued that 200 is too many, 100 is sufficient for a compiler to work. Younger engineers verified it with simulation. Actually the definition of APU ISA was what he chose for his last job at IBM.

+ The blueprint for the Cell architecture was completed in early 2001. It was 4-chip multi-chip module @ 4GHz = 32 SIMD APU = 1TFLOPS. It was decided that the development center would be established at IBM Austin campus. SCE and Toshiba began the selection of engineers. Hearing the news, Eiji Kasahara, who was once an NEC engineer working on mainframes and later designing microcontrollers at Sony America, jumped at the interview with SCE. Hiro-o Hayashi of Toshiba was invited by Mitsuo Saito. 10 years before he rejected an offer from Saito about the UltraSPARC design, so this new chance was big for him though he was hesitant to leave his then project. In reality his senior kindly allowed him to get an air ticket to Austin. Finally STI Development Center was opened in April 2001.

Part 4 excerpt

+ Haruyuki Tago of Toshiba, who managed the EE core development in the US, organized newly arrived Toshiba engineers in Austin. At first Toshiba and SCE engineers were handed Contractor ID cards but they were upgraded soon to allow their entrance after 19:00 PM. Then they requested IBM sysadmin to construct a work environment in which they can connect to intranets of their own companies in Japan.

+ IBM also began the selection of engineers to join the Cell project. Ted Maeurer, once middleware developer for mainframes, organized software engineers in IBM to create development tools and simulators. He got 50 applications a week. Michael Day, who designed IBM Convertible world-first laptop PC, AIX 3.0/4.0, encouraged Maeurer by comparing the Cell project with the importance of IBM Convertible.

+ The development was geared up in the summer of 2001. The evaluation of the Cell blueprint for mass production revealed a problem about power supply for 4-chip MCM. Suzuoki feared for how Kutaragi would respond to this news. However, when Suzuoki told they'd give up MCM, he just said "I see. There you go." Instead the team started the evaluation of multi-processor model for the 1TFLOPS mark.

+ The multiprocessor model spawned the need of fast interconnect technologies. IBM's server technologies were evaluated, but eventually Rambus Yellowstone & Redwood were adopted because of their affinity with cheap PCB.

+ The number of APU was reduced to 6 from 8 as it was too big for the target cost. The development team progressed the design process with 6 APUs as a precondition. They'd thought they already informed Kutaragi about this decision in a regular report.

 

[Titanio]

They'd thought they already informed Kutaragi about this decision in a regular report.

Ruh Roh!

Cheers one

 

[Fafalada]

Then the APU ISA was discussed between generalized VLIW and object-code-efficient SIMD, at a hotel in NY. Peter Hofstee of IBM succeeded to persuade Toshiba engineers into SIMD.

As I always figured - Deano is a covert agent for IBM's secret ops and "real men do it vertically" is their secret moto

and he argued that 200 is too many, 100 is sufficient for a compiler to work.

Why 100, can't real men do it with just 3?

 

[ERP]

Why 100, can't real men do it with just 3?

When I was at school the single instruction processor was in vogue as the ultimate risc solution

 

[one]

By the way, Kutaragi himself drew the PS3 board layout diagram and handed it to the design team engineers

 

[V3]

They aimed for the sky only to get less than a fourth of it. Considering the money that had been poured into the project, you thought they get better team of engineers, that can get the job done

 

[PC-Engine]

By the way, Kutaragi himself drew the PS3 board layout diagram and handed it to the design team engineers

I'm sure it takes a real good engineer to do that...I mean just look at how complex it would be with two chips and some memory.

Must be the asthetic thing...

Now if only KK drew the industrial design to the PS3 controller, we'd be all set.

 

[wireframe]

So the original plan was a CELL with only SPEs?

If I read it correctly that was Toshiba's suggestion. IBMs was an all Power core. They sort of compromised in the middle. It seems that both Toshiba and IBM were very active in CELL.

Interestingly, the XeCPU sounds similar to the IBM idea. Obviously it is far short of 1TFLOPs, but I do wonder if the core idea/concept/design began life as IBM's "CELL" proposal and later that idea was sold to MS?

I read it like this:

1. Toshiba has a concept that is more-or-less what Cell is today: a master core surrounded by simpler cores that the master core controls/delegates to. They go shopping for a core at IBM.

2. IBM wants to leverage its POWER4 design and proposes just multiple POWER4 cores (a la Xbox 360, but the MS piece is also probably more customized that what IBM originally proposed).

3. When Toshiba hears this they make some statement to the effect that they don't want "just a POWER4" core. They want something specialized.

4. This upsets IBM.

5. Compromise: Alter the POWER4 core slightly and let it serve as a master core and slap on the SPEs (which are the simpler cores Toshiba envisioned in the first place).

So, we end up with Microsoft accepting the original IBM proposal for "Cell" (the general idea, perhaps not the details) and Toshiba/Sony compromising by not getting a whole new custom made master core and using POWER4 to some extent in this, but adding an array of micro-cores (SPEs).

In childish words, it reads like this in my head:

Toshiba - "We have a fantastic idea for a special CPU for entertainment that uses a central core and several helper micro-cores to achieve great results. We would like your help in finalizing the design and manufacturing."

IBM - "Why would you want micro-cores when when can glue together POWER4 cores and you will have full capability in all of them? There's no need for micro-cores." (Probably not inteded as an insult, but insults the intelligence of the Toshiba engineers who came up with the idea nonetheless. IBM is obviously looking to sell what they already have).

Toshiba - "Well, maybe we don't even need a master core then..." (suggesting that if IBM cannot work with their overall concept, maybe they are the wrong people to talk to about the master core. They are not making a design philosophy change here, the way I read it).

Looks to me like Toshiba/Sony (Kutaragi) got almost exactly what they envisioned from the start. Perhaps they had other details in store for the master core and the micro-cores, but compromises always have to be made when going from concept to reality. So the forebodings of Toshiba, specializing in customized ICs, having to work with a partner that targets servers (with POWER4) were justified. IBM tried to sell them a "server chip" and Toshiba/Sony wanted a specialized chip to fulfill their vision.

EDIT:

Perhaps I read your comments incorrectly. I would definitely lean to the side of the original concept having a much smaller master core, primarily a governing logic and not a fully fledged GP core. So, don't think they actually ever thought of "only SPEs," but they may have ended up with a heftier master core than they had originally planned for.

 

[PC-Engine]

Perhaps I read your comments incorrectly. I would definitely lean to the side of the original concept having a much smaller master core, primarily a governing logic and not a fully fledged GP core. So, don't think they actually ever thought of "only SPEs," but they may have ended up with a heftier master core than they had originally planned for.

I read that the original plan was to use a MIPS core as the main core with SPEs dangling off of it.

 

[Carl B]

One I love these posts of yours! 8)

@wireframe: I think I see s imilar progression as you, but Microsoft wouldn't have received the 'rejected' Cell design so much as have a team of engineers that came to a similar conclusion, since there was zero communication between the two teams or through IBM at higher levels.

Besides, the actual XeCPU was designed by a third party firm I believe, and brought to IBM for implementation.

EDIT: I also wonder how the Cell would have turned out without Power or SIMD, two pretty fate-determining additions. Would we be seeing things like a Linux port?

 

[wireframe]

@wireframe: I think I see s imilar progression as you, but Microsoft wouldn't have received the 'rejected' Cell design so much as have a team of engineers that came to a similar conclusion, since there was zero communication between the two teams or through IBM at higher levels.

I'm not saying that Microsoft got "leftovers". I think it is more realistic to look at it through the way I think things went down with Toshiba: IBM wants to push PowerPC/POWER4. If you don't have any great ideas walking in, this is what you walk out with. It requires no specific communication between teams because this is what they have on the shelf to sell.

Furthermore, and this may just be your wording, but I see "Cell" as a Toshiba/Sony part. IBM didn't have this concept, they just helped finalize it and build it. So, IBM never proposed a "Cell" design, they proposed something else, which Sony rejected. (And, again, which MS accepted)

I'm not sure about the design part for XeCPU. I don't think they "walked in" knowing the final design. The were probably offered Power, just like Toshiba/Sony was, and they accepted. Then they made some alterations to the design to improve clocks/yield/costs. IBM cannot have been in a position to sell Cell (sing it!) because the design is not theirs and Microsoft has no grandiose plans like Cell of their own.

BTW, I am really curious about why Microsoft decided to go with an IBM part. Lucrative manufacturing deal? Sour grapes elsewhere?

 

[Carl B]

I think Microsoft felt burned and trapped by the x86 manufacturers from XBox 1. Also IBM is willing to make a core for your company; is Intel? We see Apple adopting what's already on their roadmaps to begin with afterall.

And no I completely agree with you... I see Cell as something that Sony and Toshiba brought IBM in on to assist with, but also as something that at it's heart is still Sony and Toshiba's creation. It kind of upsets me how English-speaking media is always focused on 'IBM's Cell' chip, just because they think of IBM as traditionally being a 'computer' company.

But right, I wasn't saying Microsoft got 'leftovers' either; I agree with the fact that it's not surprising two seperate teams of IBM engineers would have reached similar conclusions.

As for the Microsoft chip design however, there was an semi-obscure article a month or two ago that got some decent coverage discussing the firm Microsoft went to to conceptualize their CPU. I'm sure it was IBM's nudge that put the Power 4 in there as well, but I wish I could find this article or the company name.

EDIT: I may be confused about the whole CPU design thing after searching non-stop since posting for the mentioned article. I think I may have somehow wrapped the whole Astro Labs/Hers Experimental Design Laboratories 'designing' of the 360 into prelim work on the CPU.

Well, I'll keep searching of course! 8)

 

[phat]

+ The basic architecture of Cell shaped up in the fall of 2000. It's unprecedent that APU has no cache. Many in the development team doubted the usefulness of the tiny 128k dedicated memory called Local Store. But Takeshi Yamazaki of SCE insisted that realtime response is essential for games while cache interferes with it, as a result LS was adopted for APU.

This Yamazaki must not have done much systems programming. That realtime-ness excuse is lame. Only if you have a hard realtime problem that you barely have enough CPU to solve would the inherent quasi-non-determinism of a cache be an issue. But the Cell was targeted to be so powerful, and the problems in games are so often pared-down or rephrased to fit within a given compute budget, that a deterministic local store at the expense of a cache is very arguably a bad decision. As a systems programmer, 128kB of cache is a lot more attractive to me than 256kB or even 512kB of plain local storage.