News AMD CEO Lisa Su reminisces on helping design the PS3's infamous Cell processor at IBM

[Admin]

Lisa Su reminisces on when she helped design a processor for the PS3 while working at IBM.

AMD CEO Lisa Su reminisces on helping design the PS3's infamous Cell processor at IBM : Read more

 

[bit_user]

Cell was pretty amazing, for its time. IMO, the biggest problem it faced was a dearth of good programming frameworks for using it effectively. If OpenCL had existed, at the time, that probably would've helped immensely.

In fact, I'm not the first to think this. IIRC, George Hotz (of recent notoriety involving his Tiny.AI startup) was experimenting with OpenCL on the PS3, when Sony withdrew OtherOS support and disrupted his efforts. If I'm not mistaken, that's what sent him down the spiral of legal troubles with Sony.

(in 2001) We [IBM] started with a clean sheet of paper and sat down and tried to imagine what sort of processor we'd need five years from now." The decision that IBM, Sony, and Toshiba made was to create a CPU with an extreme focus on parallelization.

A key detail to remember is that GPU compute was in the earliest days of its infancy. GPUs were still incredibly specialized towards graphics and trying to use them for anything else was very hard and subject to heavy constraints. So, I'd say they were spot-on in identifying the trend, but just off the mark about which element of the system was best-suited to address the needs.

The interviewer points out that Sony's PlayStation 3 is viewed as one of its least successful consoles, which is true.

Um... according to wikipedia:

Worldwide sales: 87.4 million (as of March 31, 2017)
​

Okay, so the PS4 has (also, according to them) now sold "106 million (as of December 31, 2019)", but I'd argue the Total Addressable Market (TAM) for consoles grew by more than that amount, between when the PS3 was Sony's leading console and when the PS4 was. So, if we grade on a curve, I'd say the PS4 was probably the loser.

The Switch also saw Nintendo pivot to a fully Nvidia-powered SoC design for their new hybrid console focus.

The Nintendo Switch's CPU cores were designed by ARM. Nvidia made the SoC and iGPU.

 

[jeremyj_83]

The Nintendo Switch's CPU cores were designed by ARM. Nvidia made the SoC and iGPU.

Apparently AMD made a play to be the CPU/GPU in the Switch 2 but Nintendo went with Tegra again.

 

[TheyCallMeContra]

​

Worldwide sales: 87.4 million (as of March 31, 2017)​

​

Okay, so the PS4 has (also, according to them) now sold "106 million (as of December 31, 2019)", but I'd argue the Total Addressable Market (TAM) for consoles grew by more than that amount, between when the PS3 was Sony's leading console and when the PS4 was. So, if we grade on a curve, I'd say the PS4 was probably the loser.

Well, do keep in mind that the PS4 dramatically outperformed the Xbox One despite 360's lead in the previous generation. If we only look at console sales within the scope of the manufacturer alone, Nintendo and Sony definitely peaked with Nintendo DS and PlayStation 2, respectively.

Also, good catch on Arm involvement! Will submit an edit for that one soon.

 

[jeremyj_83]

Well, do keep in mind that the PS4 dramatically outperformed the Xbox One despite 360's lead in the previous generation. If we only look at console sales within the scope of the manufacturer alone, Nintendo and Sony definitely peaked with Nintendo DS and PlayStation 2, respectively.

Also, good catch on Arm involvement! Will submit an edit for that one soon.

The Nintendo Switch has sold over 140M units. It is within a Wii U of sales of tying the DS for total sales with probably 12 months before its successor is released. Very hard to say that they peaked with the DS.

 

[TheyCallMeContra]

The Nintendo Switch has sold over 140M units. It is within a Wii U of sales of tying the DS for total sales with probably 12 months before its successor is released. Very hard to say that they peaked with the DS.

"Within a Wii U of tying with DS".

Being an entire console's install base away from being equivalent to Nintendo's current peak? Objectively speaking, they did peak with Nintendo DS. You do understand the literal meaning of the word "peak", yes? You can hardly forecast in good faith that the current Switch will surpass NDS when the existence of Switch 2 is widely-known and an entire competing market of Deck + other PC handhelds exists.

Also we know it's March of 2025 now, so it's more like 9 months.

 

[jeremyj_83]

"Within a Wii U of tying with DS".

Being an entire console's install base away from being equivalent to Nintendo's current peak? Objectively speaking, they did peak with Nintendo DS. You do understand the literal meaning of the word "peak", yes? You can hardly forecast in good faith that the current Switch will surpass NDS when the existence of Switch 2 is widely-known and an entire competing market of Deck + other PC handhelds exists.

I do know what peak means but thank you for the condescending tone. Do you have any idea how few consoles the Wii U sold? The Wii U sold all of ~13M units with the Switch at 141M after 7 years it very well could catch and surpass the DS. Even after the PS5 and Xbox Series X were released people still bought the PS4 and Xbox One. That will keep happening with the Switch as well. The competition from Deck, etc... isn't the same. Not to mention there have been multiple players that tried to enter the handheld market against Nintendo and all failed. Plus Deck and such has been around for years and still only a bit player.

Also we know it's March of 2025 now, so it's more like 9 months.

Do we really know it is March 2025? Everything for the Switch 2 release date is speculation. Nintendo has only said they "will announce the new switch within its current fiscal year, which ends in March 2025." That doesn't mean lauch in March and could very well mean it isn't lauchend until the holiday season 2025.

 

[TheyCallMeContra]

I do know what peak means but thank you for the condescending tone. Do you have any idea how few consoles the Wii U sold? The Wii U sold all of ~13M units with the Switch at 141M after 7 years it very well could catch and surpass the DS. Even after the PS5 and Xbox Series X were released people still bought the PS4 and Xbox One. That will keep happening with the Switch as well. The competition from Deck, etc... isn't the same. Not to mention there have been multiple players that tried to enter the handheld market against Nintendo and all failed. Plus Deck and such has been around for years and still only a bit player.

Forgive the condescending tone, but you're trying to correct me when I'm already right. You're still making a forecast to try and disprove an argument that is currently true and has been for quite a long while. If Switch actually has another 13 mil in it, come back to this thread when it happens and I'll CashApp you a $5 or something. I seriously doubt anyone who wants a Switch doesn't already have one or isn't just waiting for the better version to drop next year, though.

re: Switch 2 release...fair I guess, but I'd be seriously shocked if they didn't meet the March 2025 release window, in alignment with the first Switch. that was leaked before Nintendo made official comment borderline confirming it.

 

[JamesJones44]

re: Switch 2 release...fair I guess, but I'd be seriously shocked if they didn't meet the March 2025 release window, in alignment with the first Switch. that was leaked before Nintendo made official comment borderline confirming it.

I know they launched the Switch in March but a lot has changed since then and companies are targeting year Q4 a lot more these days (year Q4, not company Q4). If I could bet on it, I would put money down on September/October release, but it's truly anyones guess at the moment.

 

[hotaru.hino]

According to Lisa Su, then-Director of Emerging Products at IBM in 2001, "We [IBM] started with a clean sheet of paper and sat down and tried to imagine what sort of processor we'd need five years from now." The decision that IBM, Sony, and Toshiba made was to create a CPU with an extreme focus on parallelization.

I feel like the late 1990s to early 2000s was an interesting point in computer history as far as the server/supercomputer sector was concerned. There were like 5-6 competing ISAs still vying for this market during this time. So a missing detail here is while the Cell was made famous by the PS3, I think what the STI alliance wanted was to be the top dog in the supercomputer space and the PS3 was just a mechanism to get back some of the money they threw into it. And the Cell was the top dog in the supercomputer space for a few years. But by that point, the GPGPU concept took off, so there wasn't really a point in the design philosophy of the Cell anymore.

Today, that approach is fairly common through multicore CPUs, Simultaneous Multi-Threading (SMT, or Hyper-threading under Intel marketing), and even dedicated Efficiency cores, but SMT wouldn't emerge until 2002, and the first consumer multicore CPUs from AMD and Intel wouldn't be seen until 2005. And, of course, the first-ever multicore was released by IBM for workstation and server use in 2001— the same year they were planning the PS3's Cell processo

I don't think it's appropriate to compare the Cell architecture to modern CPUs. It's hard to call the SPEs in the Cell additional CPU cores because they're isolated from the rest of the system. That is, they can't access main memory directly and everything has to live on its 256KiB of local storage. Plus they couldn't run the same programs as the main PPE part of the processor. The SPEs were meant to accelerate parts of an application, but not run whole applications themselves.

The closest thing I'd argue that PCs have with this are AMD's APUs and Intel's CPUs with integrated graphics.

 

[NeoMorpheus]

Apparently AMD made a play to be the CPU/GPU in the Switch 2 but Nintendo went with Tegra again.

Evil loves evil.

View: https://youtu.be/m2KV8MHRJlQ

Well, do keep in mind that the PS4 dramatically outperformed the Xbox One despite 360's lead in the previous generation. If we only look at console sales within the scope of the manufacturer alone, Nintendo and Sony definitely peaked with Nintendo DS and PlayStation 2, respectively.

Also, good catch on Arm involvement! Will submit an edit for that one soon.

The real reason why the PS4 overtook the Xbox One (stupid name) was the forced kinect, higher price due to the kinect and some really stupid comments from the then head of Xbox.

By the time they corrected those mistakes, it was too late.

 

[bit_user]

I feel like the late 1990s to early 2000s was an interesting point in computer history as far as the server/supercomputer sector was concerned. There were like 5-6 competing ISAs still vying for this market during this time.

Yeah and not only that, but there were still people trying lots of different approaches to parallel computing. You had vector processors and funky interconnect topologies, NUMA without cache coherency, and other things that are weird by today's standards.

So a missing detail here is while the Cell was made famous by the PS3, I think what the STI alliance wanted was to be the top dog in the supercomputer space and the PS3 was just a mechanism to get back some of the money they threw into it. And the Cell was the top dog in the supercomputer space for a few years. But by that point, the GPGPU concept took off, so there wasn't really a point in the design philosophy of the Cell anymore.

Yeah, exactly. I was thinking about mentioning the HPC aspect. I once gave a job interview to someone who optimized a few HPC apps to run on the Cell. He shared a few insights, but it was more than 15 years ago and I can't recall any...
: (

I don't think it's appropriate to compare the Cell architecture to modern CPUs. It's hard to call the SPEs in the Cell additional CPU cores because they're isolated from the rest of the system. That is, they can't access main memory directly and everything has to live on its 256KiB of local storage. Plus they couldn't run the same programs as the main PPE part of the processor. The SPEs were meant to accelerate parts of an application, but not run whole applications themselves.

True. Weird though the Cell was, the PS3 was really the last time consoles had a technical edge on PCs of the time. In terms of raw compute power, it was really out there. That showed, when the PS3 renaissance happened as game devs finally got good at exploiting it. Some PS3 games were beautiful on a 1080p monitor, whereas I never saw a 3D game on the PS2 that didn't look like garbage to me.

 

[hotaru.hino]

Yeah and not only that, but there were still people trying lots of different approaches to parallel computing. You had vector processors and funky interconnect topologies, NUMA without cache coherency, and other things that are weird by today's standards.

This is why I started to look at IA-64 in less of a harsh light now than how most people who are aware of it would normally judge it.

Besides, Intel got almost 20 years out of the ISA in actual deployments. That's got to count for something.

 

[Bar-G]

Lovely article, except some stuff is plain wrong.

Wii U should count as an 8th gen console since it's more powerful than the 360, which was the most powerful system in that gen.
While still being a lot weaker than an Xbox One for example, it should count as a next-gen since it manages to surpass the other older consoles.

On the other hand, while Cell was great at floating-point (really strong SIMD), the 360 had 3x the general-purpose performance of it, which is the most important thing on a processor anyways.

While it was useful to have high-compute on those days since GPUs weren't as flexible as today, Xenon wasn't that bad with its 115.2GFLOPS compared to the 204.8GFLOPS of Cell... all of that SIMD performance while still offering 3 general purpose cores. I believe that the 360 had the superior CPU here, and not only because it was easier to code for.

The GPU we all know how it went, more modern design (unified shaders) and more raw power (although RSX had better fillrate).

Same goes to the RAM setup, Xbox 360 had 512MB of RAM without the penalties the PS3 had when Cell tried to access the VRAM (slow as hell, and higher latency). Not to mention the eDRAM on the 360 giving it a bandwidth advantage.

Other than that, everything is good

EDIT: Corrected some info.