Cell Processor and Apple
I am very interested in the new IBM/Sony Cell processor that was anounced today. The stats look astounding.
Does anyone have any information on Apples position in using or not using this part?
Are there technical chalanges to it's emplementation?
Is it used in conjuction with other standard chips like the 970?
Does IBM have reasons to not sell these chips to Apple?
In reading the article, it mentiones 16 TF being achieved in one "rack". is this a rack of servers/workstations running together? If so, isnt the VTech supercomputer only clocked at 12TF?
!!!!
Does anyone have any information on Apples position in using or not using this part?
Are there technical chalanges to it's emplementation?
Is it used in conjuction with other standard chips like the 970?
Does IBM have reasons to not sell these chips to Apple?
In reading the article, it mentiones 16 TF being achieved in one "rack". is this a rack of servers/workstations running together? If so, isnt the VTech supercomputer only clocked at 12TF?
Quote:
Quote:
IBM, Sony Corporation (Sony) and Sony Computer Entertainment Inc. (SCEI) announced today that they have powered-on the first Cell* processor-based workstation.
The prototype workstation is the first computing application planned for the highly-anticipated Cell processor.
The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second.
Quote:
IBM, Sony Corporation (Sony) and Sony Computer Entertainment Inc. (SCEI) announced today that they have powered-on the first Cell* processor-based workstation.
The prototype workstation is the first computing application planned for the highly-anticipated Cell processor.
The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second.
!!!!
Comments
But you didn't provide a link to your article, and this subject has not been talked abut in a while so cough it up. (the link)
Originally posted by onlooker
You should use the search first. This has been discussed in length in other threads.
But you didn't provide a link to your article, and this subject has not been talked abut in a while so cough it up. (the link)
http://www.convergedigest.com/Silico...e.asp?ID=13077
http://www.tomshardware.com/hardnews...29_121627.html
Quote taken from Toms Hardware
The Cell processor, a code name of the technology being developed for the past four years, is based on a 64-bit processor core and will be available in single-core and multi-core versions. The firms said that the chip will support several operating systems for consumer electronics and HDTV's. Pilot production of the Cell processor is scheduled to launch in the first half of 2005 in IBM's 300mm wafer facility in East Fishkill (NY) with ramp-up of the processor to follow later that year or early 2006.
Looks like there could already be discussions with a maker of another OS besides MS.
Originally posted by onlooker
Looks like there could already be discussions with a maker of another OS besides MS.
PS3 uses its own OS as does home entertainment. Then there is always linux. It has nothing to do with Apple necessarily.
With the PPC going dual Core they may have incorporated some fascinating new features into it's designs already.
Even if the support cores had an ISA identical to AltiVec's, the difference between their loose coupling and AltiVec's tight coupling to the PPC core would badly break a lot a AltiVec code.
On top of that, there is enough code on the Mac that assumes that the CPU is the SPU (sole processing unit ) that any design that doesn't have a fairly beefy PPC core is going to turn in subpar performance on the bulk of the Mac's applications.
And the final nail in the coffin is the bus. RAMBUS has taken a break from suing everything that can't run away fast enough to supply a 6.4Gb/s bus to the outside world which is compatible with HyperTransport and RapidIO. That's 6.4 gigabits per second, or 800 megabytes per second. So, it weds the latency of Elastic Bus to the bandwidth of MaxBus. I'm sure this is an elegant solution well-tailored to the needs of the PS3. I'm not sure it has any business in a Mac.
But then, I'm more than happy to be corrected, since we've still only been given a 30,000 foot view of this thing.
What could you use Cell for in a Mac?
Well it's an MT (SMT or SoE?) SMP capable
64bit PowerPC, with 8 vector units.
Future 65nm Cell chips will have more
Processor Engines, 2-4 looks the most
probable.
Code will port easily enough, or at least
with no more difficulty than G3/4 to G5.
(If Apple want to do it then it can be done.)
All this assumes Cell is targetted at a broader
range of applications than embedded/games,
so far it appears so.
Apple are already producing Core Video and
Image to accelerate certain tasks with Altivec
fall-backs, they recognise that GPUs have
a huge amount of horse-power to tap.
The Cell APUs could be another target for
the Core APIs. A highly desirable target not
supported by Windows?
So it may lack altivec, but we may not care.
Apple want Intel crushing performance,
Sony might like Apple to produce Workstations
using BE derivatives, nice dev boxes right?
IBM might like Apple to take up Cell too,
rather than requesting POWER6/7 derivatives.
It's too early to tell, but there are many
interesting rumours around. Some of
which are quite amazing.
Originally posted by Amorph
Well, now that I have more than Programmer's teasing to go by, I really don't see Apple taking advantage of this particular Cell processor in a Mac. It's way too different...
But then, I'm more than happy to be corrected, since we've still only been given a 30,000 foot view of this thing.
First of all, the publicly available information is still way too thin to really draw any conclusions. I am not applying any knowledge in this message that isn't publicly known...
It is entirely possible (albeit unlikely) that this thing is a 970-derived core with a bunch of on-chip vector processor cores. I haven't seen anybody point this out yet, but the extreme clock rate quoted for the chip doesn't rule this out -- the main core could be running at 1/3rd or 1/2 the clock of the vector cores and thus be in the ~2 GHz realm of the 970.
This design is actually very similar to the rumour a year or two ago about Apple considering putting vector units into the memory controller. Even if the chip for Sony uses a different core that still doesn't rule out Apple's use of it because it will have the standard collection of PowerPC execution units (including floating point -- the PS2's core has an FPU and 2 vector units so its hard to imagine that they take a step backward). It is even possible that the main core could include AltiVec -- it is a little known fact that the PS2's core has some SIMD and 64-bit functionality built into it. An AltiVec unit could actually be very useful on the main core for prepping data going to and coming from the vector cores.
That it will appear in workstations running Linux and (apparently) other OSes implies some kind of memory management support on, at least, the main core. As a result its not much of a stretch to imagine that Apple could use this thing directly and with a minimal amount of OS changes, just by ignoring the Cell's specialized processing elements. Obviously that is fairly pointless, so if they did do this Apple's engineers would no doubt quickly port Quartz, OpenGL, QuickTime, CoreAudio, CoreImage, and other system services to leverage these extra cores. They could then publish the API needed to access those same cores and 3rd party software would be able to jump on board. They have done this before (with the av series 040 Macs), and considered doing it again (with Transmeta media processors). They are essentially doing it with the programmable GPUs in the current Macs, just in a more specialized fashion.
As catbat partially pointed out, the pin rate was actually quoted as "6.4 GHz per pin" and abused from there by the reporting agencies. This could even be double or quad data rate, we don't know. Either way, there are probably 16-32 pins in the bus so the memory bandwidth could be anything from 10-100 GB/sec, we simply don't know yet. I'm betting it'll be on the higher end of that range, and will leave some jaws hanging open (both in shock at the bandwidth, and horror at the latency).
I guess you'll just have to wait until Feb '05 to know more. Personally I hope Apple can use the technology -- it would be great to mash Intel's face in it.
? LONGHORN ?
One thing I read did send up a few RED Flags
The processor, code-named Cell, will handle vastly more memory than today's consumer chips as well as
<<enable hardware-based copyright protection>> and allow multiple operating systems to run at the same time. It also will feature multiple cores, or logic engines, on a single die.
It's unclear whether Cell will pose a threat to today's most popular processors, Intel Corp. (INTC)'s Pentium 4 and Advanced Micro Devices Inc. (AMD)'s Athlon 64. Both companies have announced plans to include features similar to Cell's in versions planned by Cell's launch.
All computer makers see home entertainment as major driver of next-generation computers and are designing systems that can serve up all types of media - including high-definition TV, standard video and audio - at once.
The consortium behind Cell suggested today's PC chips aren't up to such computation-intensive task, both in terms of processing power and the capacity of the data pipes that connect the processor to the system memory.
"The current PC architecture is nearing its limits," said Ken Kutaragi, Sony's executive deputy president and chief operating officer.
The companies released few details about Cell, saying more information will be divulged at a technical meeting next year in San Francisco. In fact, little information has been released since the $400 million partnership was announced in March 2001 with the promise of building a "supercomputer on a chip."
IBM plans to begin pilot production of Cell during the first half of 2005 at its factory in East Fishkill, N.Y. The first product will be a Cell-based workstation it is developing with Sony. Toshiba said it expects to launch a Cell-based high-definition TV in 2006.
Apparently that M$ Palladium Project may have found a home after all.
The thing is that my main concern is with the installed base of Mac applications. Theoretically, Cell sounds more or less like the ideal Core Image / Core Audio / Core Video accelerator, with the CPU delegating each filter to a support core, and enough pins hanging off the bus to stream the requisite amount of data, while the PPC core handled the application proper and the GPU took Quartz (and some of the Core Image/Core Video work as well). This is multimedia multitasking heaven.
I was also careful to restrict my comments to this Cell processor, because I'm going on the assumption that it's not a full powered 970. Neither consoles nor (traditionally) workstations have required fire-breathing CPUs. It's always been about bandwidth and support hardware, and in the case of consoles a lot of bit-banging (I imagine that by now that's changed in the most literal sense, but not in the sense that developers do a lot of painstaking, low level profiling and optimization for the console's hardware).
However, as far as I can see, the Cell architecture only mandates a PPC, and that means a 440 or a 750 or a 970 or, hell, a 7448. A Cell processor built around a PPC900 series core would indeed change the rules of the game. Especially in a duallie. The number of support cores is also theoretically negotiable (and I suppose the PPC could be dual core, too, down the road...). IBM is all about this level of modularity and customization now, so I can see this happening.
The catch is that Apple has to treat this as the next AltiVec (hey! it's AltiVec 2!). It must not be like the Quadra AV series, because hardly anyone ended up actually using those poor, lonely DSPs. Granted, Apple has an edge today in that their framework-based OS can guarantee that Cell support cores will be used by the frameworks, and thus by the software that can best take advantage of them. Nevertheless, Apple has to get this technology as widely adopted as possible, and wait for the market to be satisfied that it's sticking around so that it's widely adopted.
I'm not surprised to hear that fair use suppression circuitry has made it into Cell: It's being sold to Sony, after all, and Microsoft is an IBM customer as well, so the XBox CPU will probably inherit the same circuitry. But nothing requires Apple to use the particular Cell processors that have that circuitry. If they do, nothing requires OS X to support it. But Steve has been uncharacteristically hard to read on this topic, because he has demonstrated none of the open contempt for foolproof DRM of movies that he has for music. So we might be left at the mercy of this hardware, hoping that Apple's terms are at least liberal. At least, until Congress stops its 25-year addiction to money from Big Publishing (to whom, of course, Apple is also beholden) and remembers what American copyright law is supposed to look like.
Originally posted by FallenFromTheTree
The processor, code-named Cell, will handle vastly more memory than today's consumer chips as well as
<<enable hardware-based copyright protection>> and allow multiple operating systems to run at the same time. It also will feature multiple cores, or logic engines, on a single die.
...
Apparently that M$ Palladium Project may have found a home after all.
Doubt it. More likely it's support for future DVD copyright systems and/or what Sony will use for their PS3 games.
So, today's "real time" will be ancient in a couple years. Thats just the way it goes.
In 20 years maybe we'll have powerful enough homecomputers for "The Matrix" type worlds... but there will be technology in 20 years that will require 1000 times that power to do its job.
viscious circle
Originally posted by ZO
the problem with saying that a CPU will have enough power to do things in "real time" is that every year, technology gets much more complex. A high-end PC with souped up graphics card can render An Ants Life a hundred times faster than what was available when it came out. Today it wouldn't even come close.
I think the significance of the comment is that there are common tasks this year that can't be done in real time this year. There's no point worrying about whether today's tech will scale up five years from now.
And Cell appears to scale up pretty nicely, like most modular architectures. If you want more computational power, add more cells.
There's an interesting story the reg about Power and how IBM plans on putting it everywhere. They also mention Cell: "Cell is expected to be pitched as much at IT applications as its better-known role in Sony's next-generation video console"
Is that good news or what?
Oh the irony: MS, Sony, Redhat, IBM using an IBM processor!
Originally posted by Programmer
I'm betting it'll be on the higher end of that range, and will leave some jaws hanging open (both in shock at the bandwidth, and horror at the latency).
Actually RAMBUS latency isnt all that bad ( when it was introduced it was actually better than its competitors ). The thing about RAMBUS is that latency is dependant on the number of chips connected to a channel. When only one chip is connected latency is very good, but as more are added it gets worse and worse ( unlike SDRAM ). This actually suites consoles really well, because they have a fixed memory allocation, and the manufacturer can balance cost/bandwidth/latency how they want. In a PC things are a little different. Most RIMMs were multi chip ( bad latency right there ), and you could plug a couple of RIMMs into a single channel ( even worse ).
http://www.x86-secret.com/articles/r.../km400en-8.htm
Provides some latency information, but no details as to how many chips are running on the RDRAM system.
Personally Im very surprised that GPUs arent using RAMBUS for memory, they are a perfect match.
Originally posted by mmmpie
Actually RAMBUS latency isnt all that bad ( when it was introduced it was actually better than its competitors ). The thing about RAMBUS is that latency is dependant on the number of chips connected to a channel. When only one chip is connected latency is very good, but as more are added it gets worse and worse ( unlike SDRAM ). This actually suites consoles really well, because they have a fixed memory allocation, and the manufacturer can balance cost/bandwidth/latency how they want. In a PC things are a little different. Most RIMMs were multi chip ( bad latency right there ), and you could plug a couple of RIMMs into a single channel ( even worse ).
Unfortunately the PS2 has a pathetic cache so the latency of its RAMBus memory really kills it.
GPUs typically don't care much about latency since they can fetch way in advance. Cost and bandwidth are bigger issues for them.