German website Heise Online has received confirmation that IBM is terminating its Cell processor line. This means that no future development will take place, making the PoweXCell 8i the last Cell processor. Parts of the Cell project will still make it into future processor designs, however.
Oh, for how long had I read how the Cell processor was going to revolutionize computing. The Cell would turn up in everything form hair dries to super super computers. Sony would turn it Cell driven PS3 into the home media platform of tomorrow. Computing paradise was in our grasp.
But will you look at that. Somehow the whole venture got boffed yet again. Can they simply not learn? It boggles the mind at just how moronic these people at IBM and Sony can be, it’s a wonder they are still in business at all.
Nothing surprising here.
Cell processors cannot do much against a set of GPGPU cards doing parallel processing.
Actually, the CELL cpu was quite slow on most things. For instance, the SUN Niagara at 1.4GHz was more than 13 times faster than the 3.2GHz CELL at string pattern matching, if the working set is large. If the working set is tiny, then the 3.2GHz CELL cpu is roughly 50% faster than the 1.4GHz Niagara. But who is interested in tiny working sets, in real life there is always large working sets.
http://blogs.sun.com/bmseer/entry/extremely_fast_pattern_matching_o…
Are you kidding ?
You gave us a Sun blog entry to back up your numbers ?
Dude, if I learnt something since the release of Solaris 10, it’s that you should never trust Sun.
What happened? Did a Sun field engineer eat your sandwich?
Could you please tell us what you know, that we dont? Do you mean that the SUN blog lies about those numbers in the benchmark, the blogger just made up the numbers? When he claims that one Niagara 1.4GHz achieves 24.6 GBit/sec and two Cell 3.2GHz achieves 3.8GBit/sec – it is just lies? In reality, those two CELLs achieved maybe 50GBit/sec? And the Niagara achieved 1GB/sec?
So could give us the correct numbers in the benchmark, if he lies – as you claim? I would like to know what the true numbers are for a CELL cpu. I bet several others here would like to know the true performance. For a large IRL data set, of course. No one is interested in tiny data sets that never occur in real life. And please, give us links also.
I wouldn’t even bother trying to extract a reason out of the guys mouth; his attacking the source of the information is nothing less than a variation of ad-hom attack. He can’t address the facts displayed to him so he attacks the source of the information – the same sort of thing christian fundamentalists do with the ‘origin of the species’ who attack Darwin instead of attacking the theory itself.
[EDIT]Hummm, I put a big post about my problems in my corporate, but I am not sure it’s not risky to speak about that on a public forum. So removing it.
Ho yeah, and I am atheist.
Edited 2009-11-22 10:19 UTC
Sorry to hear about your problems in your corporation. But could you please post the true Niagara and CELL numbers? I would like to know the true benchmark results on string pattern matching.
According to the SUN blogger, he took the CELL results from IBM’s report. IBM heavily optimized the CELL code and used assembler, and all sorts of intricate optimizations. And IBM reported the CELL numbers, which the SUN blogger reports.
Whereas the SUN Niagara just did a compile of the algorithm implemented directly from pseudo code, with no optimizations at all. I bet someone here with access to the Niagara could do a straight forward implementation of the pseudo code and verify the numbers that are reported in the Niagara benchmark.
But what do I know, maybe SUN blogger lies, as you claim. So again, could you please post the true numbers? Aparently, the CELL numbers (reported by IBM) should be… what? 5? 10? times as high? Or is it 100 times higher?
I didn’t say that the blog entry lies, I said that you should not trust a guy working in a corporate making benchmark for its own company product.
I am pretty sure that if someone here put a benchmark from an IBM guy saying that it rules the Sun servers, everybody should say that it’s not true.
If Paul Thurrott’s make a benchmark telling that Windows 7 outperform all unixes, you will trust him ?
But on OSAlert, just like for Linux, you can’t say anything bad on Sun or something good on IBM / Apple.
Edited 2009-11-22 12:46 UTC
It’s a benchmark that is doing something that Niagara was designed for, that GPGPUs are very good at, and that happens to be a general weakness of the cell. There’s no good reason not to trust it. Honestly, it shows the Cell in a very good light, given how close it got. A midrange CUDA-enabled Geforce from a couple years ago would bring a smack down on both of them, I would imagine
.
“I didn’t say that the blog entry lies, I said that you should not trust a guy…”
I fail to see the difference?
“I am pretty sure that if someone here put a benchmark from an IBM guy saying that it rules the Sun servers, everybody should say that it’s not true.”
No actually, I would not mistrust a benchmark. On other sites, I have discussed with IBM guys and they have refered to benchmarks which I have trusted. No problems at all. The reason that I trust those benches, is because it is far to easy to prove/disprove those benches. Just fire up a machine and try it yourself. If someone lied about those numbers, he would loose credibility.
It is far to dumb to lie about easily proved benches. It is like I say “on p34 i the ISO standard for C++, it says that ‘C# sucks'” – well that would be dumb to lie, because it would be to easy to prove/disprove. Then I would be considered as a liar.
So, go ahead, post IBM benches about the CELL if you like. I will trust them. I dont think IBM lies about benches. Why would they? I dont think SUN lies. Why would IBM? Benches are too easy to prove/disprove. No one should dare to lie about them? Or am I wrong?
It’s not about truth or lies, it’s about marketing.
Sun picked a test which really wasn’t suited to Cell. What they didn’t do it pick a test that wasn’t suited to Niagara or was equally suited to both. If Sun had picked any benchmark with floating point Niagara would have been completely destroyed – not just by Cell, but by any processor on the market.
One thing you can do with Cell is emulate a Niagara like architecture (you use software threads and a software cache), I don’t know how much work was done in this area but it’d be interesting to see how the benchmark would turn out doing this. OTOH you can’t emulate Cell on Niagara (at least not without a great deal of difficulty).
Anyway, Niagara and Cell are both general purpose processors optimised for specific tasks. Those tasks are quite different so there really isn’t much point comparing them.
So does that mean that the SUN benches are not for real? Sure SUN picked benches where Niagara excels. SUN would be dumb not to. Do you mean IBM picks benches where CELL performs lousy, eh?
My point is, the Cell might be fast for some work loads. Niagara is also fast for some work loads. And that does not mean that benches lie. Capisce? They have different profiles, but I would not dismiss benches showing CELLs performance as lies – as some of you do here. Instead, I accept that CELL is fast on some things, and Power6 is fast on some things, and Niagara is fast on some things.
SUN has publicly stated that Niagara has it flaws, and it is good at multi threaded work. IBM has not stated Power6’s flaws. SUN specifies strengths and weaknesses of the Niagara. IBM only specifies strengths. So whose marketing do you prefer, huh?
IBM states that Power6 has 250GB/sec bandwidth, upon closer scrutiny it shows that IBM adds all bandwidth in the chip – which is wrong, because if there is a bottleneck at 1GB/sec then the chip will never be faster than that. So whose marketing do you prefer, huh?
Or, IBM states that one mainframe can consolidate 1500 x86 servers. Upon closer scrutiny, it shows that all x86 servers must idle, and Mainframe is 100% loaded! Then I can claim that my laptop can consolidate 10 Power servers, if they all idle.
And besides, according to a famous Linux expert
http://www.mail-archive.com/[email protected]/msg18587.html
we have 1 Mainframe MIPS equals 4MHz x86. So a 10.000 Mips Mainframe CPU is quite slow. One Intel Quad core at 3GHz has
4 cores x 3GHz = 12GHz = 12.000MHz
It corresponds to 3.000MIPS. So whose marketing do you prefer, huh?
Exactly.
Statistics themselves don’t lie. People just carefully choose which statistics give the most favourable overview.
Or “Lies, damned lies and statistics” for short.
Haha, you just couldn’t resist throwing rocks at Christians, could you? As if no one has ever done this sort of thing the other way around, attacking Christians themselves instead of their actual arguments. Or better yet, as if this wasn’t done by someone on every side of every argument in the history of humanity. Please check your irrational hate of Christians at the door, it has no relevance to this discussion. God Bless!
Actually, the CELL cpu was quite slow on most things.
Yeah, one specific case will prove your “theory” =)
The cell processor is not optimized for tasks that require moving a lot of data. It is designed for vector math.
That means it’s great at things like modeling physics, cryptanalysis, and the like. Scientific computing.
I’d be interested to see benchmark comparisons for something like that.
Sony failed at proving to users that the PS3 was a good Home Media Platform. The PS3 is a good (if not the best) BluRay player but the Cell processor isn’t needed for that. When it comes to home media the PS3 can’t do anything the Xbox360 can’t do. None of my rants above is the fault of the processor though.
You know, the machines run games too. If you bought one of those and only watch movies with it, I’m not really sure why you bought it.
Sadly, most bluray players have horrible upscaling for DVDs since they don’t have the processing power of the cell.
When the PS3 was $400~, $800+ bluray players had worse upscaling capabilities.
As far as i know, 360’s micro is based of Cell technology.
It’s several SMT in-order full PPC cores. Poor theoretical performance, but able to really reach it with code doing all kinds of workloads. The main processor unit may or may not be similar, but that’s the weak point of the Cell, while a strong point of the Xenon.
Unless, of course, you mean some other thing containing 360 in its name, as that’s not the most unique number to name things with, I guess (Xbox 2pi, anyone?).
Edited 2009-11-24 08:02 UTC
The Cell *did* revolutionize many things. The Cell itself is dead, but the technologies and design principles of the Cell processor are pretty much the future of all modern CPU design. It’s the same approach that both Intel and AMD are taking with their future CPUs, e.g. Larrabee and Fusion. One or more general purpose cores fused with a number of stream processing units. Future POWER CPUs will also almost certainly take a similar approach. It just won’t be the Cell.
This announcement is just the death of a specific architecture and brand name. Intel has killed off all of its outdated architectures and brand names, too, and replaced them with newer designs and brands. Pentium is no more, the architecture of the P3 and P4 and Core 1/2 are no more.
The Cell was a first cut at a new design. That design has been improved and will the newer big brother of the Cell will replace it. The fact that that was going to happen is pure common sense.
I think that what many of us are chuckling about is how OVERHYPED the Cell was, compared to its current reality.
Actually, the Cell had quite a bit of potential. It’s just that AMD/ATi and nVidia moving to GPGPUs stomps all over it for cheap and scalable. A single powerful CPU (as in accumulator-model processor, like a typical CPU) and some GPGPU components, or many weak CPU type processors in tandem with many simpler GPGPU processors (the way AMd and nVidia are going, though very differently), can do everything the Cell can, and then some. Also, those components need far less trickery on the compiler side of things (assembly guys seem to love the SPUs).
The Cell was certainly something that should never have been used for a game console, and Sony pulled their typical shenanigans there, but I think it is hardly fair to say that they should have figured it would fail from the start (at least on the IBM side). Derivatives will likely continue to pop up, supporting natively parallel languages, like OpenCL, and life will go on.
It was a worthy effort, and the actual things it did well (many computations on small- and medium-sized data sets with minimal memory bandwidth relative to a GPU’s, FI), along with the research that made it, and research that it brought about, are helping to make for better heterogeneous systems in the years to come. Except Larrabee, which seems to mimic it in many ways that do not make sense (small number of computational units with large memory, but then complicating it with things like hardware cache-coherency, a big inefficient bus, HT-like multithreading, and x86…*sigh*, they’ll brute-force it into working, eventually).
Edited 2009-11-22 10:33 UTC
Oh, for how long had I read how the Cell processor was going to revolutionize computing.
It did. CELL was on market before Intel/AMD/NV even start to speak about CPU/GPU convergence. When PC was mainly single core.
Anybody who doesn’t see it is just blind.
It boggles the mind at just how moronic these people at IBM and Sony can be, it’s a wonder they are still in business at all.
LOL. They’re are in business exactly because they don’t fear new things and do it in a way what immovable mammonths like Intel start to move their asses.
CELL was a really nice architecture, but it had some serious shortfalls:
– Somewhat strange programming model that is difficult to understand for many people.
– The CPU part of this chip was somewhat underpowered
– Memory bandwidth between PPE and SPU
– just a single memory controller
Cell was a somewhat glorified GPU introduced by IBM. Sony and IBM weren’t moronic. I think it was a good way to implement a graphic card in budget. It’s just the way, that Cell hadn’t that impact in the market that would be necessary to spend enough in R&D to compete with ATI or nVidia.
Power7 is a somewhat more reasonable approach to build a general purpose CPU, but at this processor budget at the size mentioned in the Hot Chip document this chip will be in no way a cheap one …
You forgot what CELL was targeted to game console, which must be provide good performance for low price.
– Somewhat strange programming model that is difficult to understand for many people.
Typical console game programmers (not the ones who came from PC in the hunt for quick buck) are like to work at bare metal.
– just a single memory controller
25.6GB/s of _real_ B/W is fairly enough
.
Edited 2009-11-24 08:26 UTC
All people fairly new to OSAlert have read the crazy stuff he wrote more than half a decade ago.
Most entertaining stuff I read here and I wish there were still a few wacky writers round..
Sane reporting is too easy and boring. Bring back the crazy, please.
Why would I do that? Because someone didn’t read the article and posted a completely misleading headline?
IBM have said a chip they never announced (PowerXCell32), will never go into production. FYI This was known about *months* ago.
What is interesting is that they have said they’ll use the technology elsewhere, they have talked about this before but never really confirmed anything. Welding SPEs onto a BlueGene or POWER makes a lot more sense, I thought it was quite an obvious angle really.
So actually, the story is pretty much the opposite of what the headline says.
Er, Jan 2005 is less that half a decade ago.
So wacky in fact that it was used as required reading material by quite a number of University Computer Science courses…
I don’t think Cell is going to go away anytime soon even if it is in some other form, The biggest advantage of the architecture – scalability – probably wont be seen for some years yet.
BTW The vast majority of Cells are made by Toshiba and are used by Sony, the article says nothing about this.
Why don’t you write a follow-up next Jan.. It would be interesting to revisit your predictions for the future.
?
not nearly as powerful as GPGPU, not nearly as flexible as a real CPU, it was a somewhat an underperofming in-the middle soluiton.
You must remember that these were connected to the POWER series of processors, right?
Thom, I am disappointed that you have decided to go for the most sensationalistic headline and summary possible in exchange for journalistic accuracy.
Let’s take a look at the actual contents of the article, for people who can’t read German. From the first paragraph:
“nun hat IBMs Vice President of Deep Computing, David Turek, im Interview mit heise online best~A¤tigt, dass es den geplanten Nachfolger des aktuellen PowerXCell-8i-Prozessors mit zwei PowerPC-Prozessoren und 32 SPEs nicht geben wird. Damit sei das Cell-Design aber nicht tot, hei~A"Yt es, Teile davon sollen in anderer Form wieder auftauchen.”
Translation: “IBM’s VP of Deep Computing, David Turek, has now confirmed in an interview with heise online that the planned successor to the current PowerXCell-8i processor, slated to have two PowerPC processors and 32 SPEs, will not be released. This will not, however, spell the end of the Cell design, as parts of it will still crop up in different forms.”
The article then goes on to talk about IBM’s plans to release hybrid Power and BlueGene chips, with all chips simultaneously targetable via the usage of OpenCL.
Not once does the article say that the PowerXCell-8i “will be the last Cell processor”. In case you’d forgotten, Sony and Toshiba co-developed Cell, and in fact continue to manufacture them. In fact, Toshiba is bringing out a new Cell-powered set-top box/DVR as we speak:
http://www.engadget.com/2009/10/05/toshiba-cell-regza-55×1-and-regz…
The PowerXCell 8i is just IBM’s variant that it used in its supercomputers, nothing more, nothing less. This is pretty much news just for the HPC sector, and even then, not that spectacular.
EDIT: By the way, the article that heise online links to actually *is* in English, and contains considerably more detail:
http://www.hpcwire.com/features/Will-Roadrunner-Be-the-Cells-Last-H…
Edited 2009-11-21 13:40 UTC
From what it sounds like, with the targeting of OpenCL, one is going to see the Cell appear as a co-processor. From what I understand, Toshiba I think sold a laptop with a Cell processor in it. I see that NVIDIA are selling machines that have all the PCIe slots loaded with NVIDIA video cards – are we going to see the same sort of thing in the case of Cell where one can purchase PCIe cell processing cards and used to speed up computing in the area of video and audio compression etc?
to go. Like AMD, a CPU-only OpenCL solution can bring revolution into computing (with a 2D acclerated framebuffer only). The main problem I see is that CELL is suitable for this but does not make inroads to the desktop. Using it as a co-processor sounds an overkill because you do not have a unified solution and rely on PCIe bandwidth to move the tasks. I hope to see real desktop motherboards with CELL and a CPU only OpenCL implementation.
Edited 2009-11-22 09:14 UTC
and about the PS3 I believe the choice of NVIDIA was a mistake. 2D framebuffers by XGI/SM were way better and better coupled with CELL for 3D rendering/ Decoding. If they needed NVIDIA they could use AMD/Intel why spend their money on CELL.
…er, double post. Delete if possible.
Edited 2009-11-22 10:34 UTC
They really wanted to have only Cells on the PS3, but then the Cell turned out to suck as a GPU. They needed to get something done. nVidia had nothing in any consoles, so they were certainly eager. They needed triangles pushed out, right then, not a really good elegant technological solution.
Or, to summarize: they were being Sony.
Edited 2009-11-22 10:36 UTC
Sure I read somewhere that IBM hired the US chip guys Sun dumped and that mysteriously Power7 was very ‘Niagara’ shaped…
It would make sense to roll the ideas from the Cell into that. Still funny to watch IBM eat crow and have to effectively copy their almost dead competitors ideas as well as their OS
Kind of sort of…
All server processors are going in that direction.
They all influence each other, who do you think did servers with multi-core and multi-threads first? Hint: it wasn’t Sun.
I always found it amusing that IBM built a No1 supercomputer based on Cell. This was ironic because Cray and IBM had been competitors for many years and IBM had always pushed the more complex processors, Cell is much closer to a Cray design (Cray 2 actually).
“They all influence each other, who do you think did servers with multi-core and multi-threads first? Hint: it wasn’t Sun. ”
The fun thing is that IBM has long bashed SUNs Niagara architecture. IBM has several times stated that few cores with higher clock frequency is much better than many slow cores. As proof, IBM shows that one Power6 core is faster than one Niagara core, and therefore the Power6 CPU must be faster. This is how IBM’s proof goes that Power6 is the fastest CPU. Which is wrong by the way. Because, if a core is slow, it doesnt tell anything about the whole CPU. If the fastest cpu in the world uses 10.000 slow cores, and you can show that Power6 core is faster, does that translate to the Power6 cpu being faster? Of course not. That is just marketing and FUD. Power6 is not fastest, which you can see from benchmarks. SpecINT, SIEBEL v8, etc. One Sun T5440 with four 1.4GHz Niagara is twice as fast as three IBM Power570 servers with 12 Power6 at 4.7GHz – when you compare white papers from SUN and IBM. And one T5440 costs 76.000USD and one P570 cost 413.000USD. So how can the Power6 be fastest, huh? Specially when you consider that Niagara has many world records right now.
Now it is funny to see IBM releasing many slower cores, than few cores. SUN was first to popularize and heavily promote many lower clocked cores. There might be some researcher in 1950 that spoke about multi cores, so what? He didnt take them to the mass market. SUN did.
Wrong. Many faster cores.
IBM says that there will be about 20 percent across the board improvement on a per-core basis, and 4X to 5X performance improvement when compared chip-to-chip.
http://www.techztalk.com/techwebsite/08-26-09-ibm-unveils-power7-pr…
This is interesting. I have read that something about the Power7 cores:
http://www.c0t0d0s0.org/archives/5786-About-Power7.html
“Let^A's assume the Power6 is 2-core, let^A's further assume that Power7 is a 8-core design. Obviously a single core of the Power7 has just half the performance of the Power6 core.”
So it seems that IBM has jumped on the band wagon with lots of slower cores, than few fast cores? And how much IBM bashed Niagara for doing this!
IBM uses more and more cores, and they dont get faster. Instead they get more and more. Just like SUN.
Edited 2009-11-23 15:18 UTC
IBM uses more and more cores, and they dont get faster. Instead they get more and more. Just like SUN.
Again. Where is the facts? IBM says about 20% improvements per core (even with a slower cpu freq)
In the link above all I can see is a flawed logic of biased sun blogger made even before P7 details was unveiled.
http://ideasint.blogs.com/ideasinsights/2009/08/ibm-reveals-a-few-p…
Sorry … but that article was written just on the foundation that was made available by IBM at that time … and so far nobody was able to explain me where my logic was flawed except “But IBM said otherwise”.
BTW: Even when the core is 20 percent faster, you have to take into consideration that it’s SMT4 instead of SMT2 … that makes the equation even more difficult.
In your link:
“With up to eight cores per chip, compared to the dual-core POWER6, POWER7 could achieve up to four times the performance per chip, depending on the workload. To remain within the same power/thermal envelope, POWER7 clock rates are not expected to be quite as high as POWER6”
If you took dual core Power6 and made 8 cores, you get the Power7 which gives _up to_ 4 times the performance. To me it seems that Power7 cores are not much faster than Power6 cores. IBM just collects more and more of lower clocked cores. Just what IBM bashed earlier. But when IBM later does exactly the same thing, it is the best thing since sliced bread. This is just weird.