Is IBM "cell" chip a replacement of Intel CPU or ATI/NVidia GPU?

[Ar Q]

The market seems to think the IBM "cell" chip a replacement to Intel/AMD
CPU. But from the description, it seems to me it is another GPU like ATI or
NVidia has right now? What do you guys think?

 

[TF]

This article sums it well enough.
http://biz.yahoo.com/rb/041129/tech_cell_2.html

The "I" word company (not IBM) has been looking for many years to expand
into different areas.
They have been trying to push chips/systems that run WinCE, etc. to handle
some of this, as well as putting video on-chip and developing network
conectivity.
However, they haven't done well.
Also, the standard processor market has slowed way down.
Not as many employed people to keep upgrading and those that are working,
are spending less.
Besides, the USD is worth much less now and petrol is not cheap.

This chip looks like much more than a GPU, but don't worry, IBM is hardly
alone in this area and competition will be tough, but being first with
something big and getting it into lots of products first has a huge
advantage.

TF

 

[keith]

The market seems to think the IBM "cell" chip a replacement to Intel/AMD
CPU.

I don't think anyone with a brain (which leaves out any martetroids or
news types) thinks its going to replace X86.

But from the description, it seems to me it is another GPU like ATI or
NVidia has right now? What do you guys think?

It's far more than a GPU. Make that an array og GPUs tightly coupled to an
array of 64b CPUs and you have a closer approximation. We'll see if
software can keep up with the hardware.

 

[Ar Q]

It's far more than a GPU. Make that an array og GPUs tightly coupled to an
array of 64b CPUs and you have a closer approximation. We'll see if
software can keep up with the hardware.

Did IBM hire Gray's Connection Machine team to design "Cell"?

 

[keith]

Did IBM hire Gray's Connection Machine team to design "Cell"?

Cray? I don't think so, but anything's possible.

 

[)-()-(]

The market seems to think the IBM "cell" chip a replacement to Intel/AMD
CPU. But from the description, it seems to me it is another GPU like ATI or
NVidia has right now? What do you guys think?

Sounds like a DSP type chip. That's TI's area of expertise. I don't think it's
a GPU.
I think it will interface broadband digital networks to video like HDTV.
Kind of like what the Sony VAIO PC's do now using a CPU and GPU,
but they will use a "cell', or DSP type chip to process the video data stream.

 

[ykhan]

The market seems to think the IBM "cell" chip a replacement to Intel/AMD
CPU. But from the description, it seems to me it is another GPU like ATI or
NVidia has right now? What do you guys think?

The Cell chip is basically a variant of the PowerPC chips that IBM
puts in everything from laser printers to Apple Macintoshes to IBM's
own mainframes. It's meant to be a general purpose CPU, but it does
have some elements of what a GPU has. But I think it's much closer to
the CPU role than the GPU role.

As for whether it's a replacement for the Intel/AMD chips, that's a
definite negatory. The CPU market is one thing, but the x86 CPU market
of which Intel and AMD are the two biggest members are quite another
market. No matter how much performance a competing architecture brings
to bear, it's still never good enough to be considered a replacement
for x86. The only thing that's good enough to replace an x86 is
another x86. Just look at the relative success of Opteron vs. Itanium.

Yousuf Khan

 

[Ar Q]

Did IBM hire Gray's Connection Machine team to design "Cell"?

I mean Cray.

 

[Rob Stow]

Ar Q wrote:




Sounds like a DSP type chip. That's TI's area of expertise. I don't think it's
a GPU.

The descriptions I've read so far make it sound like basically a
Power series GPU core surrounded by a bunch of much smaller DSP
cores. It is apparently going to be a true multi-core chip and
not a multi-chip package.

 

[keith]

The descriptions I've read so far make it sound like basically a
Power series GPU core surrounded by a bunch of much smaller DSP
cores. It is apparently going to be a true multi-core chip and
not a multi-chip package.

Yes. What nut would do an MCM in that market?!

 

[Mark Hahn]

As for whether it's a replacement for the Intel/AMD chips, that's a

definite negatory. The CPU market is one thing, but the x86 CPU market
of which Intel and AMD are the two biggest members are quite another
market. No matter how much performance a competing architecture brings
to bear, it's still never good enough to be considered a replacement
for x86. The only thing that's good enough to replace an x86 is
another x86.

that's a bit of an overstatement - transmeta has found a corner of
the x86 market where a non-x86 can do some damage. sure, they haven't
taken over the world, but they're certainly viable.

Just look at the relative success of Opteron vs. Itanium.

it's very intresting to compare. for instance, the it2 has been
pushed very hard ahead of demand, and hasn't exactly shone.
but the opteron has been anti-marketed by AMD (ie, yes, it does
run windows, but most windows-on-opteron is in brain-damaged
Intel-compatible mode). so the Opteron has a big bump to performance
just waiting for someone in Redmond to take advantage of it.
the it2 desperately needs a perf bump, since at least one large
company has staked its enterprise line on it. and the it2 perf bump
doesn't seem forthcoming.

look at the routes these chips took. opteron tweaked the existing,
fairly boring athlon core, put a nicer cache on it, great memory interface,
and elegant system infrastructure. considering how little AMD has improved
their core over the years, their success is amazing. quite a contrast
to the it2, whose raison-d'etre was to do a new core ISA from scratch.
Intel/HP have always claimed that EPIC's design was motivated by wanting
to scale performance into the stratosphere. alas, for current it2's at
least, where performance is good happens to be places where the entire
application+dataset fit into cache. would it2-9m be faster than an
opteron with 9M cache? I doubt it. what does that say about Intel/HP's
innovation-in-ISA?

and multicore is not going to save it2. it'll give a nice tweak to
performance under some loads, no doubt. but since it2's competitors
will probably go multicore before it, Intel gets no relative advantage.
and after all, it2-multicore is essentially just making up for the fact
that the it2 ISA designers are so hostile to out-of-order. one of those
mysteries - if OOO is such a bad thing, why have other chip design teams
been able to push it so far? no doubt the OOO designs have been harder
because of the OOO, but has the it2 taken advantage of this to push other
aspects of performance, or to have a shorter time-to-market? unless you
count obscene onchip caches as an ISA innovation, I just don't see where
the it2 is winning.

here's a match I'd love to see: AMD buys Transmeta. why? mainly because
Transmeta has higher-order, firmware-based insight into performance.
for instance, Transmeta can implement using a hard+soft combination,
more intelligence in predicting branches, or in predicting computed ld/st
addresses, or even in predicting values. suppose you can accurately predict
when a store writes to an address which won't be used any time soon. why
not write it through to dram? no sense in wasting a cacheline on that.
predicting loads, of course, is even better, since they tend to stall,
but stores can be posted. but can you afford to treat every load the same
way you do branches (with a multi-state FSM to predict the outcome)?
perhaps not in hardware, but maybe in firmware (maybe even encoded in your
internal instruction stream, since AMD already does cache partially decoded
instructions, though doesnt' go as far as Intel's trace cache...)

regards, mark hahn.

 

[Eddie Grove]

here's a match I'd love to see: AMD buys Transmeta. why? mainly because

I don't want to see AMD buy anybody. But coincidentally, there's
a quasi-AMD-Transmeta agreement today.

Fujitsu just announced that it licensed some tech from Transmeta.
The significance is that AMD and Fujitsu are partners in flash
memory in a unit called Spansion. The press release wasn't too
specific so I can't really guess if Spansion was relevant to the
agreement.


Eddie

 

[Radeon350]

The Cell chip is basically a variant of the PowerPC chips that IBM
puts in everything from laser printers to Apple Macintoshes to IBM's
own mainframes. It's meant to be a general purpose CPU, but it does
have some elements of what a GPU has. But I think it's much closer to
the CPU role than the GPU role.

As for whether it's a replacement for the Intel/AMD chips, that's a
definite negatory. The CPU market is one thing, but the x86 CPU market
of which Intel and AMD are the two biggest members are quite another
market. No matter how much performance a competing architecture brings
to bear, it's still never good enough to be considered a replacement
for x86. The only thing that's good enough to replace an x86 is
another x86. Just look at the relative success of Opteron vs. Itanium.

Yousuf Khan

Actually, Cell is much, much more than just a PowerPC variant. Yes,
Cell processors will *contain* a PowerPC or POWER variant, but most of
Cell is different from PowerPC. the APUs, or now apparantly they're
called SPUs, will do the lionshare of the processing. the PowerPC or
POWER cores within Cell processors will run the OS, and hand out the
workloads for the APUs to do, like all of the floating point and
integer processing. Cell is also very scalable and modular. moreso
than PowerPC. actually you could replace the PowerPC or POWER cores
with ARM, MIPS, Intel, AMD, etc CPUs and Cell would still be Cell.

 

[ykhan]

that's a bit of an overstatement - transmeta has found a corner of
the x86 market where a non-x86 can do some damage. sure, they haven't
taken over the world, but they're certainly viable.

Yeah, but Transmeta is just another x86. Nobody uses it for anything
other than x86, and Transmeta won't even tell you how to "go native"
-- it wants you to use x86.

look at the routes these chips took. opteron tweaked the existing,
fairly boring athlon core, put a nicer cache on it, great memory interface,
and elegant system infrastructure. considering how little AMD has improved
their core over the years, their success is amazing. quite a contrast
to the it2, whose raison-d'etre was to do a new core ISA from scratch.

Perhaps the lesson here is that the ISA was never the problem. Yeah,
it's difficult to believe after years of us being cult-indoctrinated
to believe x86 was evil; that it was about to leave this planet for
good and never show its ugly face ever again; that it's still here,
and doing better than ever.

here's a match I'd love to see: AMD buys Transmeta. why? mainly because
Transmeta has higher-order, firmware-based insight into performance.
for instance, Transmeta can implement using a hard+soft combination,
more intelligence in predicting branches, or in predicting computed ld/st
addresses, or even in predicting values.

How do you know Transmeta is doing all of this right now? If it is,
then it's not really showing through on Transmeta's own processors.

Yousuf Khan

 

[ykhan]

I don't want to see AMD buy anybody. But coincidentally, there's
a quasi-AMD-Transmeta agreement today.

Fujitsu just announced that it licensed some tech from Transmeta.
The significance is that AMD and Fujitsu are partners in flash
memory in a unit called Spansion. The press release wasn't too
specific so I can't really guess if Spansion was relevant to the
agreement.

I can't see how anything Transmeta has would be relevant to anything
Spansion does, which makes flash memory. However, Fujitsu does have a
business making and selling UltraSparc-compatible processors for its
own server division, and there was a recently announced agreement
between Fujitsu and Sun for Sun to use Fujitsu's Sparc64 processors
instead of its own UltraSparcs. This might have more to do with that
CPU/server business than flash.

Spansion is more relevant to AMD than to Fujitsu; Fujitsu is
definitely the silent partner it this venture.

Yousuf Khan

 

[keith]

that's a bit of an overstatement - transmeta has found a corner of
the x86 market where a non-x86 can do some damage. sure, they haven't
taken over the world, but they're certainly viable.

I think this is over-stated. First, Transmetta *is* x86. They haven't
released their internal ISA, so it's x86 on the outside. Second, I don't
think Transmetta is viable. I don't see it lasting long without a major
float. Indeed, I'm surprised it even got past the first announcement
hype. The fact that it still exists is nothing but amazing.

it's very intresting to compare. for instance, the it2 has been pushed
very hard ahead of demand, and hasn't exactly shone. but the opteron has
been anti-marketed by AMD (ie, yes, it does run windows, but most
windows-on-opteron is in brain-damaged Intel-compatible mode).

Anti-marked by AMD? DO you think AMD owns M$?

so the
Opteron has a big bump to performance just waiting for someone in
Redmond to take advantage of it. the it2 desperately needs a perf bump,
since at least one large company has staked its enterprise line on it.

Don't believe this hype either.

and the it2 perf bump doesn't seem forthcoming.

Agreed. Itanic lost its 15 minutes. Could-shoulda-woulda.

look at the routes these chips took. opteron tweaked the existing,
fairly boring athlon core, put a nicer cache on it, great memory
interface, and elegant system infrastructure.

"Tweaked", "fairly boring Athlon"? We live in alternat universes,
obviously (mine has capital letters, BTW, ;-). The Atlon core was
anything but boring. It was a rather agressive super-scalar OoO
architecture, one that set Intel's P3 on its butt. I hardly consider the
K8 a "tweaked" K7. Sure, there are elements of the K7, but hell there are
elements of the Eniac in it too.

considering how little
AMD has improved their core over the years, their success is amazing.

You're on drugs! I suppose you don't think IBM's Z series is more than a
"little" improved over the 360 that came our 40 years ago. Sheesh! The
parallels are there for the blind to see.

quite a contrast to the it2, whose raison-d'etre was to do a new core
ISA from scratch. Intel/HP have always claimed that EPIC's design was
motivated by wanting to scale performance into the stratosphere. alas,
for current it2's at least, where performance is good happens to be
places where the entire application+dataset fit into cache. would
it2-9m be faster than an opteron with 9M cache? I doubt it. what does
that say about Intel/HP's innovation-in-ISA?

Again, you miss the forrest for the trees. The real question is "what
does that say for evolution over revolution?". Opteron evolved, the
Itanic sunk.

and multicore is not going to save it2. it'll give a nice tweak to
performance under some loads, no doubt. but since it2's competitors
will probably go multicore before it,

Umm, PowerPC has been multicore for several years. There is nothing
new here. Pay attention!

Intel gets no relative advantage.
and after all, it2-multicore is essentially just making up for the fact
that the it2 ISA designers are so hostile to out-of-order. one of those
mysteries - if OOO is such a bad thing, why have other chip design teams
been able to push it so far? no doubt the OOO designs have been harder
because of the OOO, but has the it2 taken advantage of this to push
other aspects of performance, or to have a shorter time-to-market?
unless you count obscene onchip caches as an ISA innovation, I just
don't see where the it2 is winning.

Are we rambling here? ...or is there a point?

here's a match I'd love to see: AMD buys Transmeta.
Gack!

why? mainly
because Transmeta has higher-order, firmware-based insight into
performance.

....and it's given them exactly what? What gain would that knowledge
give AMD? Please, Transmetta is all about profiling, albeit on the fly.

for instance, Transmeta can implement using a hard+soft
combination, more intelligence in predicting branches, or in predicting
computed ld/st addresses, or even in predicting values.

Sounds great. Tastes not so.

suppose you can
accurately predict when a store writes to an address which won't be used
any time soon.

If you can *really* predict this well, why bother doing the instruction in
the interim? This is *exactly* the falacy of Itanic. The reson we
execute instructions is because we *don't* know what the result is going
to be ahead of time.

why not write it through to dram? no sense in wasting a
cacheline on that.

Good plan. All we need is the famous IEHGOD supervisor and we can have
IBM;s OS/VU (Virtual Universe), since we can now see into the future!

predicting loads, of course, is even better, since
they tend to stall, but stores can be posted. but can you afford to
treat every load the same way you do branches (with a multi-state FSM to
predict the outcome)? perhaps not in hardware, but maybe in firmware
(maybe even encoded in your internal instruction stream, since AMD
already does cache partially decoded instructions, though doesnt' go as
far as Intel's trace cache...)

Hell, if we know the answer, don't even turn the damned machine on!

Let me guess, you've never worked in processor or OS development.

 

[Tony Hill]

that's a bit of an overstatement - transmeta has found a corner of
the x86 market where a non-x86 can do some damage. sure, they haven't
taken over the world, but they're certainly viable.

Huh? Transmeta *IS* x86, they just took a really odd-ball way of
going about it. Ok, in theory I suppose someone could get another ISA
working on their chips, but no one has done so yet and the chips were
designed expressly with x86 in mind.

Also, calling them "viable" is probably giving them way too much
credit. They're rather deep in the red and look unlikely to ever pull
out of it. Their latest and greatest "Efficeon" chips are already
falling WELL behind the super-low-power solutions from Intel, leaving
them to compete in the same market as VIA, who's just creaming them in
that (already small) niche for very low powered and low-cost x86
chips.

it's very intresting to compare. for instance, the it2 has been
pushed very hard ahead of demand, and hasn't exactly shone.
but the opteron has been anti-marketed by AMD (ie, yes, it does
run windows, but most windows-on-opteron is in brain-damaged
Intel-compatible mode).

I don't think that it's AMD doing the anti-marketing here...

Also, rumor has it that as much as 60% of all Opterons are running
under Linux, as least as per IDC:

http://www.theinquirer.net/?article=19859

so the Opteron has a big bump to performance
just waiting for someone in Redmond to take advantage of it.
the it2 desperately needs a perf bump, since at least one large
company has staked its enterprise line on it. and the it2 perf bump
doesn't seem forthcoming.

Don't hold your breath for any groundbreaking increase in performance
with 64-bit software, the real key to 64-bit is that it allows you to
access much larger chunks of memory. Sure, there will be some
benefits from the 8 extra registers, normally enough to negate the
traditional performance penalty associated with 64-bit software (ie
pointers that are twice as big). In most testing AMD has only seen
about a 5-10% performance boost on average. Very nice considering
it's "free", but it's not going to dramatically change the dynamics of
things.

The Itanium's performance bump has already arrived in the form of VERY
good compilers from Intel. They've probably raised the bar in
compiler design by a HUGE margin to get the Itanium performing up to
speed.

And in some situations, the Itanium DOES perform up to speed.
Particularly in the scientific computing market it seems to do very
well. The problem for Intel is that this is another of those rather
small niches. Ok, this one deals with BIG-$$$, but extremely low unit
sales. Intel just wasn't designing the Itanium for these sorts of
economics of scale, and if they did the chip would probably cost a lot
more and be much less attractive in this market.

and multicore is not going to save it2.

Actually I think multicore is exactly what Itanium needs, though not
just one core. Itanium looks like VERY good candidate for a Sun
"Niagara" or IBM BlueGene/L style design of a WHOLE LOT of cores. The
whole idea of Itanium is that it has a fairly simple core, so it
should be relatively easy to stick a boatload (err.. ok, maybe we
should aim for 8 to start with) of those cores on a single chip.

That being said though, I don't know that multiple cores will save
Itanium. The real problem that Intel is facing is that hardware is
cheap, software and support is EXPENSIVE. The Itanium really doesn't
offer anything on the support side and the software line-up for it is
very weak relative to it's main competitors (IBM Power and Sun Sparc
and especially Intel Xeon/AMD Opteron encroaching from the low-end).

 

[chrisv]

Perhaps the lesson here is that the ISA was never the problem. Yeah,
it's difficult to believe after years of us being cult-indoctrinated
to believe x86 was evil; that it was about to leave this planet for
good and never show its ugly face ever again; that it's still here,
and doing better than ever.

Well, it's not exactly elegant, either. Our new processors would be
less complex, and thus possibly cheaper and faster, if they didn't
need to support such a "kludgy" ISA...

 

[chrisv]

I think this is over-stated. First, Transmetta *is* x86. They haven't
released their internal ISA, so it's x86 on the outside. Second, I don't
think Transmetta is viable. I don't see it lasting long without a major
float.

I agree, considering Intel will sell you an ultra-low-voltage
Pentium-M w/2MB L2 that has a Pmax of only 5W while running at 1.1GHz
- which would blow any Transmetta chip out of the water.

ftp://download.intel.com/design/mobile/datashts/30218904.pdf

(page 70)

 

[Felger Carbon]

The whole idea of Itanium is that it has a fairly simple core

How simple is it, Tony? ;-)

Seriously, how does the die area compare with that of x86 chips such
as the OOO Opteron and the non-OOO Via chips (exclusive of caches)? I
knew the Itanium was *different*, but I didn't realize it was
*simpler*.