Intel strikes back with a parallel x86 design

[Jim Brooks]

Signs and portents as JMS would say.

Stevel Jobs does a 180' and enthusiastically becomes
Intel's bedfellow on the basis of a compelling roadmap.
That roadmap has to be pretty darned interesting.

Intel claims they aren't developing Hyperthreading anymore.
But Intel now knows all the issues involved in hw threading.
Why not exploit that know-how as an advantage over AMD?
AMD has only a fraction of the resources that Intel has,
so AMD will have a hard time catching up

My speculation is that Intel will build on their HyperThread experience
to design a "parallel x86". x86 CPUs have become superscalar machines.
The next evolutionary step is a parallel machine. Dual-cores are only
an inefficient stop-gap design that wastes transistors with duplicated
or unnecessary resources (eg coherency logic between the core's caches).

My ideas for a parallel x86:

- thread quantums

The idea is to move coarse-granularity timer-driven time-slicing
into the hw so that time-slices can be instruction-granular.

- thread prioritization

The OS assigns static priorities to threads.
The hw computes dynamic priorites according to static priority
and instruction issue for a thread per quantum.

- sub-threads

Support for parallel programming.
A reduced 80386 Task-State Segment (TSS) will be defined
(avoid saving unnecessary registers such as ES/FS/GS)
A variant of JUMP [TSS] with a new Thread bit defined in the TSS
will spawn a sub-thread (analoguous to a UNIX child process).
The sub-thread can stop by IRET [TSS].
A new WAIT [TSS] will synchronize the parent with its sub-thread.

- thread exceptions

A thread can raise exceptions to end or suspend itself.

- cache lines have thread bits in addition to LRU bits

When one cache line has to be evicted, victimize the line owned
by a lower-priorty thread.

- ALUs: 8 simple, 4 complex.

- FPUs: 4 FADD, 4 FMUL, 2 FLDST.

- deprecation of FP SIMD instruction set

SIMD was a good idea for a single-thread CPU as it let the control unit
issue a single-instruction for multiple-data without resource hazards.
But a multi-threaded control unit would function optimally with
a wide window of decomposed (SISD) instructions.

 

[Tony Hill]

Signs and portents as JMS would say.

Stevel Jobs does a 180' and enthusiastically becomes
Intel's bedfellow on the basis of a compelling roadmap.
That roadmap has to be pretty darned interesting.

Intel claims they aren't developing Hyperthreading anymore.
But Intel now knows all the issues involved in hw threading.
Why not exploit that know-how as an advantage over AMD?
AMD has only a fraction of the resources that Intel has,
so AMD will have a hard time catching up

My speculation is that Intel will build on their HyperThread experience
to design a "parallel x86". x86 CPUs have become superscalar machines.
The next evolutionary step is a parallel machine. Dual-cores are only
an inefficient stop-gap design that wastes transistors with duplicated
or unnecessary resources (eg coherency logic between the core's caches).

My ideas for a parallel x86:

And just how much do you need to change existing software to work with
your new model?

Hardware is cheap, software is expensive. Make your hardware to work
well with the software that exists and you'll do a LOT better than the
next guy who is making hardware for software they hope to be invented
in a few years time.

 

[Nick Maclaren]

And just how much do you need to change existing software to work with
your new model?

Hardware is cheap, software is expensive. Make your hardware to work
well with the software that exists and you'll do a LOT better than the
next guy who is making hardware for software they hope to be invented
in a few years time.

If you can do that, you can almost certainly square circles, travel
faster than light and produce an immortality treatment.

Hardware is engineered, software is hacked up. Even those of us
with pretty powerful imaginations have been unable to think of any
hardware designs that would work well with the software that exists
and yet not violate the laws of physics.


Regards,
Nick Maclaren.

 

[Oliver S.]

I'd rather like to see an x86-architecture with a lot of cores like
Niagara. These cores shouldn't be superscalar, but should feature mul-
tiple execution-units to enable multiple threads (I think four threads
would be best) to run in parallel physically. Branch-prediction and
prefetching could be dropped because the threads would be able to sat-
urate the execution-units anyway. The FPU sould remain pipelined of
course to enable the threads to feed the adder(s) and multiplier(s)
every cycle. So the per-cycle throughput of a single thread wouldn't
be better as for an old 486 processor.
The only problem I see with this architecture is, that a lot of pro-
grams would have to be parallelized to take advantage of this. In most
cases this isn't necessary (i.e. for desktop business-applications,
word processors etc). In some cases like video-encoders or raytracers,
this is easy to handle. But in some cases like with games this is rea-
lly challangeing.

 

[Yousuf Khan]

Signs and portents as JMS would say.

Stevel Jobs does a 180' and enthusiastically becomes
Intel's bedfellow on the basis of a compelling roadmap.
That roadmap has to be pretty darned interesting.

Intel claims they aren't developing Hyperthreading anymore.
But Intel now knows all the issues involved in hw threading.
Why not exploit that know-how as an advantage over AMD?
AMD has only a fraction of the resources that Intel has,
so AMD will have a hard time catching up

I would hardly call the knowledge they gained in hyperthreading to be
knowledge that will come in handy later. Whatever they learned about HT
was only relevent to a Pentium 4, it can't apply to Pentium M.

My speculation is that Intel will build on their HyperThread experience
to design a "parallel x86". x86 CPUs have become superscalar machines.
The next evolutionary step is a parallel machine. Dual-cores are only
an inefficient stop-gap design that wastes transistors with duplicated
or unnecessary resources (eg coherency logic between the core's caches).

My ideas for a parallel x86:

- thread quantums

<snip>

You're from the comp.arch group aren't you?

Yousuf Khan

 

[Del Cecchi]

I would hardly call the knowledge they gained in hyperthreading to be
knowledge that will come in handy later. Whatever they learned about HT
was only relevent to a Pentium 4, it can't apply to Pentium M.


<snip>

You're from the comp.arch group aren't you?

Yousuf Khan

doesn't ring any bells from comp.arch
Has only a few posts on groups.google.com search, my guess is
comp.sys.intel

 

[Nathan Bates]

I would hardly call the knowledge they gained in hyperthreading to be
knowledge that will come in handy later. Whatever they learned about HT
was only relevent to a Pentium 4, it can't apply to Pentium M.

That's like saying none of the knowledge gained from
the P6 design carried over to the Williamette design.

About the Mac, Jobs looked at the roadmaps of both Intel and AMD.
There were other factors, but Jobs was more intrigued by Intel's.
Why?

Turning to the server-blade market, when you need more racks of blades,
the consideration is packing as much processing performance per watt
(and space).
There is only so much instruction parallelism in one thread of code.
Single-threaded superscalars core often stall because of
data dependencies, resource locks, cache misses,etc.
Multi-threaded (parallel) cores are the future.

 

[Anne & Lynn Wheeler]

My speculation is that Intel will build on their HyperThread experience
to design a "parallel x86". x86 CPUs have become superscalar machines.
The next evolutionary step is a parallel machine. Dual-cores are only
an inefficient stop-gap design that wastes transistors with duplicated
or unnecessary resources (eg coherency logic between the core's caches).

intel back to the future? old posting
http://www.garlic.com/~lynn/2001n.html#83 CM-5 Thinking Machines, Supercomputers

that includes pieces of some old news articles ... inclucing an NYT
article from 6/15/92 titled Foray into Mainstream for Parallel
Computing. some number of the efforts from the period were parallel
x86s ... also can anybody say Intel Paragon?

how 'bout ncube:
http://en.wikipedia.org/wiki/NCUBE

iPSC
http://www.cs.kuleuven.ac.be/museum/multiproc/myosotis-E.html

IPSC/2 hypercube
http://portal.acm.org/citation.cfm?id=255159

red asci, originally pentium pro, then updated to pentium II ..
http://en.wikipedia.org/wiki/Intel_ASCI_Red
http://www.sandia.gov/ASCI/Red/

in the mid-90s, there there were so many people tied up in working on
such stuff that one such new mpp project ... looking around for people
to lead the effort, got around to asking my wife and me. I ran thru
all the names I could think of them to ask ... and they pointed out
that everybody I named, were all busy, tied up on something else

 

[YKhan]

doesn't ring any bells from comp.arch
Has only a few posts on groups.google.com search, my guess is
comp.sys.intel

Oh well, it just seems the comp.arch group produces an abundance of
dreamers. It's like a Beatnik Speakeasy for chip geeks. It's like the
technological equivalent of artsies. All of the "Alpha coulda ruled the
world types" seem to hang out there.

Yousuf Khan

 

[Del Cecchi]

Oh well, it just seems the comp.arch group produces an abundance of
dreamers. It's like a Beatnik Speakeasy for chip geeks. It's like the
technological equivalent of artsies. All of the "Alpha coulda ruled the
world types" seem to hang out there.

Yousuf Khan

Along with a few hard headed circuit designers..... and some highly
qualified and knowledgable folks who have btdt.

del

 

[dummy]

Signs and portents as JMS would say.

Stevel Jobs does a 180' and enthusiastically becomes
Intel's bedfellow on the basis of a compelling roadmap.
That roadmap has to be pretty darned interesting.

Intel claims they aren't developing Hyperthreading anymore.
But Intel now knows all the issues involved in hw threading.
Why not exploit that know-how as an advantage over AMD?
AMD has only a fraction of the resources that Intel has,
so AMD will have a hard time catching up

My speculation is that Intel will build on their HyperThread experience
to design a "parallel x86". x86 CPUs have become superscalar machines.
The next evolutionary step is a parallel machine. Dual-cores are only
an inefficient stop-gap design that wastes transistors with duplicated
or unnecessary resources (eg coherency logic between the core's caches).

My ideas for a parallel x86:

- thread quantums

The idea is to move coarse-granularity timer-driven time-slicing
into the hw so that time-slices can be instruction-granular.

- thread prioritization

The OS assigns static priorities to threads.
The hw computes dynamic priorites according to static priority
and instruction issue for a thread per quantum.

- sub-threads

Support for parallel programming.
A reduced 80386 Task-State Segment (TSS) will be defined
(avoid saving unnecessary registers such as ES/FS/GS)
A variant of JUMP [TSS] with a new Thread bit defined in the TSS
will spawn a sub-thread (analoguous to a UNIX child process).
The sub-thread can stop by IRET [TSS].
A new WAIT [TSS] will synchronize the parent with its sub-thread.

- thread exceptions

A thread can raise exceptions to end or suspend itself.

- cache lines have thread bits in addition to LRU bits

When one cache line has to be evicted, victimize the line owned
by a lower-priorty thread.

- ALUs: 8 simple, 4 complex.

- FPUs: 4 FADD, 4 FMUL, 2 FLDST.

- deprecation of FP SIMD instruction set

SIMD was a good idea for a single-thread CPU as it let the control unit
issue a single-instruction for multiple-data without resource hazards.
But a multi-threaded control unit would function optimally with
a wide window of decomposed (SISD) instructions.

We'll see this CPU introduced in the blueberry iMac in Q3 2006 ;-)

 

[Joe Seigh]

Oh well, it just seems the comp.arch group produces an abundance of
dreamers. It's like a Beatnik Speakeasy for chip geeks. It's like the
technological equivalent of artsies. All of the "Alpha coulda ruled the
world types" seem to hang out there.

Along with the (transputer | iAPX-432 | TMS9900) coulda ruled the world
types. I only mention TMS9900 because even though it had memory mapped
registers, to a programmer it looked real good compared to the 8088 ISA.
You can substitute your favorite failed obscure processor there.

 

[Anton Ertl]

About the Mac, Jobs looked at the roadmaps of both Intel and AMD.
There were other factors, but Jobs was more intrigued by Intel's.
Why?

Intel have the Killing Roadmap (think Monty Python) that makes
executives immediately bet their company's future on the roadmap.
Worked for Compaq, SGI, HP, and now Apple.

- anton

 

[EdG]

We'll see this CPU introduced in the blueberry iMac in Q3 2006 ;-)

Maybe 2016

 

[Yousuf Khan]

That's like saying none of the knowledge gained from
the P6 design carried over to the Williamette design.

You might as well say that the knowledge they gained in Hyperthreading
will help them design an SMT Xscale. Sure, it might, but it's like
textbook knowledge -- they are still at square one the next time around.

About the Mac, Jobs looked at the roadmaps of both Intel and AMD.
There were other factors, but Jobs was more intrigued by Intel's.
Why?

That falls into the wishful thinking category. It's been stated that
Intel is going to have a new secret weapon that will just blow AMD right
out of the water next year; this has been stated in the same way for the
past three years now. So far everything Intel has turned out has been
laughably outclassed by the competition. Even the rumours so far have
Intel only developing processors next year that are still behind what
AMD already had in 2003.

Yes, I'm sure Jobs has looked at their roadmap. Perhaps the most
intriguing roadmap that he saw was Intel's MDF roadmap? That is, MDF =
market development funds.

Turning to the server-blade market, when you need more racks of blades,
the consideration is packing as much processing performance per watt
(and space).

That seems to be one of the biggest reasons why AMD is doing so well in
servers now. They have special ultra-low-power Opterons doing 30W for
this market.

There is only so much instruction parallelism in one thread of code.
Single-threaded superscalars core often stall because of
data dependencies, resource locks, cache misses,etc.
Multi-threaded (parallel) cores are the future.

Again, another reason why AMD is doing so well these days. They brought
64-bit x86 out first, which was intriguing, but they really ignited the
rockets once dual-core was introduced.

Yousuf Khan

 

[keith]

Intel have the Killing Roadmap (think Monty Python) that makes executives
immediately bet their company's future on the roadmap. Worked for Compaq,
SGI, HP, and now Apple.

You got it! They'll simply bite the leggs off the competition. ;-)

 

[keith]

On Sun, 25 Sep 2005 19:05:26 -0400, Yousuf Khan wrote:

Again, another reason why AMD is doing so well these days. They brought
64-bit x86 out first,

While Intel was dragging heels, thinking the world would adore Itanic
because it was 64bit, or sumpin'.

which was intriguing, but they really ignited the rockets once
dual-core was introduced.

I disagree, sorta. Intel cooled its jets with Itanic. They haven't
yet recovered from that disaster plan. OTOH, AMD has executed their
plan rather well.

 

[Del Cecchi]

You got it! They'll simply bite the leggs off the competition. ;-)

leggs? The competition wears nylon stockings or pantyhose? kinky, dude.

 

[keith]

leggs? The competition wears nylon stockings or pantyhose? kinky, dude.

I thought you knew me better. ;-)

 

[Del Cecchi]

I thought you knew me better. ;-)

never saw you without long pants. Too cold for shorts up there in the
green mts? Or you just can't spell?