Intel Larrabee GPU / GPGPU based on the old P54C Pentium ?

Intel's Larrabee GPU based on secret Pentagon tech, sorta [Updated]

By Jon Stokes

In a recent interview with German tech site heise.de, Intel's Pat
Gelsinger let slip* there's some interesting information that I've
been sitting on since last year, but haven't published: the processor
cores that make up Intel's forthcoming Larrabee GPU are based on the
original Pentium core design. Larrabee is, in essence, a bunch of P54C
(i.e. pre-MMX) Pentium cores that have been enhanced with very wide
vector floating-point resources and ganged together to make the brains
of a flexible, x86-based GPU.


*Update: Intel contacted me to clarify that Gelsinger did not, in
fact, tell heise.de that Larrabee was based on P54C. This was
something that was apparently mangled in the Babelfish translation,
and Andreas at heise.de was pulling the P54C detail from some other
source. Nonetheless, I had already heard the information about the
Larrabee-Pentium-Pentagon connection from a source of my own, so I do
vouch for it.

You're probably wondering why Intel picked the old P54C to base
Larrabee off of, instead of, say, the later P55 part with MMX. The
reason, which Gelsinger did not reveal to heise.de and which comes
from a source of mine, is a bit surprising, in that it involves the
Pentagon.
Say "Pentagon Pentium" five times fast

Quite some time ago, after the Pentium was obsolete and Intel had
moved on, the company gave the RTL code for the processor to the
Pentagon so that the military could continue to fabricate a radiation
hardened version of it for use in military applications. Trailing-edge
hardware like the Pentium has the advantage of having been thoroughly
tested and debugged (cf. the P54C's infamous FDIV bug), and at the
time the military had its own fab facilities that could do some low-
volume fabrication. (I'm sure they still have such facilities for
prototyping.) So the Pentagon cleaned up the P54C's RTL code and began
producing a rad-hard version of the chip for military use.

A few years later, when the Pentagon had moved on from the P54C, they
offered the RTL back to Intel. So Intel took the core, which has a
very small footprint and by this time had been pretty thoroughly
debugged, and modified it for use in the many-core chip that later
became Larrabee.

Speaking of five times fast

While I'm just spilling all kinds of Larrabee beans, I might as well
drop another, performance-related tidbit that came my way. In an
upcoming SIGGRAPH paper, Intel will claim that Larrabee has 20x the
performance per watt of a Core 2 Duo and half the single-threaded
performance. It also has a 4MB coherent L2, and three-operand vector
instructions.

Note that I don't have any more context for the information I just
gave, so I'm not sure if "half the performance of Core 2 Duo" is a
clock-for-clock figure or not. If it is, recall that the Pentium's
pipeline is less than half the depth of Core 2's, so if the GPU does
debut in the 1.7GHz to 2.5GHz range, then that will help it in Core 2
comparisons. The other big unknown is the type of workload that Intel
is using for this ballpark performance figure. When it comes to in-
order (i.e., Larrabee) vs. out-of-order (i.e., Core 2) and short
pipeline vs. long pipeline, the type of workload involved makes all
the difference. So these few details that I've given tell you a lot
less than you might think at first; but something is better than
nothing—and nothing is what we have so far about Larrabee's
peformance.


http://arstechnica.com/news.ars/post...ech-sorta.html


Hmmm
If Intel is using old Pentium CPU technology in Larrabee, I hope they
also use old GPU tech for the rasterisation portions of the chip.
Specifically, Lockheed Martin Real3D graphics technology, which Intel
once owned the patents to, and probably still has knowledge of.

* AirRaid:
*Update: Intel contacted me to clarify that Gelsinger did not, in
fact, tell heise.de that Larrabee was based on P54C. This was
something that was apparently mangled in the Babelfish translation,

That's why it's rather stupid to write about an article that you even
don't understand.

Say "Pentagon Pentium" five times fast

The whole Pentagon storylone is just crap. Military agencies don't
produce microprocessors, they just buy them from certain vendors (intel
being one of them).

So just another AirRaid post on the usual bottom-low level.

Benjamin

On Jul 14, 9:42*am, Benjamin Gawert wrote:
* AirRaid:

*Update: Intel contacted me to clarify that Gelsinger did not, in
fact, tell heise.de that Larrabee was based on P54C. This was
something that was apparently mangled in the Babelfish translation,

That's why it's rather stupid to write about an article that you even
don't understand.

Say "Pentagon Pentium" five times fast

The whole Pentagon storylone is just crap. Military agencies don't
produce microprocessors, they just buy them from certain vendors (intel
being one of them).

So just another AirRaid post on the usual bottom-low level.

Benjamin

I didn't WRITE that you moron. Notice I included a ****ing LINK, you
ass.

Also, Military / Government agencies DO make their own
microprocessors, in smaller quantities than consumer/industrial,
etc.

* AirRaid:

[useless fullquote deleted]

I didn't WRITE that you moron. Notice I included a ****ing LINK, you
ass.

This was generally said you stupid retard. Not that I expect a
crossposting robot to understand that.

Also, Military / Government agencies DO make their own
microprocessors, in smaller quantities than consumer/industrial,
etc.

Really? So exactly which microprocessors were made by
government/military agencies?

Benjamin

Also, Military / Government agencies DO make their own
microprocessors, in smaller quantities than consumer/industrial,
etc.

Milspec processors are mfg by Texas Instruments, Intel, AMD, etc. and NOT by
any government agency. Hardening against EMP, X-Ray and Gamma radiation,
large thermal extremes, amongst other things are some of many differences.

Benjamin Gawert pravi:
Really? So exactly which microprocessors were made by
government/military agencies?

Chips on industrial boards made in Israel.

Just kidding...


-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFIfP6kB6mNZXe93qgRAqA/AKCwSqIWEQETTViwBEl1fnfTcv89QQCdHMzC
SvYijPc0sjUzfjGnOVsOEMA=
=Iszi
-----END PGP SIGNATURE-----

'Benjamin Gawert' wrote, in part:
Really? So exactly which microprocessors were made by government/military
agencies?
_____

Have you never heard of a 'Field FAB'? The 'Field FAB' is air
transportable, delivered by low-level parachute extraction, and uses NATO
7.62 mm wafers B^)

'Skybuck' and 'AirRaid' both have logorrhea, but at least the 'Skybuck'
makes up his own stuff.

Phil Weldon

"Benjamin Gawert" wrote in message
...
* AirRaid:

[useless fullquote deleted]

I didn't WRITE that you moron. Notice I included a ****ing LINK, you
ass.

This was generally said you stupid retard. Not that I expect a
crossposting robot to understand that.

Also, Military / Government agencies DO make their own
microprocessors, in smaller quantities than consumer/industrial,
etc.

Really? So exactly which microprocessors were made by government/military
agencies?

Benjamin

Augustus wrote:
Also, Military / Government agencies DO make their own
microprocessors, in smaller quantities than consumer/industrial,
etc.

Milspec processors are mfg by Texas Instruments, Intel, AMD, etc. and NOT by
any government agency. Hardening against EMP, X-Ray and Gamma radiation,
large thermal extremes, amongst other things are some of many differences.


I know that in the 80's and early 90's some chips were made by producers
who do not make any chips for commercial sale. When you are making odd
stuff in low volume there are a lot of place which can make low volume
chips and which have very secure development area. I would be amazed if
places like Sandia and Los Alamos didn't have in-house and very secure
small fabs.

This is not to say that commercial vendors aren't involved, because I
don't know, but for many of these applications a single wafer is a
lifetime supply, and research facilities, government, commercial and
educational, can do processing in that volume. Hardened sometimes means
"large process" to stand alpha hits, so it need not be the latest 350
angstrom process.

No commercial fab is going to shut down and convert to a RadHard or Mil
Spec process to knock out a few wafers, or even a few hundred thousand
copies of a specialty chip, set up is too expensive, so I think most
specialty chips are going to be done in specialty fabs.

When I used to sell computers to LANL, we did mil spec gas tight wire
wrap on the floppy drives instead of jumpers. Yes that was back in Z80
days, the original PC.

--
Bill Davidsen
He was a full-time professional cat, not some moonlighting
ferret or weasel. He knew about these things.

* Bill Davidsen:

I know that in the 80's and early 90's some chips were made by producers
who do not make any chips for commercial sale. When you are making odd
stuff in low volume there are a lot of place which can make low volume
chips and which have very secure development area. I would be amazed if
places like Sandia and Los Alamos didn't have in-house and very secure
small fabs.

They don't. None of the military or governmental agencies has any such
facilities. In all military projects (no matter how secret they are)
there are always commercial vendors included in one form or another.

This is not to say that commercial vendors aren't involved, because I
don't know, but for many of these applications a single wafer is a
lifetime supply, and research facilities, government, commercial and
educational, can do processing in that volume. Hardened sometimes means
"large process" to stand alpha hits, so it need not be the latest 350
angstrom process.

Chip fabrication is a extremely delicate process. Just setting up a fab
isn't enough, it takes a lot of effort to get the process running
effectively enough to not only produce key rings.

No commercial fab is going to shut down and convert to a RadHard or Mil
Spec process to knock out a few wafers, or even a few hundred thousand
copies of a specialty chip, set up is too expensive, so I think most
specialty chips are going to be done in specialty fabs.

While it might look to be like this to you it simply isn't the case.
There are enough commercial vendors who are security certified and also
have no problem doing small series of specialty chips. Companies like
intel or TI have much more than the fabs where they make their mass
market chips. And not to forget that most MIL-Std ICs while not reaching
the numbers of standard versions in fact are mass made articles today.

Benjamin