How many props equal a Cray 1a Supercomputer?

Humanoido

There are various issues on floating point, speed, data size, etc.

humanoido

Post Edited (humanoido) : 3/20/2010 10:07:54 AM GMT

heater

More than I can afford. Even if I sold my home.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.

Leon

It's a silly question, no amount of Propellers (or any other MCU) could have the same performance as a supercomputer. Which Cray, BTW, there were lots of them.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Leon Heller
Amateur radio callsign: G1HSM

Post Edited (Leon) : 3/19/2010 11:35:47 AM GMT

Dave Hein

According to Wikipedia, Cray delivered a system with 1.75 petaflops -- that's 1.75e+15 floating point operations per second.· A prop has a capability of 160 MIPs.· It would require 10 million props to equal 1.6e+15 instructions per second, and it requires many instructions to implement a 64-bit floating-point operation.· So I would guess the number is around a billion props to equal a Cray.

Leon

It wouldn't work, though. The Propeller doesn't have an external memory interface, and how are they going to communicate? Also, the Crays were vector processors with eight 64 x 64-bit vector registers, and could perform operations on vectors in one instruction. That sort of thing would be impossible with Propellers.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Leon Heller
Amateur radio callsign: G1HSM

Post Edited (Leon) : 3/19/2010 1:41:34 PM GMT

Humanoido

I was referring to the first Cray ever made. It cost $8 million and performed at 80 million operations a second. One prop can do 200 MIPs. I think the question is reasonable. We compare apples and oranges over the decades. Agreed, there are some approximations to make for such a comparison between two computers that reach their goals using different methods.

humanoido

John R.

As Lean points out, raw MIPS is only part of the story. Communication between components (memory, other processors, etc.) and/or "throughput" (however you want to define that) are probalby more significant parts of the equation (no pun intended).

Extending the conversation, let's take inter prop communication. Assuming you will need some pins for other tasks, the best you could do would be a 16 pin parallel bus. That means 4 communication cycles to get one 64 bit number transferred.

Could you make it work, sure, but not anywhere close to "supercomputing" speeds.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
John R.
Click here to see my Nomad Build Log

wjsteele

Hmmm... I worked on a Cray X-MP several years ago. (Do I really have to say how many?) Anyway, I remember when we upgraded to the Y-MP. That was a vast change. I can't even see a Prop being anywhere near the power of a single X-MP processor. The X-MP/48 we used had 4 CPUs... each was significantly faster (at work, not clock speed) doing it's thing.

I think the X-MP only ran at something like 105mhz, but it could address the memory blazingly fast. (In fact, it could read and write to the memory simultaneously due to a dual bus arrangement.) It was a 64 bit bus as well.

We were able to do some complex 3D simulations (for their time) in incredible speeds... days instead of years. The processor actually had instructions on doing matrix transformations right in, like today's video cards do. Back then, that was amazing.

Now, considering that a simple video card of today can do way more work than the X-MP or even the Y-MP, that is amazing.

Bill

LoopyByteloose

Video cards! Apparently some Russian hackers have found that one can use Nvidia cards to crack encryptions. A comparison of them to a Cray might be more appropriate than a Propeller/Cray comparison. In some cases, the video card is more computer than the PC it supports.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Ain't gadetry a wonderful thing?

aka G. Herzog [noparse][[/noparse] 黃鶴 ] in Taiwan

Leon

I put one of the new nVidia cards in my desktop PC, the chip has 48 processors on it. They are working on one with 512 processors.

The last time I was in Paris I visited the Musee des Arts et Metiers. They have an early Cray there, but I can't remember which one, and I can't find it on the web site. It was brown, so it was probably a Cray-1:

en.wikipedia.org/wiki/File:Cray-1-deutsches-museum.jpg

That looks like it, anyway. It was fenced off, so I couldn't sit on the seating over the power supplies.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Leon Heller
Amateur radio callsign: G1HSM

Post Edited (Leon) : 3/19/2010 3:53:12 PM GMT

heater

I presume the Crays 80 million operations a second were floating point operations. So divide the props MIPS by 50 or a hundred, whatever it takes to do similar ops.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.

Dave Hein

humanoido said...
I was referring to the first Cray ever made. It cost $8 million and performed at 80 million operations a second. One prop can do 200 MIPs. I think the question is reasonable. We compare apples and oranges over the decades. Agreed, there are some approximations to make for such a comparison between two computers that reach their goals using different methods.

Oh, I thought you were talking about the current Cray computer.· Yes, IC technology has improved vastly over the past 25 years.· A small video camcorder today has more compute than a video system I worked on 25 years ago that took up an entire equipment rack.
The Cray-1A had a theoretical instruction rate of 160 MIPS -- the same as the prop at 80 MHz.· It was capable of about 250 MFLOPs.· So for non-floating point operations, the prop is about the same as a Cray-1A.· For 64-bit floating point operations, which is what the Cray was designed to do, you would need something on the order of 50 to 100 props.· And, as others have mentioned, you would need an efficient way to do data I/O.

Phil Pilgrim (PhiPi)

Dave Hein said...
Oh, I thought you were talking about the current Cray computer.

He probably was, but changed his mind when the absurdity was pointed out.

-Phil

HollyMinkowski

Forget the prop....use multi core video card processors instead.

But it will be the software to utilize a bunch of them that will be hard to create

Really powerful processors are actually pretty cheap now....game consoles contain
great processors at a very low cost.

Leon

The Cell chips used in the Playstation have eight vector processors.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Leon Heller
Amateur radio callsign: G1HSM

Humanoido

Dave Hein said...
Oh, I thought you were talking about the current Cray computer.
Phil said...
He probably was, but changed his mind when the absurdity was pointed out.

* I was never talking about the current Cray computer
and should have specified Cray 1 in the subject (will edit it now).
Thanks for pointing this out.

Dave Hein said...
The Cray-1A had a theoretical instruction rate of 160 MIPS -- the same as the prop at 80 MHz. It was capable of about 250 MFLOPs. So for non-floating point operations, the prop is about the same as a Cray-1A. For 64-bit floating point operations, which is what the Cray was designed to do, you would need something on the order of 50 to 100 props. And, as others have mentioned, you would need an efficient way to do data I/O.

* This is more on the order of what I was thinking about. I think modeling the Cray 1a with Prop chips would be a very interesting retro project.

* Other ideas hit upon using the PlayStation and Graphics cards. These are powerful and would make good prop coprocessors, but they cannot be used - this theoretical project would have modeling only from props.

* Some executable ideas for cross modeling floating point and increasing I/O efficiency from the experts would be helpful.

* humanoido

localroger

There is no "current Cray computer" in the sense of cool C-shaped cabinets with Freon running through their circuit boards. All recent supercomputers depend on massive parallelism; that "current Cray" is a cabinet full of AMD Opteron Linux boxes. I understood that humanoido was musing about the Cray-1, as the question didn't make much sense otherwise.

pharseid

I don't know if you would consider this cheating, but I would use FPGA's with flash multipliers to create vector math units and handle communications. If you had pipelined floating point units, a single unit would probably be on a par with the original Cray1, but programming it would be a headache. I also don't think a little FPGA on a board alongside the Prop would give you the visual you're looking for.

-phar

Humanoido

"The Cray-1 was a supercomputer developed during the early 1970's by Seymour Cray and his company Cray Research, it is considered to be one of the most famous and successful supercomputers of all times. The specs were extraordinary in 1976: the Cray-1 was a 64-bit system running at 80 MHz, addressing was 24-bit for a maximum of 8 MB of main memory.

High performance microprocessors didn't exist yet, therefore the new machine used a large number of high speed integrated circuits (ICs) with a total of about 200,000 gates, a complexity comparable to the Intel 80386 which became available 10 years later. The main register set consisted of eight 64-bit scalar registers and eight 24-bit address registers, plus 64 shadow registers and eight 64 bit vector registers. The system contained four buffers that could pipeline 64 instructions and feed the 12 functional units.

The indicated performance was 160 MIPS, when execution real world applications the system generally offered a performance of about 136 megaflops, with peaks of up to 250 megaflops when running highly optimized software. Since 1978 the Cray-1 was running the Cray Operating System (COS), later machines were running UNICOS, Cray's UNIX derivate.

...The system weighed 5.5 tons including the freon refrigeration system, the complete system consumed an incredible 250 kW of power when running.

Today a common Mac or PC is about 100 times faster than a Cray-1."
www.kirps.com/web/main/_blog/all/cray-1-a-supercomputer-way-ahead-of-its-time.shtml

The CRAY 1-A was a single processor computer capable of a sustained performance of 50 Megaflops (50 million arithmetic operations per second).
www.ecmwf.int/about/computers.html

Humanoido

Some insight into the internal wiring of the Cray 1.

Phil Pilgrim (PhiPi)

This thread has diverted (abetted by the OP, no less) from any substantive connection to the Propeller. Perhaps it should be moved to the Sandbox.

-Phil

heater

Phil, not quite. All that white wiring jumble in the Cray picture looks like the spaghetti monster that was my first home made Prop Demo Board style board.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.

Humanoido

Phil said...
This thread has diverted (abetted by the OP, no less) from any substantive connection to the Propeller. Perhaps it should be moved to the Sandbox.

It looks like a prop topic to me as we are examining the architecture of the Cray-1 and looking at ways to create a retro model using a certain number of Propeller chips. If determining how many prop chips equals a Cray 1a supercomputer is a sandbox topic, then by all means the moderator may move it. But as they say, if it looks like a prop, acts like a prop, walks like a prop, and quacks like a prop... it must be a prop. [noparse]:)[/noparse]

Grouping some basic ideas together, I think there's a plural modeling approach to the Cray 1 that determines prop configuration. Some elements are:

1) Matching Speed
2) Matching Architecture
3) Making an Approximation

humanoido