i realize this a SGI thread but maybe the price is going down because TERA willbe on board soon???
many references are made in this thread to CRAY so i thought i could get a few opinions here:
If anyone else is acturally reading this, Grazak's comments
and questions are preceeded by 1 or 3 ">" marks
and mine are preceeded by 0 or 2 ">" marks.
>>> OK, so one MTA CPU can do an integer sort
>>> faster than a Cray CPU. Big deal.
Yes, I remember this "question about our company".
Sort of set the tone for your entire note.
>> Ahem. It's faster than any other CPU in the world.
>It's also more expensive than any other commercial CPU in the world.
Nope, Cray T90's are more expensive per CPU.
And we're faster on this benchmark. And since they're our
main competition, it is a big deal.
>I'll take a dozen Pentium-Pro or PowerPC CPUs running parallel over
>a single Tera CPU. They may not be as fast but they're certainly
>cheaper.
We're not aiming at problems that can be well served by a dozen
micros. We're aiming at markets that are not well served by anything.
We want to avoid competition so we can establish a virtual monopoly
and make a lot of money for our shareholders.
>And because they're so much cheaper, I can scrap the lot >when something more advanced comes along. In the end I'll do much
>better than if I had put my entire equipment budget for the next
>five years into one machine from a financially insecure, virtually
> unknown company.
So, you propose assembling your own machine (from dozens or hundreds
or thousands) of micros, writing your own OS and compilers,
trying to get your applications to runs fast on this kludge,
and then throw the whole mess away every five years?
Most companies don't want to get involved in such a project.
They want to buy a machine that runs fast, has software that works,
and helps them get they're work done (versus going into the computer
development business).
>>>What's that got to do with multi-threading or multi-processing
>> Everything.
>Nothing. The sorting benchmark was performed in serial, not parallel.
>Has Tera released any parallel benchmarks that I don't know about?
The sorting benchmark was definitely run in parallel.
There's no way we (or anyone else) could run this fast without
parallelism.
Basically, you are confusing "1 processor" with "serial".
Are processors are multithreaded. Mean each processor runs multiple
threads in parallel, trying to stay busy. Cray's use vector
processors, meaning they try to stay busy by crunching vectors
at a time. And most micros rely on instruction-level parallelism
for their speed. For this particular benchmark, multithreading
turns out to be the most effective form of parallelism.
>Considering how fast the PC industry responds (six-month
>obsolescence cycles), how much time do you think Tera has
>left before there's a cost-effective, scalable parallel desktop
>computer.
Years. And even then, we don't care. We're not trying to put
a machine on a desktop. We're trying to build the fastest
general-purpose machine in the world. When people are putting
16 GFlop machines on the desktop, (currently, PCs are about 100 MFlops, so they've got a ways to go) we'll be working on 10 TFlop
machines.
>>> Cray CPU's aren't particularly good with integers but
>>> they're second-to-none for floating point calculations.
>> Where do you get these assertions? And why do you feel the need
>> to share them with us? Don't you imagine we know what Cray
>> machines can do?
>Besides, you haven't refuted the assertion.
Check out
Computer Architecture: A Quantitative Approach
John Hennessy and David Patterson
Morgan-Kauffman, 1990
In Chapter 7, Vector Processors,
section 8, Fallacies and Pitfalls,
at the bottom of page 391, they write
Fallacy: The scalar performance of the best supercomputers is low.
The supercomputers from Cray Research have always had good
scalar performance. [then goes on to compare the performance
of an old Cray with a modern workstation on a scalar benchmark.
The Cray was much faster.]
>If you think the Tera MTA is competing with the Cray T90 you're
>deluding yourself - it's possible to buy a working T90 right now.
>Tera is more likely to find itself competing with the Origin2000.
How can you think this? Certainly not from anything Tera has ever
said. We cost about the same as a T90, processor for processor.
And we perform about the same as a T90, processor for processor.
The difference is that we plan to build bigger machines.
If you really belive we intend to compete with SGI in the area of
their mid-level machines (the Origin series), you've seriously
misunderstood our goals.
>Will the MTA be faster and cheaper than the competition's latest
>effort?
In terms of price/performance (or bang for the buck, if you prefer),
we expect to be cheaper than our competition.
>Are you actually shipping those machines? How much do they cost
>and how many have you sold so far?
We've announced all our sales and we'll annouce when we ship.
I don't think we've announced our prices.
Anyway, I'm getting a tired of arguing all these points so I'll stop.
I've got work to do and I doubt I'm making any headway.
I think the central question we all wonder about is "wll Tera suceed?"
I can't tell you, and debate at this point is pretty useless.
We'll deliver the machine to San Diego and those guys will
wring it out, running all their favorite benchmarks and applications
on it. They'll publish their impressions and measurements.
And we'll see what happens. Until then, speculation seems
fruitless.
Preston Briggs |
