RISC or CISC?

Is the current incarnation of nVidia's latest GPU's closer to RISC or
CISC architecture?
Since modern GPUs have such floating point computational power, I was
think they're closer to RISC.

* Justin:
> Is the current incarnation of nVidia's latest GPU's closer to RISC or
> CISC architecture?
> Since modern GPUs have such floating point computational power, I was
> think they're closer to RISC.

RISC has nothing to do with "floating point computational power".

The differentiation between RISC and CISC was a valid one maybe 15 years
ago, but not anymore. Traditional RISC processors became more and more
CISC, and traditional CISC platforms like x86 became much more RISC-like
since the Pentium Pro 13 years ago. Today CPUs are somewhere in between
these two categories.

GPUs are not really comparable to general purpose CPUs but more to DSPs.
But generally, the same that is valid for modern CPUs is true for GPUs.

Ben

* Justin:
> Is the current incarnation of nVidia's latest GPU's closer to RISC or
> CISC architecture?
> Since modern GPUs have such floating point computational power, I was
> think they're closer to RISC.

RISC has nothing to do with "floating point computational power".

The differentiation between RISC and CISC was a valid one maybe 15 years
ago, but not anymore. Traditional RISC processors became more and more
CISC, and traditional CISC platforms like x86 became much more RISC-like
since the Pentium Pro 13 years ago. Today CPUs are somewhere in between
these two categories.

GPUs are not really comparable to general purpose CPUs but more to DSPs.
But generally, modern GPUs are probably closer to VLIW than to RISC or CISC.

Ben

Benjamin Gawert wrote:
> * Justin:
>> Is the current incarnation of nVidia's latest GPU's closer to RISC or
>> CISC architecture?
>> Since modern GPUs have such floating point computational power, I was
>> think they're closer to RISC.
>
> RISC has nothing to do with "floating point computational power".
>
> The differentiation between RISC and CISC was a valid one maybe 15 years
> ago, but not anymore. Traditional RISC processors became more and more
> CISC, and traditional CISC platforms like x86 became much more RISC-like
> since the Pentium Pro 13 years ago. Today CPUs are somewhere in between
> these two categories.
>
> GPUs are not really comparable to general purpose CPUs but more to DSPs.
> But generally, modern GPUs are probably closer to VLIW than to RISC or
> CISC.
>
> Ben

I htink I understand. CISC and RISC are outdated terms; and now the
latest is Very Long Instruction word. I remember those corns Apple
commercials where they said their G4 was twice as fast as the then
prevalent Pentium II. Then I looked at Sun's UltraSPARC T2 Plus which
broke the 1 TeraFlop record back in 2007.
Is there a chart somewhere where I can see each processor and their
respective FLOPs

* Justin:

> I htink I understand. CISC and RISC are outdated terms; and now the
> latest is Very Long Instruction word.

Sorry, but I don't think you really understand. CISC, RISC and VLIW are
different concepts, all of them are quite old, and all of them have
advantages and disadvantages. None of them is necessarily better than
than the other.

> I remember those corns Apple
> commercials where they said their G4 was twice as fast as the then
> prevalent Pentium II.

Apple said a lot of things which weren't necessarily true.

> Then I looked at Sun's UltraSPARC T2 Plus which
> broke the 1 TeraFlop record back in 2007.

The UltraSPARC T2 is more in the area of 11GFLOPS:
<http://blogs.sun.com/deniss/entry/floating_point_performance_on_the>

> Is there a chart somewhere where I can see each processor and their
> respective FLOPs

There are websites which show LINPACK (the tool which is used to measure
the FLOPS performance of supercomputers) results for various computers.
However, the FLOPS performance depends on many factors and especially on
the memory architecture. Besides that, it is a highly theoretical number
that says very few about about real-world performance (for example, the
Japanese supercomputer MDGRAPE-3 which does over one PFLOPS can't run
the Linpack benchmark), and it says *nothing* about the real-world
performance outside specialized HPC (high performance computing)
applications. In short, if you are not building the next world largest
supercomputer, FLOPS are pretty meaningless.

Benjamin

Benjamin Gawert wrote:
> * Justin:
>
>> I htink I understand. CISC and RISC are outdated terms; and now the
>> latest is Very Long Instruction word.
>
> Sorry, but I don't think you really understand. CISC, RISC and VLIW are
> different concepts, all of them are quite old, and all of them have
> advantages and disadvantages. None of them is necessarily better than
> than the other.
>
>> I remember those corns Apple commercials where they said their G4 was
>> twice as fast as the then prevalent Pentium II.
>
> Apple said a lot of things which weren't necessarily true.
>
>> Then I looked at Sun's UltraSPARC T2 Plus which broke the 1 TeraFlop
>> record back in 2007.
>
> The UltraSPARC T2 is more in the area of 11GFLOPS:
> <http://blogs.sun.com/deniss/entry/floating_point_performance_on_the>
>
>> Is there a chart somewhere where I can see each processor and their
>> respective FLOPs
>
> There are websites which show LINPACK (the tool which is used to measure
> the FLOPS performance of supercomputers) results for various computers.
> However, the FLOPS performance depends on many factors and especially on
> the memory architecture. Besides that, it is a highly theoretical number
> that says very few about about real-world performance (for example, the
> Japanese supercomputer MDGRAPE-3 which does over one PFLOPS can't run
> the Linpack benchmark), and it says *nothing* about the real-world
> performance outside specialized HPC (high performance computing)
> applications. In short, if you are not building the next world largest
> supercomputer, FLOPS are pretty meaningless.
>
> Benjamin


So a CPU with a high FLOP rating could be slow in the real world.
Sort of like having a 1,500hp Toyota Supra gets beat by a 600hp
Corvette. If the CPU can't apply its power to real world applications
its basically spinning its wheels at the start line.

Steve wrote:
> In article <hbl4fn$np9$-
> september.org>,
> says...
>> Is the current incarnation of nVidia's latest GPU's closer to RISC or
>> CISC architecture?
>> Since modern GPUs have such floating point computational power, I was
>> think they're closer to RISC.
>
>
> Maybe you would be interested in...
>
> http://www.cpubenchmark.net/index.php
>
>
>
>
>
>
> s

Interesting, page. I wish they included the Sparcs and various
different technologies. But that's that's still fascinating.

* Justin:

>> Maybe you would be interested in...
>>
>> http://www.cpubenchmark.net/index.php
>
> Interesting, page. I wish they included the Sparcs and various
> different technologies.

How can they if this are the results of their own benchmark program
which runs on Windows only, an operaing system which simply does not run
on SPARC and most other non-x86 architectures? Besides that, the list
shows the results users have when running this synthetic benchmark on
their systems, which is not necessarily representative of real-world
application performance.

f you want a benchmark that covers more architectures look at the SPEC
CPU2006 suite:

<http://www.spec.org/results_search.html>

It gives you results for integer and floating point performance based on
a standardized suite of tests, and unlike the Passmark CPU benchmark
from cpubenchmark.net the SPEC Suite is a widely recognized benchmark.
However, like with all benchmarks, real world performance can differ.

Benjamin

* Justin:

> So a CPU with a high FLOP rating could be slow in the real world.
> Sort of like having a 1,500hp Toyota Supra gets beat by a 600hp
> Corvette. If the CPU can't apply its power to real world applications
> its basically spinning its wheels at the start line.

A benchmark does only show how fast a certain system (not only the CPU!)
is at running that benchmark and nothing else. How relevant this is to
the real world depends on how good the benchmark simulates a real
application environment. The LINPACK which is used to determine FLOPS
ratings is quite old and simulates only a very specific area of
synthific computing. All a FLOPS number tells you is how fast a system
is in this specific area, it tells you *nothing* about the performance
for other operations. GPUs are a good example because they are fast at
only very specific operation. For example, a Geforce 8800GTX can do
518GFLOPS, most Core 2 Duo processors are in the 2-4GFLOPS range.
However, a 8800GTX can only do 518GFLOPS for very few operatings in
single precision, in dual precision the FLOPS value falls rock-bottom. A
88000GTX GPU would also be ****ing slow in running any standard
application, much slower than the Core 2 Duo. If you even get to run it
on the GPU, that is, because a GPU is something very different from a
CPU, a GPU is more like a DSP.

Benchmarking ismuch more than just comparing numbers. You need to really
understand what you test, how the systems work, you need a standardized
test environment, and you need to understand the limits of the
benchmark. Just comparing two numbers to know which system is faster
doesn't work, period.

Ben