Sign in with
Sign up | Sign in
Your question

Are cpus losing their clout? - Page 2

Tags:
Last response: in CPUs
Share
November 7, 2009 6:37:39 PM

Even if Fermi does nothing more arch wise compared to the 280, itll still be 3x faster, thats not counting the DP at 8x, not the error corrections, and Im talking better than GDDR5 corrections,actually the entire arch is better faster, and more geared for this, it wouldnt surpise me on this workload, wich makes it 25% faster than LRB at 32 cores here
m
0
l
November 7, 2009 6:42:29 PM

The flops listed are DP, and thats up 8x, plus 2.13 more shaders, I think, plus all the error corrections and approach, and the much improved,larger cache
m
0
l
November 7, 2009 6:43:38 PM

I hate having to think sometimes, last figures are closer to reality, sorry heheh
m
0
l
November 7, 2009 6:48:47 PM

Plus, Elmo, being familiar with the gfx market, by the end of 2010, the new Fermi will be 60% faster
m
0
l
a b à CPUs
November 7, 2009 6:58:23 PM

JAYDEEJOHN said:
The flops listed are DP, and thats up 8x, plus 2.13 more shaders, I think, plus all the error corrections and approach, and the much improved,larger cache

All that and it still cannot run a large portion of the C++, Fortran and C library.

Intel took a slightly different approach. Intel went MIMD rather than SIMD with Larrabee. That is to say that Larrabee contains full x86 cores each capable of Multiple Instruction streams and Multiple Data streams. Therefore each core can do multiple things at once (vs. SIMD which can only handle a Single Instruction stream per core).

And if you want to push it further RV870 is a Super Scalar SIMD design (Capable of Multiple Threads per Instruction).

Fermi is nothing fancy at all (to be quite honest). It will output around 1.25-1.5TFlops of performance in SP and around 600-677GFLOPS in DP (RV870 can handle 2.72TFLops in SP and 544GFlops in DP).

Fermi isn't as giant of a leap forward as some make it out to be.
m
0
l
November 8, 2009 12:37:35 AM

but from the rasterized graphics standpoint, it will still probably have the fastest gpu on the market (depending on whether ATI/AMD does some kind of 4890 thing with the 5870)
m
0
l
November 8, 2009 1:22:04 AM

No one knows how well Fermis compute density will turn out, and whether they catch ATI or not.
If they do, and having the advantages they look to be implementing in the new arch, with what we already know about the old arch, itll be close to what I said earlier.
Theres too many unknowns at this point
m
0
l
November 8, 2009 1:37:53 AM

JAYDEEJOHN said:
No one knows how well Fermis compute density will turn out, and whether they catch ATI or not.
If they do, and having the advantages they look to be implementing in the new arch, with what we already know about the old arch, itll be close to what I said earlier.
Theres too many unknowns at this point

as with any unreleased technology
m
0
l
November 8, 2009 2:17:37 AM

Thing is, this is a whole new arch, with a better approach to gpgpu, with more arch improvements and more than double what the older one had, which is maybe 1/3 as fast as LRB at 32 cores.
Some arch changes we know shows 2.13 times improvements, plus the claimed 8x DP improvements.
It does get fuzzy, as even these numvers should work much better on the new arch, it appears it could very well be a monster.
Addressing the SW issues, alot of what this HW is for is already being done in nVidias SW, and theres designs/apps out there already using it.
For the HPC scenario, its all case specific anyways, and there too, its already there and killing cpu setups.
Not making claims here, just going on whats been known, and whats the possible perf of Fermi, and even so, I could be underestimating it as well, or way overestimating it, tho, I dont think a new arch geared for this, showing at least 4x the old less prepared arch, will somehow do worse
m
0
l
November 8, 2009 2:27:28 AM

well for gpgpu to be the important and "got to have" feature with fermi and lrb, there has to be a lot more software that uses gpgpu. I havent come across one that i use personally (except Avivo video encoding, which i rarely use). They need to push for a central computing language (OpenCl hopefully)
m
0
l
November 8, 2009 2:50:16 AM

Certainly they do, as CUDA runs with it. As could LRB as well, offering more approaches, and could also show improvements as well using OCL, depending on its libraries.
Thats the point here.
The longer LRB takes, and weve recently mlearned that, guess what? the SW is causing some delays.
So, the SW solution given as superior isnt necesarily the case, and having HKMG, a smaller process and if it doesnt dominate, Fermi will once these too are applied.
Theres just alot more growth potential process wise for Fermi, as well as setting a good SW base, so the sooner the better for LRB, or itll never get traction, which goes to my comments in other threads about the ever moving target of gpus
m
0
l
a c 127 à CPUs
November 8, 2009 3:19:07 AM

I personally don't think the CPU was over emphasized but more that until recent years a lot of the other hardware evolved slowley.

Now thanks to advanced process technology and advanced RAM and so on other components have been able to grow as fast as CPUs have.

GPUs have only grown this fast since I would sy the 9700Pro hit and made nVidia think twice.
m
0
l
November 8, 2009 3:23:32 AM

Cant wait for HKMG on gpus, itll be like a free shrink!!!!!!!
m
0
l
      • 1
      • 2 / 2
!