So my G5 have 4 years old !!
last G5 is the PPC970MP Bi-core run at 2,5Ghz in Apple box but can run at 2,7ghz (3Ghz in single-core) and it's almost 2 years old and 90nm fab against 65nm for core 2 duo.
The POWER5 have 3 years old run at 2,3Ghz (two times faster than G5 in SPEC2000 results) 90nm fab
the POWER6 is very recent (april 2007) run at 4,7Ghz (two times faster than POWER5 in SPEC2006.) 65nm fab
The G5 is a modified POWER4 core (6 years old 1,9Ghz max in 130nm)
POWER5 and 6 consume lot of power 160 watts but they are mainframe's CPU.
Folding@home, which is not a BOINC project, has reached a total computing power of 1109 teraflops, that is 1.109 petaflops. 76% of it comes from the CELL processor of Playstation3. It is said that Sony is going to sell fabrication and development of CELL to Toshiba. Go figure!
Tullio
Folding@home, which is not a BOINC project, has reached a total computing power of 1109 teraflops, that is 1.109 petaflops. 76% of it comes from the CELL processor of Playstation3. It is said that Sony is going to sell fabrication and development of CELL to Toshiba. Go figure!
Tullio
CELL have been developped by IBM,Sony and toshiba and it's a PowerPC CPU :)
I have been told that BOINC reaches a total computing power of 562 Tflops, while Folding@home reaches 1110 Tflops all by itself. But I am wondering whether Folding@home crunches data in single precision, while many BOINC projects use double precision. Are we not comparing apples with oranges?
Tullio
FAH developed a new app to crunch in SP to take advantage of the Cell/Gpu architectures. AFAIK their main apps are all DP because the SP one needs significantly more FLOPs to complete a given amount of work.
FAH developed a new app to crunch in SP to take advantage of the Cell/Gpu architectures. AFAIK their main apps are all DP because the SP one needs significantly more FLOPs to complete a given amount of work.
This reinforces my suspicion that our (BOINC) Flops are not comparable to their CELL Flops since they need more flops to perform a given amount of work. Is it so?
Tullio
Many BOINC apps use double precision float numbers, afaik. the Cell CPU cannot do those on the hardware.
Folding can run parts of the calculation on high end video cards too btw.
p.s.: This thread has the flops count values for the Cell, as you can see, it is great on single precision and falls behind current x86 based CPUs on double precision.
There are BOINC applications that need a good integer speed too or high memory throughput for rearranging arrays. I don't think that there is any ideal CPU for all projects, each pair of CPU/project has to be evaluated separately.
I've been using the SMP app at FAH for some time on a pair of x360's and a Z Pro. In general it works pretty good, although it does get a little flaky at times. Was getting 10 cpu days worth of work done in 2.5 days. The SMP app seems to be quite sensitive to L2 and L3 cache size, the bigger the better. They have a good range of app's tailored to take advantage of the strengths of the different processors.
Many BOINC apps use double precision float numbers, afaik. the Cell CPU cannot do those on the hardware.
Folding can run parts of the calculation on high end video cards too btw.
p.s.: This thread has the flops count values for the Cell, as you can see, it is great on single precision and falls behind current x86 based CPUs on double precision.
There are BOINC applications that need a good integer speed too or high memory throughput for rearranging arrays. I don't think that there is any ideal CPU for all projects, each pair of CPU/project has to be evaluated separately.
That is incorrect. Cell can do doubles in hardware with IEEE 754 precision, but it cannot do them as efficiently as possible because it is optimized for games. The double precision support is an afterthought on those chips used to broaden the chip's possible uses. However, the sheer number of FPUs on that chip (eight mini vector FPUs exist, one in each of the eight SPU's, and one scalar FPU in the scalar CPU, which might also have another mini vector FPU in it as well for a total of nine or ten FPUs per Cell Broadband Engine) makes it attractive for scientific computation.
The Cell Broadband Engine's FPUs cannot handle single precision floats correctly because they contain optimizations that make them less precise on extremely small numbers so they can be faster. These optimizations speed up games and MPEG-2/H.264/VC-1 decoding where maximum accuracy does not matter, but make single precision math useless for scientific work (because they introduce another source of error).
So my G5 have 4 years old
)
So my G5 have 4 years old !!
last G5 is the PPC970MP Bi-core run at 2,5Ghz in Apple box but can run at 2,7ghz (3Ghz in single-core) and it's almost 2 years old and 90nm fab against 65nm for core 2 duo.
The POWER5 have 3 years old run at 2,3Ghz (two times faster than G5 in SPEC2000 results) 90nm fab
the POWER6 is very recent (april 2007) run at 4,7Ghz (two times faster than POWER5 in SPEC2006.) 65nm fab
The G5 is a modified POWER4 core (6 years old 1,9Ghz max in 130nm)
POWER5 and 6 consume lot of power 160 watts but they are mainframe's CPU.
Folding@home, which is not a
)
Folding@home, which is not a BOINC project, has reached a total computing power of 1109 teraflops, that is 1.109 petaflops. 76% of it comes from the CELL processor of Playstation3. It is said that Sony is going to sell fabrication and development of CELL to Toshiba. Go figure!
Tullio
RE: Folding@home, which is
)
CELL have been developped by IBM,Sony and toshiba and it's a PowerPC CPU :)
I have been told that BOINC
)
I have been told that BOINC reaches a total computing power of 562 Tflops, while Folding@home reaches 1110 Tflops all by itself. But I am wondering whether Folding@home crunches data in single precision, while many BOINC projects use double precision. Are we not comparing apples with oranges?
Tullio
FAH developed a new app to
)
FAH developed a new app to crunch in SP to take advantage of the Cell/Gpu architectures. AFAIK their main apps are all DP because the SP one needs significantly more FLOPs to complete a given amount of work.
RE: FAH developed a new app
)
This reinforces my suspicion that our (BOINC) Flops are not comparable to their CELL Flops since they need more flops to perform a given amount of work. Is it so?
Tullio
Many BOINC apps use double
)
Many BOINC apps use double precision float numbers, afaik. the Cell CPU cannot do those on the hardware.
Folding can run parts of the calculation on high end video cards too btw.
p.s.: This thread has the flops count values for the Cell, as you can see, it is great on single precision and falls behind current x86 based CPUs on double precision.
There are BOINC applications that need a good integer speed too or high memory throughput for rearranging arrays. I don't think that there is any ideal CPU for all projects, each pair of CPU/project has to be evaluated separately.
I've been using the SMP app
)
I've been using the SMP app at FAH for some time on a pair of x360's and a Z Pro. In general it works pretty good, although it does get a little flaky at times. Was getting 10 cpu days worth of work done in 2.5 days. The SMP app seems to be quite sensitive to L2 and L3 cache size, the bigger the better. They have a good range of app's tailored to take advantage of the strengths of the different processors.
RE: Many BOINC apps use
)
That is incorrect. Cell can do doubles in hardware with IEEE 754 precision, but it cannot do them as efficiently as possible because it is optimized for games. The double precision support is an afterthought on those chips used to broaden the chip's possible uses. However, the sheer number of FPUs on that chip (eight mini vector FPUs exist, one in each of the eight SPU's, and one scalar FPU in the scalar CPU, which might also have another mini vector FPU in it as well for a total of nine or ten FPUs per Cell Broadband Engine) makes it attractive for scientific computation.
The Cell Broadband Engine's FPUs cannot handle single precision floats correctly because they contain optimizations that make them less precise on extremely small numbers so they can be faster. These optimizations speed up games and MPEG-2/H.264/VC-1 decoding where maximum accuracy does not matter, but make single precision math useless for scientific work (because they introduce another source of error).
Thanks for explaining, my
)
Thanks for explaining, my (mis)information has been from the time before PS3 was out - probably just gossips back then.