Computer Construction Conundrum
-
Nicholas Butler wrote:
What would you do with 48 cores?
i'd wonder what was going to become of those two bags of apples.
-
I'm in a niche ( ATM ) market, as I work from home writing concurrent software. The equivalent Intel chips are the 7500 Beckton series. These have 8 cores and 16 threads which according to reports beat the 12-core AMD 6100s. However, I can't find any in stock and the retail prices that are published are prohibitive.
Dan Neely wrote:
the guy who ~4 years ago got a 9x speedup vs the stock app by hacking the binary
Impressive :omg: Nick
Nicholas Butler wrote:
Dan Neely wrote: the guy who ~4 years ago got a 9x speedup vs the stock app by hacking the binary Impressive OMG
Yeah. It was really nice for a while since we were getting 9x as many points/day which helped a lot in the bragging rights contest. IIRC he got about ~4-5x from algorithmic changes that were incorporated in the next official app (moving stuff into lookup tables and tweaking working sets to fit inside L1 caches) and ~2x from writing better hot loops to eliminate pipeline stalls and incorporating SSE/SSE2 instructions. The original was pure x87, his replacements interleaved both x87 and SSE instructions to keep both of their pipelines fully loaded. It was eye opening after years of reading that a good compiler is almost impossible to beat these days. Granted that a good C/C++ programmer could have done most of the improvements in his code and used compiler options to get basic SSE1/2 versions, but IIRC we've had 30-50% speedups in new apps after they were regarded as stable enough to start mucking around with asm. Overkill for most uses probably, but at supercomputer work loads even a few percent can help; especially when the total problem set is large enough that probabilistic algorithms are needed to prune the data massively early on. Even with those at present signals need to be an order of magnitude larger than the sensors thresholds to be detected (below that the odds decline roughly linearly), the previous algorithm's sure detection threshold was +2 magnitudes. This is a major challenge when the strongest signal expected from a known pulsar is right at the theoretician detection threshold and is near a massive noise spike.
3x12=36 2x12=24 1x12=12 0x12=18
-
I don't understand why you feel the need to build a new machine. Compiler tech certainly is just barely using what you have now. My current quad-core box is 3-4 years old, and I feel no need to upgrade the hardware.
.45 ACP - because shooting twice is just silly
-----
"Why don't you tie a kerosene-soaked rag around your ankles so the ants won't climb up and eat your candy ass..." - Dale Earnhardt, 1997
-----
"The staggering layers of obscenity in your statement make it a work of art on so many levels." - J. Jystad, 2001 -
Nicholas Butler wrote:
Dan Neely wrote: the guy who ~4 years ago got a 9x speedup vs the stock app by hacking the binary Impressive OMG
Yeah. It was really nice for a while since we were getting 9x as many points/day which helped a lot in the bragging rights contest. IIRC he got about ~4-5x from algorithmic changes that were incorporated in the next official app (moving stuff into lookup tables and tweaking working sets to fit inside L1 caches) and ~2x from writing better hot loops to eliminate pipeline stalls and incorporating SSE/SSE2 instructions. The original was pure x87, his replacements interleaved both x87 and SSE instructions to keep both of their pipelines fully loaded. It was eye opening after years of reading that a good compiler is almost impossible to beat these days. Granted that a good C/C++ programmer could have done most of the improvements in his code and used compiler options to get basic SSE1/2 versions, but IIRC we've had 30-50% speedups in new apps after they were regarded as stable enough to start mucking around with asm. Overkill for most uses probably, but at supercomputer work loads even a few percent can help; especially when the total problem set is large enough that probabilistic algorithms are needed to prune the data massively early on. Even with those at present signals need to be an order of magnitude larger than the sensors thresholds to be detected (below that the odds decline roughly linearly), the previous algorithm's sure detection threshold was +2 magnitudes. This is a major challenge when the strongest signal expected from a known pulsar is right at the theoretician detection threshold and is near a massive noise spike.
3x12=36 2x12=24 1x12=12 0x12=18
Dan Neely wrote:
It was eye opening after years of reading that a good compiler is almost impossible to beat these days
yup. i beat it almost every time i try, and i'm a complete ASM novice. and the 'intrinsics' are nearly more trouble than they're worth. i was struggling this weekend to come up with x64 equivalents to some old
_asm {... MMX stuff...}blocks in our code, so i tried the intrinsics. but they ended up being like 4x slower than my hand-written ASM because things were constantly being shoved in and out of the MMX registers instead of being held registers for reuse. -
So, it's time to replace my main dev box. Request for advice &| experience please :) I built my current box for £1200 2.5 years ago. It's a dual quad-core Xeon and I've learnt just about all I can from it. It's been the best £1 per day I've ever spent, but hey! nothing lasts forever... So I've looked around a bit and AMD have released a new range "Magny-Cours" of 8-core and 12-core Opterons at sensible prices. They are basically two quad- or hex-core CPUs in one package for half the price. They have are NUMA ( albeit cache coherent ) as each half of each socket has it's own memory channel. And they all support 2-socket and 4-socket configurations. Very interesting! I'm sorely tempted to build a 48-core box just for fun, but I'm having difficulty convincing myself as my last client still only had up to 8-core boxes. On the other hand, I think that will change soon, and certainly over the next 2.5 years. BTW, I would be using this for writing concurrent business server software. What would you do with 48 cores? Cheers, Nick
Nicholas Butler wrote:
What would you do with 48 cores?
Fully describe the high power DE stray energy problem Recreate my college "Lana" (Lexical Analysis) and add AI Make some killer raytraces.... Increase my RF-LOS work even more.... Massive amounts of soft-body physics.... "wouldn't you love to get your hands on one of those gibsons...." :-D
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
modified on Monday, September 20, 2010 3:09 PM
-
Nicholas Butler wrote:
What would you do with 48 cores?
i'd wonder what was going to become of those two bags of apples.
Chris Losinger wrote:
i'd wonder what was going to become of those two bags of apples.
Apple cobbler!
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
-
From comments made elsewhere I don't think is using all zillion cores to compile on, he's writing apps that scale across a zillion cores.
3x12=36 2x12=24 1x12=12 0x12=18
Dan Neely wrote:
he's writing apps that scale across a zillion cores.
The only way to fly!! :-D :-D
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
-
Chris Losinger wrote:
i'd wonder what was going to become of those two bags of apples.
Apple cobbler!
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
pie would be acceptable, also.
-
probably a fan of the banana.
-
pie would be acceptable, also.
Chris Losinger wrote:
pie would be acceptable, also.
true... but since he is considering cobbling together his own computer... I figured a cobbler was slightly more accurate. :)
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
-
Chris Losinger wrote:
pie would be acceptable, also.
true... but since he is considering cobbling together his own computer... I figured a cobbler was slightly more accurate. :)
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
good point. cobbler it is!
-
Nicholas Butler wrote:
What would you do with 48 cores?
Fully describe the high power DE stray energy problem Recreate my college "Lana" (Lexical Analysis) and add AI Make some killer raytraces.... Increase my RF-LOS work even more.... Massive amounts of soft-body physics.... "wouldn't you love to get your hands on one of those gibsons...." :-D
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
modified on Monday, September 20, 2010 3:09 PM
El Corazon wrote:
Massive amounts of soft-body physics
your internet connection will have more of an effect on that than sheer processing power.
image processing toolkits | batch image processing
modified on Monday, September 20, 2010 4:31 PM
-
El Corazon wrote:
Massive amounts of soft-body physics
your internet connection will have more of an effect on that than sheer processing power.
image processing toolkits | batch image processing
modified on Monday, September 20, 2010 4:31 PM
Chris Losinger wrote:
your internet connection will have more of an effect on that then sheer processing power.
I wonder how many heads that will go over?? :laugh:
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
-
So, it's time to replace my main dev box. Request for advice &| experience please :) I built my current box for £1200 2.5 years ago. It's a dual quad-core Xeon and I've learnt just about all I can from it. It's been the best £1 per day I've ever spent, but hey! nothing lasts forever... So I've looked around a bit and AMD have released a new range "Magny-Cours" of 8-core and 12-core Opterons at sensible prices. They are basically two quad- or hex-core CPUs in one package for half the price. They have are NUMA ( albeit cache coherent ) as each half of each socket has it's own memory channel. And they all support 2-socket and 4-socket configurations. Very interesting! I'm sorely tempted to build a 48-core box just for fun, but I'm having difficulty convincing myself as my last client still only had up to 8-core boxes. On the other hand, I think that will change soon, and certainly over the next 2.5 years. BTW, I would be using this for writing concurrent business server software. What would you do with 48 cores? Cheers, Nick
48 AMD cores? Boiling water for tea in 12 seconds!
Agh! Reality! My Archnemesis![^]
| FoldWithUs! | sighist | WhoIncludes - Analyzing C++ include file hierarchy -
So, it's time to replace my main dev box. Request for advice &| experience please :) I built my current box for £1200 2.5 years ago. It's a dual quad-core Xeon and I've learnt just about all I can from it. It's been the best £1 per day I've ever spent, but hey! nothing lasts forever... So I've looked around a bit and AMD have released a new range "Magny-Cours" of 8-core and 12-core Opterons at sensible prices. They are basically two quad- or hex-core CPUs in one package for half the price. They have are NUMA ( albeit cache coherent ) as each half of each socket has it's own memory channel. And they all support 2-socket and 4-socket configurations. Very interesting! I'm sorely tempted to build a 48-core box just for fun, but I'm having difficulty convincing myself as my last client still only had up to 8-core boxes. On the other hand, I think that will change soon, and certainly over the next 2.5 years. BTW, I would be using this for writing concurrent business server software. What would you do with 48 cores? Cheers, Nick
All I can contribute is that I and many other developers have long ago switched to developing on a virtual workstation using something like VMWare Workstation or Virtual Box so any and all hardware decisions I make always revolve around optimal performance for VMWare workstation machines. If you or other readers have not considered this you should. That being said my secondary consideration these days seems to be power consumption since almost any decent hardware is powerful enough for what I do so there is no reason not to consider power consumption pretty high on the list and the electric bill is not an insignificant business expense.
“If you want to build a ship, don't drum up people together to collect wood and don't assign them tasks and work, but rather teach them to long for the endless immensity of the sea” - Antoine de Saint-Exupery
-
So, it's time to replace my main dev box. Request for advice &| experience please :) I built my current box for £1200 2.5 years ago. It's a dual quad-core Xeon and I've learnt just about all I can from it. It's been the best £1 per day I've ever spent, but hey! nothing lasts forever... So I've looked around a bit and AMD have released a new range "Magny-Cours" of 8-core and 12-core Opterons at sensible prices. They are basically two quad- or hex-core CPUs in one package for half the price. They have are NUMA ( albeit cache coherent ) as each half of each socket has it's own memory channel. And they all support 2-socket and 4-socket configurations. Very interesting! I'm sorely tempted to build a 48-core box just for fun, but I'm having difficulty convincing myself as my last client still only had up to 8-core boxes. On the other hand, I think that will change soon, and certainly over the next 2.5 years. BTW, I would be using this for writing concurrent business server software. What would you do with 48 cores? Cheers, Nick
Fold faster.
-
Nicholas Butler wrote:
What would you do with 48 cores?
Fully describe the high power DE stray energy problem Recreate my college "Lana" (Lexical Analysis) and add AI Make some killer raytraces.... Increase my RF-LOS work even more.... Massive amounts of soft-body physics.... "wouldn't you love to get your hands on one of those gibsons...." :-D
_________________________ John Andrew Holmes "It is well to remember that the entire universe, with one trifling exception, is composed of others." Shhhhh.... I am not really here. I am a figment of your imagination.... I am still in my cave so this must be an illusion....
modified on Monday, September 20, 2010 3:09 PM
El Corazon wrote:
"wouldn't you love to get your hands on one of those gibsons...."
Actually, there were some 48-core supercomputers in 1995 ;) Thanks for your suggestions, but wouldn't GPU(s) be better for those applications? How would a 48-core computer perform against a couple of the new Tesla C2050 cards, for example? I am more interested in C# LOB apps that use concurrency. Which, as far as I can tell at the moment, is mostly server software. I was hoping that someone was actually using the cores available on client machines. Nick
-
48 AMD cores? Boiling water for tea in 12 seconds!
Agh! Reality! My Archnemesis![^]
| FoldWithUs! | sighist | WhoIncludes - Analyzing C++ include file hierarchyMy current Intel box can already do that :) Nick
-
All I can contribute is that I and many other developers have long ago switched to developing on a virtual workstation using something like VMWare Workstation or Virtual Box so any and all hardware decisions I make always revolve around optimal performance for VMWare workstation machines. If you or other readers have not considered this you should. That being said my secondary consideration these days seems to be power consumption since almost any decent hardware is powerful enough for what I do so there is no reason not to consider power consumption pretty high on the list and the electric bill is not an insignificant business expense.
“If you want to build a ship, don't drum up people together to collect wood and don't assign them tasks and work, but rather teach them to long for the endless immensity of the sea” - Antoine de Saint-Exupery
Interesting John, thanks :) I presume all modern CPUs ( especially server ones ) fully support virtualization, no? I do use VMWare Workstation, but not for my main dev box as it is limited to 4-cores [edit: just checked - it now supports up to 8. I'm sure it was 4 last time I looked ] My last client was using 8-core boxes, and part of my job is to be ahead of the curve a bit. So it's not really an option for me, although I do agree they are a good choice for normal development. Nick
-
My current Intel box can already do that :) Nick
:-D
Agh! Reality! My Archnemesis![^]
| FoldWithUs! | sighist | WhoIncludes - Analyzing C++ include file hierarchy