I picked up the book.
I opened it to the first page.
He said he hadn't read that page.
I hadn't either.
He read it.
Then I read it too.
We both read the page.
Then we talked about it.
The page had information.
We talked about the information.
I'm sure Sun has good hardware, but my personal experience at one shop was pretty weak and it was due to having non-ECC memory in these servers (this wasn't a customer choice, it was what was offered from Sun, we had some 4500's and some other similar number). The things crashed randomly with memory problems once a month. They kept replacing memory but problem never went away.
We had as400's, hp9000's, etc., those sun boxes were the only ones that ever went down.
The T2 chips do compete with x86 in the single chip arena, but with Intel's volume and the rate of advancement, that won't last much longer. On the high end (multi-proc systems), T2 performs very poorly against Power (which is designed well for multi-proc systems).
Intel is going to win this war, IBM will hold out the longest, not much hope for anyone else due to the expense of the cpu business and trying to compete with Intel's volume.
There is no way he didn't know and walk through the pro's and con's of this acquisition with his people well in advance. All of those guys know far more than anything reported in the media, and because they are sitting in the same position they can pretty quickly do the analysis to figure out who might or might not be in a position to acquire a company like Sun.
The writing is on the wall, with Intel's volume it is going to be (already is) very difficult to compete in the processor market. I would be very surprised if Sun cpu's lasted very long. Maybe Oracle will keep building systems, but as Intel continues to improve, I'm guessing it will be with Intel cpu's.
The other processors perform equally well with workloads that require single thread performance because they didn't sacrifice that performance. The point is that Sun targeted a specific workload and is limited to those types of workloads. The other vendors did not and can play in both both arenas.
Benchmarks
I was specifically talking about your claim that 1 Power6 core was only 43% faster than 1 T2 core. Let me illustrate:
T2 operates at 1.4ghz
has 1 floating point ALU per core shared by all threads
has 2 integer ALU's per core shared by all threads
The Power6 operates at 4.7ghz
has 2 floating point ALU's per core shared by all threads
has 2 integer ALU's per core shared by all threads
I think you can see why I am skeptical about that 43% claim, that's why I asked you for number to back it up. The analysis I read says Power6 cores are about 4x a T2 core, T2 has 4x as many cores so they are close, as long as the workload support it. When the workload shifts to requiring more single threaded performance the T2 is a poor choice.
Yes it's a niche. But it's a HUGE niche. Enterprise server consolidation, webserving, java application servers, pretty much any multiuser application.
Not any multi-user application. That's why it's a niche. It was designed for web serving type workloads with many many threads and very small amount of processing per thread at any time. That's why the cores share so much of the logic units, it's a trade-off (like everything).
According to this IBM blog post [ibm.com], a Power6 core offers only 43% more performance than a UltraSparc T2 core.
Do you have specific benchmarks that back that up? I have not seen any numbers (spec.org?) anywhere showing T2 outperforming any other processors for any compute intensive benchmark. I looked into it when purchasing a system to perform compute intensive simulations (small system) and at 1.2ghz with a round-robin style of processing quickly eliminated any performance gains I thought would be there (I was also looking at a network of PS3's, Intel and GPU's).
IBM recently announced their Nehelem xeon processor which seems similar to the T2 chips
You mean Intel's Nehalem. Do you mean similar to T2 in that it has multiple cores? Well yes, ever since IBM first introduced a multi-core chip everyone is pretty much going down the same path. The difference is that none of the other chips backed off their speed down to 1.2ghz.
Agreed that the switch can be costly and I could certainly see them producing T2's for as long as the install base supports it. But I would be surprised to see new development.
Have you been in a coma or something? This is where all the major chip vendors are going.
The difference between T2 and the other chip vendors offering multi-core/multi-thread is that the other vendors are not sacrificing single thread performance. They are making use of the additional space due to shrink to increase core's while retaining single thread performance because many workloads will always require the single thread performance.
Sun designed these processors for a specific niche, it's not a secret, they were pretty clear about it.
They want the T2 processors, which are selling bangup to folks who have legacy SPARC source code commitments
I doubt they want the T2. It's a niche product (many threads, small compute per thread, very poor single thread) that is fighting against intel with massive economies of scale from the low end, and it isn't competetive with Power in the traditional workload environment.
Kill the hardware that's common to both the iSeries and pSeries (and I'm not certain how much of the hardware also is shared with the zSeries)? Doubtful.
I'm pretty sure they added instructions into POWER6 that get's them closer to the stuff the mainframe has (transaction rollback etc.) so it's not there yet, maybe by POWER7
As someone who has extensive hands-on use of Oracle eBusiness, I can say it's a steaming turd with some authority
Curious if you have specifics. I've worked both ends of the spectrum (completely integrated and best of breed for all apps with interfaces), I definately prefer the integrated approach. What you gain with best of breed you quickly lose in complexity of interfaces, loss of functionality because data from one app's model can't be easily represented in the other app's model and loss of a comprehensive view across all aspects of the business.
Mine maxed out at 32k. An EE I knew figured out how to get 64k, he piggy-backed 2 sets of 32k mem chips on top of each other in the sockets, one of them had 1 leg bent up with a wire connecting it to some spot on the board. Internally I could set a bit to flip between seeing the upper 32k as the factory installed ROM, or as usable RAM.
If you read the statements on their web page you will see that they aren't missing any of that and aren't over-hyping. They themselves state that we are 50 years away from just the hardware capabilities to match a brain let alone the understanding. In addition they point out this this is really just a tool for further research in understanding of non-turing style computing, which is absolutely right. There is more than one goal surrounding all of this neural stuff: 1) understanding how the brain computes, 2) Understanding the advantages/disadvatages of neural network style computing, and others I'm sure. Although the IBM brain simulator will help understand how the brain operates (to some degree), this project is really more towards the understanding of non-turing computing.
It's true that some tests are closer than others, that's why I said generally, but look at all the tests (including cintrate and cfprate), I see many where power6 is about double. As for mutiple cores, set # cores equal "2" and look through all tests, I see Power6 showing scores of around 60/50 with other procs around 30.
But isn't there some provision in patent law excluding things that are so obvious that if an "average person" can come up with it without specialized knowledge, it's not covered?
On these trivial patents, the part that is non-obvious is that you CAN actually get a patent for them, anyone that figures that out is rewarded with a patent. See what they did there, actually pretty clever!
They are going to become competitors whether they like it or not, and both Intel and nVidia already know this and both are actively working to be in a position of strength.
Computing is rapidly going many core, Intel is moving that direction from a position of good single thread performance but poor multi-thread performance (meaning something like the 16,000 threads that a GTX280 executes), GPU mfg's are coming from the opposite direction, good multi-thread performance but poor single thread performance.
Both types of computing are needed and more and more we have the technology to combine both on a single chip. If either Intel or nVidia sat back and figured business would continue as it is today, then they would end up getting left in the dust.
Except it wasn't an object as I understood things it was a table. Data within could be accessed through row operations or SQL. Pretty cool, I thought. The best of SQL with the best of the old hierarchical systems.
It was an object based system. Everything was an object with well defined methods that could be called.
It contained an integrated database, every file was automatically relational (not hierarchical, that's the mainframe).
One of their tech guys was on the CUDA programming forum, so I assume they used CUDA for NVIDIA. Not sure why the article assumes that the "proprietary process" assumes the interface language is proprietary. I assumed it meant their method/algorithm was proprietary (given the fact that they patented it, they probably want to make it seem unique).
Slashdot Mirror
I picked up the book.
I opened it to the first page.
He said he hadn't read that page.
I hadn't either.
He read it.
Then I read it too.
We both read the page.
Then we talked about it.
The page had information.
We talked about the information.
I'm sure Sun has good hardware, but my personal experience at one shop was pretty weak and it was due to having non-ECC memory in these servers (this wasn't a customer choice, it was what was offered from Sun, we had some 4500's and some other similar number). The things crashed randomly with memory problems once a month. They kept replacing memory but problem never went away. We had as400's, hp9000's, etc., those sun boxes were the only ones that ever went down.
The T2 chips do compete with x86 in the single chip arena, but with Intel's volume and the rate of advancement, that won't last much longer. On the high end (multi-proc systems), T2 performs very poorly against Power (which is designed well for multi-proc systems).
Intel is going to win this war, IBM will hold out the longest, not much hope for anyone else due to the expense of the cpu business and trying to compete with Intel's volume.
There is no way he didn't know and walk through the pro's and con's of this acquisition with his people well in advance. All of those guys know far more than anything reported in the media, and because they are sitting in the same position they can pretty quickly do the analysis to figure out who might or might not be in a position to acquire a company like Sun.
The writing is on the wall, with Intel's volume it is going to be (already is) very difficult to compete in the processor market. I would be very surprised if Sun cpu's lasted very long. Maybe Oracle will keep building systems, but as Intel continues to improve, I'm guessing it will be with Intel cpu's.
What is this "article" thing you refer to?
SUVs are stupid because they are inferior to the alternatives, and SUV owners are stupid because they were convinced as much by advertising
You sound angry about this topic. What's got you down, can't afford an SUV?
Yeah... Too bad the interweb never caught on.
The other processors perform equally well with workloads that require single thread performance because they didn't sacrifice that performance. The point is that Sun targeted a specific workload and is limited to those types of workloads. The other vendors did not and can play in both both arenas.
Benchmarks
I was specifically talking about your claim that 1 Power6 core was only 43% faster than 1 T2 core. Let me illustrate:
T2 operates at 1.4ghz
has 1 floating point ALU per core shared by all threads
has 2 integer ALU's per core shared by all threads
The Power6 operates at 4.7ghz
has 2 floating point ALU's per core shared by all threads
has 2 integer ALU's per core shared by all threads
I think you can see why I am skeptical about that 43% claim, that's why I asked you for number to back it up. The analysis I read says Power6 cores are about 4x a T2 core, T2 has 4x as many cores so they are close, as long as the workload support it. When the workload shifts to requiring more single threaded performance the T2 is a poor choice.
Yes it's a niche. But it's a HUGE niche. Enterprise server consolidation, webserving, java application servers, pretty much any multiuser application.
Not any multi-user application. That's why it's a niche. It was designed for web serving type workloads with many many threads and very small amount of processing per thread at any time. That's why the cores share so much of the logic units, it's a trade-off (like everything).
According to this IBM blog post [ibm.com], a Power6 core offers only 43% more performance than a UltraSparc T2 core.
Do you have specific benchmarks that back that up? I have not seen any numbers (spec.org?) anywhere showing T2 outperforming any other processors for any compute intensive benchmark. I looked into it when purchasing a system to perform compute intensive simulations (small system) and at 1.2ghz with a round-robin style of processing quickly eliminated any performance gains I thought would be there (I was also looking at a network of PS3's, Intel and GPU's).
IBM recently announced their Nehelem xeon processor which seems similar to the T2 chips
You mean Intel's Nehalem. Do you mean similar to T2 in that it has multiple cores? Well yes, ever since IBM first introduced a multi-core chip everyone is pretty much going down the same path. The difference is that none of the other chips backed off their speed down to 1.2ghz.
Agreed that the switch can be costly and I could certainly see them producing T2's for as long as the install base supports it. But I would be surprised to see new development.
Have you been in a coma or something? This is where all the major chip vendors are going.
The difference between T2 and the other chip vendors offering multi-core/multi-thread is that the other vendors are not sacrificing single thread performance. They are making use of the additional space due to shrink to increase core's while retaining single thread performance because many workloads will always require the single thread performance.
Sun designed these processors for a specific niche, it's not a secret, they were pretty clear about it.
They want the T2 processors, which are selling bangup to folks who have legacy SPARC source code commitments
I doubt they want the T2. It's a niche product (many threads, small compute per thread, very poor single thread) that is fighting against intel with massive economies of scale from the low end, and it isn't competetive with Power in the traditional workload environment.
Kill the hardware that's common to both the iSeries and pSeries (and I'm not certain how much of the hardware also is shared with the zSeries)? Doubtful.
I'm pretty sure they added instructions into POWER6 that get's them closer to the stuff the mainframe has (transaction rollback etc.) so it's not there yet, maybe by POWER7
As someone who has extensive hands-on use of Oracle eBusiness, I can say it's a steaming turd with some authority
Curious if you have specifics. I've worked both ends of the spectrum (completely integrated and best of breed for all apps with interfaces), I definately prefer the integrated approach. What you gain with best of breed you quickly lose in complexity of interfaces, loss of functionality because data from one app's model can't be easily represented in the other app's model and loss of a comprehensive view across all aspects of the business.
The parent post was referring to Oracle applications, the ERP software they sell that competes with SAP, not the database.
Setting aside the brainless rumors of Sun being bought
Why would you call them brainless, do you honestly think there is nothing to them?
Mine maxed out at 32k. An EE I knew figured out how to get 64k, he piggy-backed 2 sets of 32k mem chips on top of each other in the sockets, one of them had 1 leg bent up with a wire connecting it to some spot on the board. Internally I could set a bit to flip between seeing the upper 32k as the factory installed ROM, or as usable RAM.
If you read the statements on their web page you will see that they aren't missing any of that and aren't over-hyping. They themselves state that we are 50 years away from just the hardware capabilities to match a brain let alone the understanding. In addition they point out this this is really just a tool for further research in understanding of non-turing style computing, which is absolutely right. There is more than one goal surrounding all of this neural stuff: 1) understanding how the brain computes, 2) Understanding the advantages/disadvatages of neural network style computing, and others I'm sure. Although the IBM brain simulator will help understand how the brain operates (to some degree), this project is really more towards the understanding of non-turing computing.
My feeling is that "Web 2.0" was a giant waste of time
I completely agree, the quicker we get to Web 3.0 the better.
It's true that some tests are closer than others, that's why I said generally, but look at all the tests (including cintrate and cfprate), I see many where power6 is about double. As for mutiple cores, set # cores equal "2" and look through all tests, I see Power6 showing scores of around 60/50 with other procs around 30.
You may want to go back to that sight and try again. You're looking at 1 Power6 core against 2, 4 and 8 cores on the procs getting similar scores.
You need to do 2 things, first set the # Cores "equal" to 1 and look at the list so you are comparing apples to apples.
Now set the # Cores "equal" to 2 and do the same thing, just so you can see how the procs scale with multiple cores.
I generally see the Power6 score 50% to 100% higher than others.
On these trivial patents, the part that is non-obvious is that you CAN actually get a patent for them, anyone that figures that out is rewarded with a patent. See what they did there, actually pretty clever!
They are going to become competitors whether they like it or not, and both Intel and nVidia already know this and both are actively working to be in a position of strength.
Computing is rapidly going many core, Intel is moving that direction from a position of good single thread performance but poor multi-thread performance (meaning something like the 16,000 threads that a GTX280 executes), GPU mfg's are coming from the opposite direction, good multi-thread performance but poor single thread performance.
Both types of computing are needed and more and more we have the technology to combine both on a single chip. If either Intel or nVidia sat back and figured business would continue as it is today, then they would end up getting left in the dust.
It was an object based system. Everything was an object with well defined methods that could be called.
It contained an integrated database, every file was automatically relational (not hierarchical, that's the mainframe).
One of their tech guys was on the CUDA programming forum, so I assume they used CUDA for NVIDIA. Not sure why the article assumes that the "proprietary process" assumes the interface language is proprietary. I assumed it meant their method/algorithm was proprietary (given the fact that they patented it, they probably want to make it seem unique).