Assume identical disks because that's what happens when you have systems doing identical tasks - thus disks can be ignored in a comparison. I thought that would be obvious so I just stated that as "disks ignored". Is that enough to satisfy that nitpick?
No, because the price ratios are completely different once you add in disks. X over Y is not equal to X+N over Y+N if we're dealing with positive numbers.
To bring in a car analogy (I know, everyone's favorite thing, right?), this is like subtracting out the cost of non-engine parts and claiming an electric car is 10 times more expensive than a similar non-electric car. Well, such price comparisons are disingenuous, because no one ever buys just an engine; they need the entire rest of the car!
You can't ignore disks and claim order of magnitudes of price difference, just like you can't ignore costs of power and cooling and software. You won't be running without disks, electricity, or software, I trust?
Please do. Otherwise, you might as well be describing invisible pink unicorns. Your statements about pricing are wildly divergent from the reality I am seeing.
AMD doesn't exist outside of the low end server space. Precisely what sector were you talking about, where you can price out an AMD machine on the web? The machines I was talking about scale out to greater than 4 sockets, and possess features aimed at the highest end of the market such as RAS and multi-year service/software support contracts. AMD's latest Interlagos models don't even support more than 4 sockets without custom interconnects and glue logic, and if you're going for more sockets than that, you're in Top500 supercomputer territory, where cost is not a consideration anyway. If it were, AMD would be doing better than it is, as that is the only reason to buy an AMD server nowadays.
The cost difference is not nearly an order of magnitude when you take into account all the other components of the server besides CPUs you have to pay for anyway.
In desktop systems and low-end servers, the processor is often the most expensive component (or second most expensive, behind the video card), and can easily be 25-30 percent of the total system cost. But as the server becomes bigger and more expensive, with hundreds of gigabytes of RAM and terabytes of storage, the processor can become a smaller part of the total cost. For example, Dell's PowerEdge R910, a 4-socket Xeon server, lets you spend up to about $22,000 on processors, if you get four of the most expensive parts offered (the Xeon E7-4870). That's a lot, but it's nothing compared to the $185,000 that equipping the machine with 2TB RAM would cost.
If you calculate the price of a Dell R710 with the Xeon E5649 and compare it with a Dell R715 with the Opteron 6276 with similar specs, you end up more or less the same acquisition cost. However, the E5649 is an 80W TDP and should thus consume a bit less power. That is why we argued that the Opteron 6276 should at least offer a price/performance bonus and perform like an X5650. The X5650 is roughly $220 more expensive, so you end up with the dual socket Xeon system costing about $440 more. On a fully speced server, that is about a 10% price difference.
When we look at the higher end OLTP and the non low end ERP market, the cost of buying server hardware is lost in the noise. The Westmere-EX with its higher thread count and performance will be the top choice in that case: higher thread count, better RAS, and a higher number of DIMM slots.
AMD also lost the low end OLAP market: the Xeon offers a (far) superior performance/watt ratio on mySQL. In the midrange and high end OLAP market, the software costs of for example SQL Server increase the importance of performance and performance/watt and make server hardware costs a minor issue. Especially the "performance first" OLAP market will be dominated by the Xeon, which can offer up to 3.06GHz SKUs without increasing the TDP.
The competitive picture has not improved for AMD since then, with the release of the Intel Xeon E5's that are Sandy-Bridge based 6 months ago, as opposed to the Westmere-based Xeons that were benchmarked in the above reviews.
And since server/workstation hardware costs are dwarfed by server/workstation software license costs, which tend to be per core, that means that the total cost of ownership of fewer Intel cores is actually less than the TCO of more, slower AMD cores that deliver the same performance. It turns out that due to Amdahl's Law, single-threaded performance matters even more in the many-core space.
What? AMD is hovering around 6% market share in the server market, with Intel making up the other 94%! Intel has been dominating there ever since they were the first to introduce a quad-core server processor (Clovertown,) back in November 2006. Intel's average selling price (ASP) is also 33% higher. AMD's chips just do not perform anywhere near as well as Intel's on most server tasks. AMD really needs Bulldozer to improve their competitive position if they want to remain a player in the server processor market.
My god, there are too many errors in that article.
Eg, the first one that jumped out at me: "Both Athlon and Hammer have a 128KB, two-way set associative, on-die L1 cache."
The original athlon had a 2-way associative cache, but it wasn't on-die, or 128kB. Later chips (also "K7" and "Athlon", but with a redesigned core) were 16-way, 128kB, on die. Which are they comparing against? Do they even know?
Um, what about that statement was in error? L1 cache has been on-die since the 486. And AMD has never made 16-way set associative L1 cache. I have to question your knowledge of this area if you cannot make a distinction between L1 and L2 cache. And all versions of the Athlon had 64kB of L1 instruction cache and 64kB of L1 data cache, adding up to 128kB.
May I suggest knowing what you're talking about before speaking again?
Why would the threads bump into each other? SMT is supposed to utilize execution units that are otherwise idle, and no processor fully utilizes their execution units all the time, not even so-called brainiac designs like K8 and Core. That requires perfect ordering of instructions and their dependencies and ensuring that you always have (say) 3 integer instructions, 2 floating point instructions, a load, and a store, in every clock cycle. An IPC of 2 that you declare as "good", usually means a utilization of maybe a quarter to a half of the execution units a modern x86 CPU has available.
It's just the way it is. Until compilers manage to write code with no dependencies and sprinkle the correct percentage of each instruction type in the in-flight window to keep every execution unit filled on every clock cycle, there will be free slots in the execution units. There is no way around it. You might as well take advantage of the free slots.
There was a lot of good discussion about SMT in that thread. You should read it.
By far the dominant CPU activity in most code these days is waiting for load/stores to be fulfilled. And processors like Sun's Niagara address this by allowing up to 4 threads to be scheduled in a single core at the same time, so that it can batch memory requests from all four threads at once. The SPARC cores in Niagara though, really have only one execution pipeline, so only one thread is actually executing per clock cycle. But the throughput of the CPU as a whole is still increased, since the core will switch the executing thread out when one hits a stall in memory access. Simultaneous multithreading is even better in that it actually lets instructions from more than one thread execute within the same clock cycle, utilizing wide CPU architectures better.
There are implementation caveats, sure. The Pentium 4 had certain microarchitectural features (load address speculation and replay) that caused some threads to greatly decrease in efficiency. But this is not a problem inherent to multithreading, and indeed processors like IBM's Power6 and Intel's Itanium2 have decided to implement it, to great success. Pretty much every high-performance CPU company these days makes a chip that implements core-level multithreading in one form or another. AMD is the only one that doesn't.
And the reason why Alpha does not currently have multithreading, was that EV8 (what would have been the 21464) was cancelled when Compaq/HP's Alpha Design Group was transferred to Intel. EV8 (nicknamed Arana, from the Spanish for "spider") would have implemented 4-way simultaneous multithreading in a high-performance, short pipeline, wide architecture.
Anyway, sorry if I come off as high-handed, but I couldn't let a ridiculously naive statement like "The core2 and AMD64 don't have enough bubbles to warrant it." stand.
But for Intel, having multicore in mind doesn't mean quite the same thing that it means for Sun or IBM. Specifically, "multicore" doesn't mean "throw out out-of-order execution and scale back single-threaded performance in favor of a massively parallel architecture that can run a torrent of simultaneous threads." Such an aggressive, forward-looking approach is embodied in designs like STI's Cell and Sun's Ultrasparc T1.
Also, contrary to the first poster's statements, Intel's Itanium architecture (IA-64) uses an in-order implementation, relying on the compiler to extract instruction-level parallelism. So there is a definite trend in the industry to forgo the hardware complexity of out-of-order execution in favor of software methods of extracting parallelism.
It was an FX-60 (overclocked to 2.8 GHz), which is pretty much identical to the Athlon64 X2 series with the sole exception of having unlocked multipliers to support overclocking. Hexus did a review of the same machines over at http://www.hexus.net/content/item.php?item=4843.
And when someone calls me Orientalist, then maybe I might have cause to be offended. "Oriental", though, has absolutely no negative connotations, and ascribing it a new one seems to me to be yet another attempt by western thought to fall into the same errors it ironically claims to try to avoid.
There are no G5 notebook computers. The G5 chips burn far too much power to be used in a notebook, unless it's one of those desktop replacement notebooks that last maybe an hour on battery. The inability of IBM to provide a low-power G5 was a major factor in Apple's switch to Intel chips.
manga is not a style, as the style of comics in japan is far too varied to be a genre.
Of course it can be a genre. This is like claiming that since there are so many different types of animals, there can be no classification "animal". It's a logical fallacy. There are still commonalities, and manga has distinctive stylistic qualities which distinguish it from other forms of sequential art.
you even say that mimicking the "style" of japanese manga means an american comic can be called "manga", therefore you ADMIT that manga does mean japanese...
No. If someone says that mimicking the impressionist style of painting means an American painting can be called "impressionist", it does not mean he's saying "impressionist" is French. It's recognizing that "impressionist" can be orthogonal to "insert nationality here".
I didn't use to have a strong opinion on this issue, but you're definitely making a convincing argument for the other side.
The personal gaming/high-performance market is going the way of the dinosaur. Most people I know are more interested in building small-form-factor/silent systems.
What I don't understand is this: when you just listen, what is the difference between you and a wall?
You're a fellow human being.;)
But seriously, there doesn't have to be a difference. Sometimes just talking about a problem (or writing about it in a journal) makes someone feel better. It's akin to the phenomenon where naming a problem makes someone feel they have a better handle on it.
Slashdot Mirror
Assume identical disks because that's what happens when you have systems doing identical tasks - thus disks can be ignored in a comparison. I thought that would be obvious so I just stated that as "disks ignored". Is that enough to satisfy that nitpick?
No, because the price ratios are completely different once you add in disks. X over Y is not equal to X+N over Y+N if we're dealing with positive numbers.
To bring in a car analogy (I know, everyone's favorite thing, right?), this is like subtracting out the cost of non-engine parts and claiming an electric car is 10 times more expensive than a similar non-electric car. Well, such price comparisons are disingenuous, because no one ever buys just an engine; they need the entire rest of the car!
You can't ignore disks and claim order of magnitudes of price difference, just like you can't ignore costs of power and cooling and software. You won't be running without disks, electricity, or software, I trust?
Please do. Otherwise, you might as well be describing invisible pink unicorns. Your statements about pricing are wildly divergent from the reality I am seeing.
AMD doesn't exist outside of the low end server space. Precisely what sector were you talking about, where you can price out an AMD machine on the web? The machines I was talking about scale out to greater than 4 sockets, and possess features aimed at the highest end of the market such as RAS and multi-year service/software support contracts. AMD's latest Interlagos models don't even support more than 4 sockets without custom interconnects and glue logic, and if you're going for more sockets than that, you're in Top500 supercomputer territory, where cost is not a consideration anyway. If it were, AMD would be doing better than it is, as that is the only reason to buy an AMD server nowadays.
Top 500 stats by processor family:
x86-64 - 435
Intel Xeon - 372
AMD Opteron - 62
If you look at the historical graphs, AMD has been falling in share ever since 2009.
The cost difference is not nearly an order of magnitude when you take into account all the other components of the server besides CPUs you have to pay for anyway.
http://arstechnica.com/business/2011/11/bulldozer-server-benchmarks-are-here-and-theyre-a-catastrophe/
In desktop systems and low-end servers, the processor is often the most expensive component (or second most expensive, behind the video card), and can easily be 25-30 percent of the total system cost. But as the server becomes bigger and more expensive, with hundreds of gigabytes of RAM and terabytes of storage, the processor can become a smaller part of the total cost. For example, Dell's PowerEdge R910, a 4-socket Xeon server, lets you spend up to about $22,000 on processors, if you get four of the most expensive parts offered (the Xeon E7-4870). That's a lot, but it's nothing compared to the $185,000 that equipping the machine with 2TB RAM would cost.
Not to mention that Intel wins on performance/watt and performance/core and thus in total cost of ownership when power consumption and software licensing costs are taken into account. http://www.anandtech.com/show/5279/the-opteron-6276-a-closer-look/12
If you calculate the price of a Dell R710 with the Xeon E5649 and compare it with a Dell R715 with the Opteron 6276 with similar specs, you end up more or less the same acquisition cost. However, the E5649 is an 80W TDP and should thus consume a bit less power. That is why we argued that the Opteron 6276 should at least offer a price/performance bonus and perform like an X5650. The X5650 is roughly $220 more expensive, so you end up with the dual socket Xeon system costing about $440 more. On a fully speced server, that is about a 10% price difference.
When we look at the higher end OLTP and the non low end ERP market, the cost of buying server hardware is lost in the noise. The Westmere-EX with its higher thread count and performance will be the top choice in that case: higher thread count, better RAS, and a higher number of DIMM slots.
AMD also lost the low end OLAP market: the Xeon offers a (far) superior performance/watt ratio on mySQL. In the midrange and high end OLAP market, the software costs of for example SQL Server increase the importance of performance and performance/watt and make server hardware costs a minor issue. Especially the "performance first" OLAP market will be dominated by the Xeon, which can offer up to 3.06GHz SKUs without increasing the TDP.
The competitive picture has not improved for AMD since then, with the release of the Intel Xeon E5's that are Sandy-Bridge based 6 months ago, as opposed to the Westmere-based Xeons that were benchmarked in the above reviews.
This hasn't been true for about 4 years. Intel is now so crushingly ahead of AMD in the server space performance-wise and energy-efficiency-wise that they hold 94.5% of the server/workstation market. http://www.tomshardware.com/news/amd-intel-cpu-processor,15041.html
And since server/workstation hardware costs are dwarfed by server/workstation software license costs, which tend to be per core, that means that the total cost of ownership of fewer Intel cores is actually less than the TCO of more, slower AMD cores that deliver the same performance. It turns out that due to Amdahl's Law, single-threaded performance matters even more in the many-core space.
I'd say less than 10%. The vast majority of games and encoders in Windows are compiled with Microsoft's Visual C.
It makes a difference of 1-2% on x264. Nothing earth-shattering. http://forum.doom9.org/showthread.php?p=1530754#post1530754
Er, no. Gainestown (the 2-socket version of Nehalem) was released in March 2009, with system availability by the summer.
What? AMD is hovering around 6% market share in the server market, with Intel making up the other 94%! Intel has been dominating there ever since they were the first to introduce a quad-core server processor (Clovertown,) back in November 2006. Intel's average selling price (ASP) is also 33% higher. AMD's chips just do not perform anywhere near as well as Intel's on most server tasks. AMD really needs Bulldozer to improve their competitive position if they want to remain a player in the server processor market.
Psst... http://shop.amd.com/US/Pages/ShopHome.aspx . You're welcome. ;)
You can get 64-core Intel servers: http://www.sap.com/solutions/benchmark/sd2tier.epx
NEC Express5800 Model A1080a-E, 8 Processors / 64 Cores / 128 Threads, Intel Xeon Processor X7560, 2.26 Ghz, 64 KB L1 cache and 256 KB L2 cache per core, 24 MB L3 cache per processor
18185 users, 99450 SAPS
HP ProLiant DL980 G7, 8 Processors / 64 Cores / 128 Threads, Intel Xeon Processor X7560, 2.26 Ghz, 64 KB L1 cache and 256 KB L2 cache per core, 24 MB L3 cache per processor
18180 users, 99320 SAPS
AMD's best result is:
HP ProLiant BL685C G7, 4 Processors / 48 Cores / 48 Threads, AMD Opteron Processor 6174, 2.2 Ghz, 128 KB L1 cache and 512 KB L2 cache per core, 6 MB L3 cache per 6 cores
8675 users, 47420 SAPS
Even restricting Intel to 4 processors and 32 cores, it soundly beats AMD's best result in SAP-SD by 20%!
HP ProLiant DL580 G7, 4 Processors / 32 Cores / 64 Threads, Intel Xeon Processor X7560,, 2.26 Ghz, 64 KB L1 cache and 256 KB L2 cache per core, 24 MB L3 cache per processor
10445 users, 57020 SAPS
I think ShapeGSX says it best in the following Arstechnica thread: http://episteme.arstechnica.com/eve/forums/a/tpc/f /174096756/m/823008258731/p/3
It's just the way it is. Until compilers manage to write code with no dependencies and sprinkle the correct percentage of each instruction type in the in-flight window to keep every execution unit filled on every clock cycle, there will be free slots in the execution units. There is no way around it. You might as well take advantage of the free slots.
There was a lot of good discussion about SMT in that thread. You should read it.
By far the dominant CPU activity in most code these days is waiting for load/stores to be fulfilled. And processors like Sun's Niagara address this by allowing up to 4 threads to be scheduled in a single core at the same time, so that it can batch memory requests from all four threads at once. The SPARC cores in Niagara though, really have only one execution pipeline, so only one thread is actually executing per clock cycle. But the throughput of the CPU as a whole is still increased, since the core will switch the executing thread out when one hits a stall in memory access. Simultaneous multithreading is even better in that it actually lets instructions from more than one thread execute within the same clock cycle, utilizing wide CPU architectures better.
There are implementation caveats, sure. The Pentium 4 had certain microarchitectural features (load address speculation and replay) that caused some threads to greatly decrease in efficiency. But this is not a problem inherent to multithreading, and indeed processors like IBM's Power6 and Intel's Itanium2 have decided to implement it, to great success. Pretty much every high-performance CPU company these days makes a chip that implements core-level multithreading in one form or another. AMD is the only one that doesn't.
And the reason why Alpha does not currently have multithreading, was that EV8 (what would have been the 21464) was cancelled when Compaq/HP's Alpha Design Group was transferred to Intel. EV8 (nicknamed Arana, from the Spanish for "spider") would have implemented 4-way simultaneous multithreading in a high-performance, short pipeline, wide architecture.
Anyway, sorry if I come off as high-handed, but I couldn't let a ridiculously naive statement like "The core2 and AMD64 don't have enough bubbles to warrant it." stand.
It's essentially diesel fuel, but colored a nice pink.
But for Intel, having multicore in mind doesn't mean quite the same thing that it means for Sun or IBM. Specifically, "multicore" doesn't mean "throw out out-of-order execution and scale back single-threaded performance in favor of a massively parallel architecture that can run a torrent of simultaneous threads." Such an aggressive, forward-looking approach is embodied in designs like STI's Cell and Sun's Ultrasparc T1.
Also, contrary to the first poster's statements, Intel's Itanium architecture (IA-64) uses an in-order implementation, relying on the compiler to extract instruction-level parallelism. So there is a definite trend in the industry to forgo the hardware complexity of out-of-order execution in favor of software methods of extracting parallelism.
It was an FX-60 (overclocked to 2.8 GHz), which is pretty much identical to the Athlon64 X2 series with the sole exception of having unlocked multipliers to support overclocking. Hexus did a review of the same machines over at http://www.hexus.net/content/item.php?item=4843.
And when someone calls me Orientalist, then maybe I might have cause to be offended. "Oriental", though, has absolutely no negative connotations, and ascribing it a new one seems to me to be yet another attempt by western thought to fall into the same errors it ironically claims to try to avoid.
There are no G5 notebook computers. The G5 chips burn far too much power to be used in a notebook, unless it's one of those desktop replacement notebooks that last maybe an hour on battery. The inability of IBM to provide a low-power G5 was a major factor in Apple's switch to Intel chips.
http://primes.utm.edu/notes/faq/one.html
General linear groups: http://en.wikipedia.org/wiki/General_linear_group Algebraic structures: http://en.wikipedia.org/wiki/Algebraic_structures
Wolfram Research has some interesting explication on historical methods of solving the quintic: http://library.wolfram.com/examples/quintic/main.h tml
manga is not a style, as the style of comics in japan is far too varied to be a genre.
Of course it can be a genre. This is like claiming that since there are so many different types of animals, there can be no classification "animal". It's a logical fallacy. There are still commonalities, and manga has distinctive stylistic qualities which distinguish it from other forms of sequential art.
you even say that mimicking the "style" of japanese manga means an american comic can be called "manga", therefore you ADMIT that manga does mean japanese...
No. If someone says that mimicking the impressionist style of painting means an American painting can be called "impressionist", it does not mean he's saying "impressionist" is French. It's recognizing that "impressionist" can be orthogonal to "insert nationality here".
I didn't use to have a strong opinion on this issue, but you're definitely making a convincing argument for the other side.
http://www.overclockers.com/tips00827/
What I don't understand is this: when you just listen, what is the difference between you and a wall?
;)
You're a fellow human being.
But seriously, there doesn't have to be a difference. Sometimes just talking about a problem (or writing about it in a journal) makes someone feel better. It's akin to the phenomenon where naming a problem makes someone feel they have a better handle on it.