What do you guys think? Any ovbious mistakes? any missign info? any suggestions?
David Craig
Processor developments
CSCI 230
Prof. Herath
The most powerful computer in the world currently is ASCI white, which is made by IBM using 8192 of their Power3 chips running at 375 MHz. ASCI white cost the government approximately $110 Million, and gave them computing power of about 12,300 Gflops, nearly $9000 a Gflop. Each processor delivers approximately 1.5 Gflops. The number 2 computer, the Terascale Computing System (TCS), uses the Compaq Alpha EV68 processor (Alpha was formerly part of DIGITAL, and is now part of Intel) with 3,024 processors and delivers 6,048 Gflops. Each alpha processor delivers 2 Gflops of performance. Power3 or Alpha based supercomputers dominate most of the top500.org chart.
Later this year we should see a nice jump in processor power for high-end systems. Both IBM and Compaq (Alpha) are releasing new processors. IBM is bringing out the Power4 chip, which should run at 1100MHZ and has 2 cores on a single die. The dies are then fitted into 4 chip package, Giving 8 cores per package. The Power4 will have 62Kb L1 cache and 1.5MB of L2 cache. Alpha is coming out with the EV7, which has been long awaited. It will debut at 1200MHZ. The EV7 will have 62Kb L1 cache and 1.75MB of L2 cache. Both the Power4 and Alpha should be about 400mm2 and both will be on the .13.
An unfortunate development is that there will be no EV8, which was touted as being an incredibly powerful chip. Compaq has decided to sell off all of alpha?s people and patents to Intel. Compaq will then switch to making servers based on Intel?s IA64 architecture. So far one IA64 chip has been produced, the Itanium. So far most people have called this chip a flop, the servers based on it are expensive and do not deliver much higher performance than the Pentium 4. However the next generation IA64 chip, McKinley is rumored to be quite powerful, It should be available late 2003, or early 2004. One of the main improvements will be ?Hyperthreading? or ?Simultaneous multithreading (SMT)? or ?Jackson technology? as it is called inside Intel. Originally to be used in the EV8 this techology makes the software think that a single chip is actually two. The chip then switches between threads efficiently using as much of its execution units as possible. It is speculated to produce a 20-30% increase in performance.
One interesting development of late is use of x86 clusters. One shining example of this is KLAT2. KLAT2 consists of 64 700MHZ AMD Athlon processors. What makes this system so good is its use of flat network architecture to link the computers together inexpensively. Flat network architecture uses common inexpensive 100MBit/s Components, rather than expensive 1Gbit/s hardware. Without going into the details of how FLA works, basically each computer has 4 NICs and is connected to 4 routers, reducing the hops between computers. This cluster currently achieves 64Gflops at a cost of under $1000/Gflop, over 9 times better price/performance than the leading systems
The difference between consumer and server systems gets smaller in 2003. In 2003 AMD will introduce the Hammer architecture. Hammer is an extension of the x86 architecture to 64bits, AMD is calling it x86-64. Hammer will consist of 2 lines of chips; the sledgehammer series and the clawhammer series. The clawhammer will be for home computer systems and dual processing workstations. Sledgehammer will be for 4 and 8 way server systems. A large cluster of 8 way sledgehammers systems should easily make the top 500 list.
In the world of PC processors there are 2 major players, Intel and AMD. They rule the x86 side of PC?s. Cyrix (owned by Via) makes x86 Processors as well however they make < 1% of the market. IBM and Motorola make the G3 processor for apple, which accounts for about 7% of the PC market. A new player to watch in the x86 realm is Transmeta, which makes low power processors for laptops, and small mobile devices.
Currently the ?fastest? x86 processor made is the Intel Pentium 4A (P4A) at 2.2 Ghz. It is made on a .13m process and is 127mm². It has 512Kb of level 2 cache (2x more than the original p4) AMD makes a 1.67GHz processor called the Athlon XP 2000+. It is built on the .18m process and has a die size of 128mm². The reason I put quotes around ?fastest? is because AMD manages to do more instructions per clock (IPC) so the AMD 1.67 GHZ processor is actually is as fast or faster than the P4 2.2 in most applications.
Transmeta takes a different approach to x86, they are looking to create a decent performing processor that is low power and low heat. They plan to achieve this goal using their ?code morphing technology? the processor they developed. Their processor is actually a VLSI Processor that ?morphs? x86 instructions into instructions the processor can understand, and then back again. They claim that their processor, which maxes out at 800MHZ currently, ??consumes 60 to 70 percent less power and runs much cooler than competing chips??1
The computing industry processors have been following ?Moore?s law? since microprocessors were first invented. Moore?s law states that the number of transistors per square inch on integrated circuits had doubled every 18 months since the integrated circuit was invented. This has been modernized into saying that the power of current processors doubles every 18 months. This is readily apparent with the P4, which is expected to take a year to go from 2 Ghz to 3 Ghz, and then hit 4 Ghz by summer 2003.
A major part of moving forward is the ?process? used to make a processor. Currently the standard technology to make processors can make .18m lines. The P4 2.2Ghz is using a brand new .13m process, allowing Intel to put twice the cache on the chip as before. This has made the chip more powerful. It would take a 2.4 Ghz .18m P4 to match a P4A 2.2 Ghz. There are also other factors in manufacturing that can make a chip faster. AMD and IBM have been using copper interconnects (used to connect the silicon to the rest of the chip) that can withstand higher heat and speed than normal aluminum interconnects. Another exiting technology that AMD and IBM are moving forward with is SOI (Silicon on insulator) wafers. SOI reduces the capacitance in a MOS (metal oxide semiconductor) and therefore increases the speed, as it takes less time to discharge. SOI is expected to have a 25-35% improvement over standard processes the same size.
David Craig
Processor developments
CSCI 230
Prof. Herath
The most powerful computer in the world currently is ASCI white, which is made by IBM using 8192 of their Power3 chips running at 375 MHz. ASCI white cost the government approximately $110 Million, and gave them computing power of about 12,300 Gflops, nearly $9000 a Gflop. Each processor delivers approximately 1.5 Gflops. The number 2 computer, the Terascale Computing System (TCS), uses the Compaq Alpha EV68 processor (Alpha was formerly part of DIGITAL, and is now part of Intel) with 3,024 processors and delivers 6,048 Gflops. Each alpha processor delivers 2 Gflops of performance. Power3 or Alpha based supercomputers dominate most of the top500.org chart.
Later this year we should see a nice jump in processor power for high-end systems. Both IBM and Compaq (Alpha) are releasing new processors. IBM is bringing out the Power4 chip, which should run at 1100MHZ and has 2 cores on a single die. The dies are then fitted into 4 chip package, Giving 8 cores per package. The Power4 will have 62Kb L1 cache and 1.5MB of L2 cache. Alpha is coming out with the EV7, which has been long awaited. It will debut at 1200MHZ. The EV7 will have 62Kb L1 cache and 1.75MB of L2 cache. Both the Power4 and Alpha should be about 400mm2 and both will be on the .13.
An unfortunate development is that there will be no EV8, which was touted as being an incredibly powerful chip. Compaq has decided to sell off all of alpha?s people and patents to Intel. Compaq will then switch to making servers based on Intel?s IA64 architecture. So far one IA64 chip has been produced, the Itanium. So far most people have called this chip a flop, the servers based on it are expensive and do not deliver much higher performance than the Pentium 4. However the next generation IA64 chip, McKinley is rumored to be quite powerful, It should be available late 2003, or early 2004. One of the main improvements will be ?Hyperthreading? or ?Simultaneous multithreading (SMT)? or ?Jackson technology? as it is called inside Intel. Originally to be used in the EV8 this techology makes the software think that a single chip is actually two. The chip then switches between threads efficiently using as much of its execution units as possible. It is speculated to produce a 20-30% increase in performance.
One interesting development of late is use of x86 clusters. One shining example of this is KLAT2. KLAT2 consists of 64 700MHZ AMD Athlon processors. What makes this system so good is its use of flat network architecture to link the computers together inexpensively. Flat network architecture uses common inexpensive 100MBit/s Components, rather than expensive 1Gbit/s hardware. Without going into the details of how FLA works, basically each computer has 4 NICs and is connected to 4 routers, reducing the hops between computers. This cluster currently achieves 64Gflops at a cost of under $1000/Gflop, over 9 times better price/performance than the leading systems
The difference between consumer and server systems gets smaller in 2003. In 2003 AMD will introduce the Hammer architecture. Hammer is an extension of the x86 architecture to 64bits, AMD is calling it x86-64. Hammer will consist of 2 lines of chips; the sledgehammer series and the clawhammer series. The clawhammer will be for home computer systems and dual processing workstations. Sledgehammer will be for 4 and 8 way server systems. A large cluster of 8 way sledgehammers systems should easily make the top 500 list.
In the world of PC processors there are 2 major players, Intel and AMD. They rule the x86 side of PC?s. Cyrix (owned by Via) makes x86 Processors as well however they make < 1% of the market. IBM and Motorola make the G3 processor for apple, which accounts for about 7% of the PC market. A new player to watch in the x86 realm is Transmeta, which makes low power processors for laptops, and small mobile devices.
Currently the ?fastest? x86 processor made is the Intel Pentium 4A (P4A) at 2.2 Ghz. It is made on a .13m process and is 127mm². It has 512Kb of level 2 cache (2x more than the original p4) AMD makes a 1.67GHz processor called the Athlon XP 2000+. It is built on the .18m process and has a die size of 128mm². The reason I put quotes around ?fastest? is because AMD manages to do more instructions per clock (IPC) so the AMD 1.67 GHZ processor is actually is as fast or faster than the P4 2.2 in most applications.
Transmeta takes a different approach to x86, they are looking to create a decent performing processor that is low power and low heat. They plan to achieve this goal using their ?code morphing technology? the processor they developed. Their processor is actually a VLSI Processor that ?morphs? x86 instructions into instructions the processor can understand, and then back again. They claim that their processor, which maxes out at 800MHZ currently, ??consumes 60 to 70 percent less power and runs much cooler than competing chips??1
The computing industry processors have been following ?Moore?s law? since microprocessors were first invented. Moore?s law states that the number of transistors per square inch on integrated circuits had doubled every 18 months since the integrated circuit was invented. This has been modernized into saying that the power of current processors doubles every 18 months. This is readily apparent with the P4, which is expected to take a year to go from 2 Ghz to 3 Ghz, and then hit 4 Ghz by summer 2003.
A major part of moving forward is the ?process? used to make a processor. Currently the standard technology to make processors can make .18m lines. The P4 2.2Ghz is using a brand new .13m process, allowing Intel to put twice the cache on the chip as before. This has made the chip more powerful. It would take a 2.4 Ghz .18m P4 to match a P4A 2.2 Ghz. There are also other factors in manufacturing that can make a chip faster. AMD and IBM have been using copper interconnects (used to connect the silicon to the rest of the chip) that can withstand higher heat and speed than normal aluminum interconnects. Another exiting technology that AMD and IBM are moving forward with is SOI (Silicon on insulator) wafers. SOI reduces the capacitance in a MOS (metal oxide semiconductor) and therefore increases the speed, as it takes less time to discharge. SOI is expected to have a 25-35% improvement over standard processes the same size.