hard hat area | white paper AMD Multi-Core Processors Providing Multiple Benefits For The Future

y now you probably have heard of variations this year. In 2007 the chipmakers (page 50). This month, we’ll focus on what dual- and multi-core processors. expect to introduce several multi-core AMD has planned for 2006 and beyond. B Both AMD and Intel released chips, beginning with quad-core offerings. dual-core chips in 2005, and both have We discussed Intel’s multi-core proces- Dual-Core Chips plans to release dozens more dual-core chip sors in the January 2006 issue of CPU In April 2005, AMD released its first dual-core chip aimed at the server and market, the proces- Direct Connect Architecture vs. sor. There are several Opteron dual-core Legacy Systems 90nm processors ranging in clock speeds from 1.6GHz to 2.4GHz.

Direct Connect Architecture lives up to its name by providing a for the past few years in its single-core chips. direct connection between the processor, the memory But now AMD has extended the use of Direct Connect Architecture to controller, and the I/O area to improve overall system connect the cores on a dual- or multi-core chip die and to connect each performance. AMD has used Direct Connect Architecture core to its memory controller. ▲

With multi-core architecture, each core on the chip has its own memory con- When using a dual-processor legacy architecture, troller, which significantly improves memory performance. Using Direct Connect however, two processors then have to share the same Architecture to make a connection with the memory controller eliminates most memory control hub, which creates bottlenecks in data bottlenecks and makes multitasking easier. Also, connecting the processor cores transfers at the FSB. The two processors aren’t connected, together lets data flow freely and reduces latency problems. either, which can lead to latency problems.

Source: AMD

Athlon 64 X2 Processor Clock Speed L2 Caches Max Temp Price* (GHz) (°C) Dual-Core Chips 3800+ 2 512KB (x2) 71 $328 4200+ 2.2 512KB (x2) 65 $408 As with other families of AMD processors, a larger model number 4400+ 2.2 1MB (x2) 65 to 71 $507 for an 64 X2 dual-core chip equals better perfor- 4600+ 2.4 512KB (x2) 65 $643 mance on that processor vs. those with smaller model numbers. Each X2 90nm chip listed here runs at 1.35 to1.40V and 4800+ 2.4 1MB (x2) 65 $803 uses a socket. ▲ *Price as of January 2006, for direct AMD customers in 1,000-unit quantities

Source: AMD hard hat area | white paper

The first desktop dual-core processor Direct Connect Architecture. AMD dual-core chips and motherboards that sup- from AMD appeared in May 2005 under plans to enhance this technology in dual- port the AM2 socket will then be able to the brand name. AMD has and multi-core chips to improve data upgrade to a quad-core chip, which also several variations of its Athlon 64 X2 transfer connections among the cores on will use the AM2 socket when AMD intro- processors. (See the “Athlon 64 X2 Dual- the chip. (For more information see the duces it (scheduled for 2007). Core Chips” chart for some examples.) “Direct Connect Architecture vs. Legacy Pacifica technology. AMD’s Pacifica AMD’s first dual-core processors for Systems” sidebar.) technology will improve performance, relia- notebooks and the mobile market should HyperTransport 3.0. AMD and the bility, and security for virtualization hard- appear sometime in the first half of 2006. HyperTransport Consortium continue to ware environments. It should appear in the Multi-core processors from AMD should develop HyperTransport 3.0, which may first half of 2006 and use dedicated transis- initially appear in 2007. offer about three times the bandwidth of tors to deliver the new features while run- version 2.0. HyperTransport 2.0 can offer ning as part of dual- and multi-core chips. Dual-Core Technologies aggregate bandwidth up to 22.4GBps. As (See the “Using Pacifica Technology” side- As AMD releases its dual- and multi- with previous versions of the technology, bar for more information.) core chips, the company will introduce HyperTransport 3.0 will offer direct con- Presidio technology. Presidio technolo- and improve upon several technologies. nections between the CPU and the I/O gy will give users advanced security fea- Cool’n’Quiet. When AMD’s system area. The new version will provide direct tures at the chip level. Although AMD has is running Cool’n’Quiet technology, connections among the cores in the dual- released very little information about how it adjusts the speed of the system fan and multi-core chips. Presidio will work or when it may appear, and the voltage and clock speed of the AM2 socket technology. During 2006 think of Presidio as creating a protected cores on the processor based on the AMD will release new dual-core chips that area on the processor where it can store system case temperature. (See the use an AM2 socket technology, which and process sensitive data. ▲ “Cool’n’Quiet Technology” sidebar for uses a different pin configuration than more information.) AMD’s 939 socket. Customers who choose by Kyle Schurman

Inside AMD’s Dual-Core Processor

CPU 1 and CPU 2. The dual AMD64 64KB D-Cache. The L1 data cores on this chip can run 32- and internal caches are low latency. 64-bit computing simultaneously. 64KB I-Cache. The L1 instruction System Request Queue. internal caches This area manages how offer low latency. each core accesses the crossbar switch.

1MB L2 Cache. The L2 second- Link 1/2/3. level internal These three links caches offer connect the I/O low latency. area to the processor using HyperTransport Crossbar. The crossbar makes the technology. connection between each core and the HyperTransport I/O area and memory interfaces. The also can provide a crossbar is a key component of the chip, high-speed letting the core access the data it needs to link among run software and perform calculations. processors in a multi-processor configuration. Integrated DDR Memory Controller. This feature reduces the latency associated with accessing memory vs. using an FSB architecture to access 72 bit. This DDR400 memory interface memory. Each core has its own memory controller. provides data transfer rates up to 6.4GBps.

Source: AMD hard hat area | white paper

Multi-Core Benefits

Improved performance is pushing the migration to multi-core proces- • Users can add more computing power without the cost of adding sors, but that’s not the only reason: Multi-core processors are also another computer. This feature especially benefits commercial users, appearing out of necessity. letting them add more processing power without adding more servers. The technological advances that have driven manufacturers to double the By adding fewer servers, companies will need less real estate to hold number of transistors on a chip every 18 to 24 months (fulfilling Moore’s the servers. Also, the costs of electrical power to run those servers and Law) are beginning to reach their physical limits. Chipmakers have continu- the cost of cooling the servers will decrease. ally shrunk the manufacturing process for transistors over the years. They’re • For those who compile software code, dual- and multi-core processors currently shifting the manufacturing process from 90 to 65nm, which lets seriously improve compiling efficiency. AMD says its current dual-core them squeeze more transistors onto each chip. An IDC report, however, processors reduce the time needed to compile code by as much as 50% says that once the chip manufacturing process reaches about 16nm in size, compared to a single-core processor. the processors won’t be able to control the flow of electrons as the flow • Game developers can add more features and cutting-edge graphics to moves through the transistors. This means that transistors eventually will their games because dual- and multi-core chips will more easily and reach a size where chipmakers can no longer make them smaller. Ever efficiently handle multithread software designs. smaller and denser transistors on a chip generate more heat, causing pro- • Multi-core processors don’t consume more power or generate more cessing errors. But multi-core processors can improve computing power heat vs. a single-core processor, which will give users more pro- and limit some of the problems that shrinking transistors are causing. cessing power without the drawbacks typically associated with The one drawback to multi-core technology is the increase in cost such increases. ▲ for systems and chips. However for many users the benefits will out- weigh the cost factor.

A processor with two or more cores works faster and more efficiently than a single-core When running a single application, dual- processor for When running two processor- cores can share the processing load, several reasons: When multitasking, users will intensive applications, each one improving overall system performance. experience fewer bottlenecks can access its own core, which The system also can shut down portions than with single-core systems. makes it easier to burn a CD while of the cores that aren’t in use, saving running a virus scan, for example. power and generating less heat.

Source: AMD AMD’s Multi-Core Processor Future

When AMD designed the AMD64 processor during K8 Dual Core. This is the server and workstation Santa Rosa. Expected to appear in the first the late 1990s, the company had dual- and multi- Opteron dual-core chip that debuted in April 2005. half of 2006, this dual-core chip should be part core technologies in mind. Such planning has made K8L. This quad- or eight-core chip will poten- of the Opteron 800/200 series. AMD’s evolution into multi-core products an easier tially follow K9 in 2007 or 2008, although little is Taylor. Aimed at the mobile market, this process. For example, AMD designed some of its known about this project. dual-core chip should launch under the Turion current dual-core processors to fit in the same sock- K9. Due to launch in 2007, K9 should be a 64 brand name. Taylor should appear early ets as its single-core offerings, such as the 940-pin quad-core chip appearing under the Opteron in 2006, use an AM2 socket, and support Opteron processor and the 939-pin Athlon 64 brand name in servers. It will feature a new core DDR2 memory. processor. Also, all of the software that runs under design, DDR3 memory, and contain L3 cache, Toledo. This is the desktop Athlon 64 X2 x86 and AMD64 processors will work with AMD’s which will let server designers build systems dual-core chip that debuted in May 2005. multi-core processors; no coding changes needed. using up to 32 processors. (Current cache set- Trinidad. Due sometime in 2006 and intend- During the past couple of years, AMD has tings allow for only eight processors per system.) ed for the mobile market, this 90nm AMD64 moved away from releasing timelines listing detailed AMD road maps indicate quad-core chips for dual-core chip will use an AM2 socket. plans for its future processors complete with code desktops will initially appear in 2008. Windsor. This 90nm dual-core chip should names. Instead, AMD’s beginning to give more gen- Santa Ana. This dual-core chip should be part appear early in 2006, use an AM2 socket, and eral names to overall chip projects, rather than split- of the Opteron 100 series, use an AM2 socket, and support DDR2 memory. ▲ ting out code names for individual chip plans. appear in the second half of 2006.

Sources: AMD, AMDboard.com, and Endian.net hard hat area | white paper

Using Pacifica Technology

Virtualization uses software to share and manage workloads at the By building virtualization technology inside the processor, as AMD will processor level, making it appear as though the server contains a multi- do with Pacifica technology, multi-core processors could then provide processor system. A single server can run multiple OSes and applications better overall performance vs. a single-core processor trying to run virtu- using virtualization. Implementing the idea of virtualization also means alization. Pacifica, which AMD will introduce later this year, would simplify companies don’t need to have as much hardware performance capability the implementation of virtualization and let it more easily take advantage held in reserve among their servers to meet peak situations. of the multiple cores on a processor. ▲

In a system without Pacifica With Pacifica running on an AMD technology, the x86 processor dual- or multi-core processor, there hardware contains no virtu- would be fewer layers and less alization capabilities. When cre- complexity, improving application ating a virtual machine in this performance. Pacifica would use type of system, the virtualiza- Hypervisor as its virtualization tion software must manage the software, which would manage the resources between the host OS virtual machines. Hypervisor also and the guest OS. Because this would track the availability of extra layer causes additional physical hardware, letting overhead and complexity, ap- applications take advantage of the plication performance suffers. hardware as it becomes available.

Source: AMD Cool’n’Quiet Technology

AMD has carried its Cool’n’Quiet tech- The minimum state works well for nology into its dual-core offerings. It cuts computing tasks involving a lot of idle down on the heat and noise that a proces- processor time, such as word process- sor generates and helps the processor use ing and emailing. Intermediate states less power by cutting back on the its clock work well for tasks requiring a lot of speed and voltage when the computer continuous processor access, such as user’s demands are less. AMD says users system scans with no other tasks rarely will notice any system performance occurring or processes running. degradation while using Cool’n’Quiet As the Cool’n’Quiet driver works because it can adjust the clock speed and with the motherboard to measure sys- voltage up to 30 times per second. tem temperature and current processor To run Cool’n’Quiet, a PC needs a workload (based in part on the type of heatsink with a variable speed fan, which software a user is running), it places adjusts the fan’s rotation speed based on the processor in the most appropriate the computer case’s air temperature. The state. At the same time, the variable system also needs a Cool’n’Quiet driver. speed fan speeds up or slows down to When running Cool’n’Quiet, the CPU match the change in the processor’s operates in one of three basic states: max- state. As the processor decreases its imum, intermediate, or minimum. The in- clock speed and uses less power, the termediate state can have more than one processor and power supply generate setting, and the more settings available, less heat. As the temperature in the the more flexibility the processor has to case falls, the system can decrease save power and reduce heat. the fan’s rotation speeds, leading to For example, AMD’s Athlon 64 3500+ less noise from the fan and from processor can operate in one of four air turbulence. states, measured by the clock speed and In contrast, when a user cranks up voltage in use. system performance to the maximum Maximum: 2.2GHz; 1.5V state, Cool’n’Quiet increases the fan’s Intermediate A: 2GHz; 1.4V rotation speed to remove the addition- Intermediate B: 1.8GHz; 1.3V al heat it now generates, increasing Minimum: 1GHz; 1.1V system noise. ▲

Source: AMD