AMD (cont’d)

2007 marks the beginning of AMD’s transition from 90-nm to their smaller 65-nm manufacturing process. The smaller process should allow AMD to produce more CPU cores per wafer, but another tidbit about AMD’s 65-nm process that will help boost production is the fact that AMD will begin using 12-inch, 300mm wafers at 65-nm. Today’s 90-nm CPUs from AMD are built on smaller, 8-inch 200mm wafers.

By going from 200mm to 300mm wafers, AMD will be able to produce more than twice as many CPUs per wafer, as well as reduce energy costs. Back when Intel transitioned from 200mm wafers to 300mm wafers, they claimed to use 40% less energy and waste all thanks to the larger wafers. This should help AMD compete better with Intel on cost.

AMD actually plans to begin sampling their first 65-nm chips by the end of this year, with the official introduction set for Q1’07.

As has become their tradition when introducing a new manufacturing process, AMD’s playing it conservative with their first wave of 65-nm parts. Rather than use the new process on an unfamiliar, untested core, AMD’s 65-nm Brisbane CPUs are essentially die-shrunk derivatives of today’s Windsor CPUs. In other words, AMD’s introducing no new features with these chips, and they’ll ship with the same 2x512KB L2 cache configuration used today on the 5000+ and other CPUs, as well as similar clock speeds. Take a look at the table below:

AMD 65-nm Brisbane Athlon 64 X2 CPUs
Model Number	Clock Speed (GHz)	L2 Cache Size
5000+	2.6GHz	2x512KB
4800+	2.5GHz	2x512KB
4400+	2.3GHz	2x512KB
4000+	2.1GHz	2x512KB

As you can see, AMD’s 65-nm variant of the 4800+ is a little different than today’s 90-nm 4800+. The 65-nm chip is clocked 100MHz higher, at 2.5GHz, versus the 2.4GHz of today’s 4800+. In exchange though the 65-nm chip ships with a smaller 2x512KB L2 cache, in comparison to the 2x1MB L2 cache used on today’s 4800+ model. The same applies for the 4400+ and 4000+ chips, they gain an extra 100MHz in clock speed but ship with smaller 2x512KB L2 caches.

Thanks to the smaller process the CPUs will consume less power, AMD lists a TDP (thermal design power) of just 65W for them, in comparison to the 89W of today’s 4800+ and 5000+.

K8L

While the aforementioned chips will definitely be nice performers, AMD’s true answer to Intel’s Core 2 CPU won’t come until next year, in the form of K8L. K8L brings with it a number of improvements. Here are the highlights:

Native Quad-core design

New L3 cache

128-bit wide SSE units

Improved memory addressing (Up to 48-bit memory addressing)

Better prefetching (Going from 16 bytes to 32)

New Extensions added to SSE3

Real World Technologies has a good preview that goes into more details on K8L, and AMD has complete K8L slides here, but the main talking points are that K8L is AMD’s first quad-core design, and that it’s a native solution, unlike Intel’s first generation quad-core which is simply two processor dies dropped into one package.

With four CPU cores inside K8L, power is obviously a huge concern. To reduce power consumption, the CPU’s four cores can run their voltages independently of each other: if you’re running a single-threaded app that’s only using one core for example, the other three processing cores can run at lower voltages to save power.

K8L also boasts a new L3 cache that can be shared across all four cores. The L3 cache will be 2MB. In addition to the L3 cache, each processing core will continue to have its own 64KB L1 cache and 512KB L2 cache. The memory interface between L1 and L2 caches has also been widened to 256 bits. Like Core 2, K8L will support single-cycle SSE: all 128-bit SSE, SSE2, and SSE3 instructions will complete within one cycle, which effectively doubles the execution speed for these instructions. AMD also plans new extensions to SSE3.

The first K8L chip, codenamed Barcelona, is rumored to make its debut in Q2’07. This date has changed quite a bit over the past 12 months though, and is likely largely contingent on how smoothly AMD is able to transition over to 65-nm, a process which has already been delayed by AMD before. Barcelona is expected to be compatible with today's AM2 motherboards.

DigiTimes reports that in Q3’07 AMD will then follow-up Barcelona with “Altair” AMD’s first K8L chip with HyperTransport (HT) 3.0.




HT 3.0 runs at 2.6GHz, providing up to 20.8GB/sec of peak bandwidth. In comparison Athlon 64's HT 1.0 tops out at 1GHz, yielding up to 8.0GB/sec peak bandwidth. According to DigiTimes, the new Altair chips will reside in AMD’s Socket “AM2+”. In all honesty though, we’re not 100% confident in the codename of this new CPU, as it breaks with AMD’s current naming convention. Back in June of this year, AMD announced a collaborative effort with Altair Engineering, which may have led to some of the confusion. In any case, the DigiTimes article mentions two Altair lines, one high-end SKU for enthusiasts that will sell under the high-end Athlon 64 FX brand and a desktop variant of Altair that will go under the Athlon 64 X4 designation. The DigiTimes article also says that these Altair chips are expected to be compatible with AM2 motherboards, but if you want the added bandwidth provided by HT 3.0, you'll want to get an AM2+ motherboard.

Also launching around the same time as Altair is AMD's dual-core variant of K8L, this CPU is codenamed Antares. Antares will support Socket AM2 and AM2+, just like Altair.


Update 10/4/06: This news article from HKEPC is similar to the DigiTimes report, yet it provides more info on Altair and Antares. According to the HKEPC article the Altair CPU intended for the FX line will be built on a new "F+" socket, presumably F+ being AMD's 1207-pin F socket that has been adapted to include support for HT 3.0. This FX CPU would be AMD's first quad-core 4x4 CPU: two quad-core CPUs running in one PC for a grand total of eight processing cores between the two CPUs! The desktop X4 Altair chip would run at the same clock speeds as the FX Altair (2.7-2.9GHz), only it would be used in AM2 and AM2+ systems, with just 1 CPU in the system.

The HKEPC article also gives more guidance on Antares. Motherboard manufacturers told HKEPC that Antares will be clocked between 2.0-2.9GHz and have a 512KB L2 cache and 2MB L3 cache.

Finally, in Q4'07 AMD will introduce two additional K8L variants according to HKEPC, Arcturus and Spica. Arcturus will be contain a 512KB L2, but won't include support for an L3 cache, and will come in speeds ranging from 2.1-2.3GHz. Spica is intended for the value segment, with just a single processing core. Presumably it will ship at slower clock speeds and contain less cache, but HKEPC doesn't provide any specifics.