AMD CPU Roadmap Update

We haven't updated our AMD roadmap for a few months, in part due to overlapping NDAs and product launches. Unfortunately we are just about to go under NDA for AMD's next low clock X2 processor - but rather than postpone another roadmap we decided to carry on just without that information. Our last detailed AMD roadmap was actually back in December, and we have quite a bit of new information to cover. We'll focus primarily on the upcoming parts, which will be highlighted in red in our tables, but we'll also list details on many of the currently shipping parts. Let's start with the desktop sector.

Introducing M2

The big news of course is the upcoming transition from socket 939 to socket M2. M2 will have 940 pins, and while the actual layout of motherboards with the new socket will be very similar to socket 939, processors will obviously not be pin compatible. The same goes for current socket 940 processors: they won't work or fit in socket M2. That makes sense, considering that M2 will require the use of DDR2 memory. The similarity in layout will allow motherboard makers to quickly adapt 939 designs to M2, however.

Initially slated to support up to DDR2-667, the socket M2 processors will launch in the second quarter of 2006. Codenamed Orleans for the single core processors, the roadmaps include few details on the clock speeds or cache amounts of the new parts. Without speculating too much, we would assume that the specifications will be the same as equivalent 939 processors, but we could be wrong. Another possibility is that the DDR2 parts will outperform the socket 939 parts, allowing AMD to modify their names (again). Either option seems equally likely, and we wouldn't be surprised to see the 4000+ M2 ship as a 2.6 GHz 512K part, as that would allow the use of a single design for the three listed chips.

Processors running on the new M2 socket are dubbed Revision "F". Revision "F" processors carry all the features of existing "E" CPUs, but along with the new socket comes some significant changes in power. For example, a typical Revision "E" San Diego Athlon 64 utilizes 80 amps with a maximum TDP around 90W. The new Revision "F" CPUs will use 95 amps instead. However, although the processors require a bump in the current, the power remains the same. Unfortunately, expect massive increases in TDP. From the roadmap we expect the FX M2 processors to have a max TDP of 125W, 110W for dual core and 104W for single core. Revision "F" implements a new Bi-Directional PROCHOT so that the motherboard can throttle the processor dynamically, but the CPU can also override and throttle itself if necessary.

The main changes to the processor diagram only appear to be the integrated DDR2 controller instead of the integrated DDR1 controller.

The Desktop

Besides the single core M2 processors, we also have the dual core parts, codenamed Windsor. We have a 4800+ part for socket M2, as well as parts greater than or equal to 5000+ for both socket 939 and M2. As with the single core variants, we see two likely possibilities. The first is that the M2 parts will directly match the 939 parts in features, making the 4800+ a 2.4 GHz 1MB per core design. On the other hand, improvements in the performance of the platform through the use of DDR2 may allow AMD to use a 2.4 GHz 512K part for the 4800+. Part of the attractiveness of such a change is that the 2x512K parts would have a smaller die size, decreasing the manufacturing costs. Time will tell what AMD's precise plans are, and at present all we have are model names with which to speculate.

The last addition to the performance desktop market is a new FX model, the FX-59. While in the past we have speculated that AMD would switch the FX to a dual core design, the roadmap instead indicates that the FX-59 is intended to use a San Diego core running at 3.0 GHz. Other details in the roadmap show that AMD intends for the FX line to offer top single core performance. AMD market the FX line as being "ideal for 3D games and single-threaded applications." The X2 line on the other hand is "designed for sophisticated power users who run multiple processor intensive applications simultaneously," and is "ideal for digital content creation as well as for listening and viewing entertainment." It seems pretty clear from such statements that the FX line will remain single core for a while longer.

M2 isn't just a switch to DDR2 memory support, though. Besides DDR2, AMD indicates that the Orleans and Windsor processors will also have "Security and Virtualization" features. Previously codenamed Pacifica, the virtualization technology adds hardware support that allows a system to run multiple operating systems simultaneously. It was possible to do this in the past with such tools as VMware, but hardware support should dramatically improve performance. The security technology mentioned is likely to be very similar to Intel's LaGrande Technology and was code named Presidio. Whether or not you'll need a tin foil hat to operate your computer is yet undecided.

You'll notice that socket 754 is no longer even represented in the desktop arena. There are a couple parts that will work on socket 754 motherboards, but they are no longer marketed as mainstream or performance desktop processors. That brings us to the desktop Sempron processors.

There are quite a few upcoming parts in the value desktop computing sector. Launching very soon will be the Ex stepping of the Palermo core, which will add - or rather enable - 64-bit extensions on all the Sempron parts. All of the presently available Sempron parts are for socket 754 (and even a few older socket A models that are being phased out), but we also have Sempron parts targeting socket 939 that should be launching any time now. Information contained within the roadmap actually seems to indicate that 3000+ and 3200+ Sempron parts have already started shipping for socket 939, though we haven't seen any. It could be that they are going straight to OEMs, as the same parts show up as DTR (Desktop Replacement) notebook processors.

Looking forward to 2006, we still have a few more speed bumps to the Sempron line on both sockets with the 3500+, 3600+, and 3700+ parts. Somewhat interesting to note is that the Sempron 3700+ will apparently be a 2.2 GHz 256K cache part on both sockets - normally AMD has rated the equivalent clock speend and cache socket 939 parts slightly higher than socket 754 parts due to the dual-channel memory support of 939.

Besides the above value processors, we also have some information on the upcoming M2 value processor. Code named Manila, the processor won't support the virtualization and security features that other M2 processors have. That isn't too surprising, as frankly we expect virtualization technologies to be far more useful for the performance sector. However, it will support dual channel DDR2; typically we don't see dual channel support on value processors, so that's a welcome change.

Mobile and Transportable Processors

AMD has been trailing in the mobile performance per Watt competition ever since Intel launched the Pentium M. While they have announced a name change to their Mobile Athlon 64 lineup, they still appear to be using a low power Athlon 64 derivative as opposed to something designed from the ground up for the mobile sector. The good news is that AMD's base designs are generally far less power hungry than Intel's Pentium 4, so a separate design isn't really required. Like Intel, AMD has three categories of mobile processors. First are the DTR models which are essentially just desktop parts put into a laptop, although they may run with slightly lower voltage requirements. The next level up the mobility scale is referred to as Transportable, and the power requirements are 62W or less. Finally, the true Mobile parts are classified as either 35W or 25W parts, with the 25W targeting the Thin and Light category of notebooks. The DTR processors can basically be any of the regular desktop parts, so we'll skip that segment and start with the Transportable chips.

At the high end of the Transportable segment are the Athlon 64 Mobile parts. All of these parts feature 1MB of L2 cache, and most have been available for some time. AMD has recently transitioned from the older 130nm Odessa core - basically a lower power version of the Clawhammer - to the 90nm SOI Newark core. The only new Mobile Athlon 64 part is the 4000+, which runs at the same clock speed as the FX-55 processor but only supports single channel memory. As virtually any socket 754 board should have no trouble supporting these mobile variants, 754 owners looking to upgrade for additional CPU performance might be interested in checking out these parts. The price is generally a bit higher, and a switch to socket 939 is probably the better course of action if you can manage it, but a change from the 1.8 GHz 512K 2800+ to the 4000+ would provide a substantial boost to performance.

The Mobile Sempron is also undergoing a change from the older Georgetown core to the new Albany core (and before Georgetown was the Dublin core). You might still see some of the older core versions out there, but the Albany core is the preferred model, as the 90nm SOI provides much better power and thermal characteristics. While the desktop Sempron parts are slated to get 64-bit support, the mobile parts will remain 32-bit only for the time being. To quote the roadmap, "AMD will introduce 64-bit enabled Mobile AMD Sempron only when it makes sense for our value notebook customers." In other words, most value notebooks ship with lower end components, so 64-bit addressing isn't going to be terribly important for a laptop with 256 or 512 MB of RAM.

In the future, the Mobile platform will also transition to DDR2 support, only the thin and light notebooks will use socket S1 rather than socket M2. S1 appears to mostly be targeting a smaller package size, as it will also be a dual-channel DDR2 platform with support for the security and virtualization enhancements. Two code names appear on the roadmap for S1 parts, the Taylor core will be a dual core Turion 64 processor with security and virtualization support while the Keene will be a single core part without Pacifica and Presidio technologies. For the full size notebooks and desktop replacements (DTR), M2 will be used and the Trinidad core bears the Mobile Athlon 64 moniker with specs identical to Windsor - though cache sizes could differ, we admit. That means dual core, dual-channel DDR2, and security plus virtualization technologies. A Mobile Sempron part based off of the Richmond core drops the virtualization support and runs on a single core. All of these DDR2 mobile parts are scheduled for a Q1'06 release. Having mentioned the Taylor core and Turion 64 brings us to the next group of mobile processors.

We had a small article on the Turion 64 Launch, but there hasn't been a lot to say since then. OEMs haven't been quick to jump on the Turion bandwagon, though there are laptops with the processor available now. The Turion naming scheme is similar to that of the Opteron in that a two digit model indicates relative performance, with higher numbers being better. The second letter indicates suitability for mobile use, with "A" being less suitable and "Z" being ideal. On present models there are only two letters used, L and T, and they correlate to the typical maximum power requirements. "T" models have a maximum TDP of 25W while the "L" models have a maximum TDP of 35W. While both are higher than the 22W rating of Pentium M parts, the T variants ought to perform similarly with Cool n' Quiet enabled. As you can see in the table, the number designation at present is the same for similarly clocked Turion chips, though the T models (model Ts?) will cost more.

All of the Turion parts currently use the Lancaster core, which is a 1MB socket 754 part made on 90nm SOI. As we mentioned before, dual core Turion parts made with the Taylor core will move to socket S1 in mid 2006, competing against the dual core Yonah parts from Intel. Which part will actually be better is anyone's guess right now, as we expect both platforms to offer several improvements over current mobile parts. The one advantage that AMD does have is 64-bit support - we haven't heard anything about Yonah support EM64T yet, so we would guess that initial shipments will remain 32-bit. With Longhorn also scheduled to ship in 2006, the move to 64-bit applications may finally pick up steam in the mainstream market. Turion isn't the only low power processor AMD has, though, so let's take a look at the Thin and Light Sempron parts.

All of the Thin and Light Semprons have a TDP of 25W, the same as the Turion MT chips. They also run on socket 754, like the Turion, though future parts will move to S1. Previous Mobile Sempron chips have used the Dublin and Sonora cores - mobile variants of the Newcastle and Paris desktop cores, respectively. The interesting thing is that the earlier Dublin-based Mobile Semprons used to bear the Athlon XP name, even though they were Athlon 64 derivatives since they ran on socket 754. AMD is now switching to the Roma core, which is the Mobile version of the Palermo core. It adds SSE3 support to the design, among other small tweaks. Most of the Roma chips should be available shortly, with the only new model being the Sempron 3400+ scheduled for early 2006.

There are a lot of overlapping code names and models in AMD's mobile sector, and to be honest we're not exactly sure what differences - if any - exist between the Turion Lancaster and the Mobile Athlon 64 Newark. They may be identical with the difference being binning, or there may be some small architectural tweaks that separate them. The same holds for the Albany vs. Roma and Georgetown vs. Sonora comparisons. But enough about mobile chips; let's get on to the last sector.

Server and Workstation Processors

Not much has changed in the server and workstation segment other than the actual launch and availability of the dual core Opterons. Here's what we have looking at the present and near future.

We've separated the 8xx, 2xx, and 1xx Opteron models into their individual parts for a reason. While older Opterons all targeted socket 940, AMD is apparently moving the 1xx models towards socket 939. The most likely reason is for price/performance advantages of the platform, or perhaps that yields of the newer parts have reached the point where everything works in two socket or higher configurations. 152 is the last single socket Opteron appearing on the roadmap, though determined single socket 940 holdouts can always drop in the more expensive 2xx parts if they want.

Looking at the future, the only new Opterons coming out will be the single core x54 models. As we just mentioned, 154 will actually be a socket 939 part (along with the now available 165, 170, and 175). The switch to socket 939 also means that the 154 will use the San Diego core rather than the Venus core that previous 1xx 90nm parts have been based off. 254 and 854 will also exist, and we must admit we're a little surprised - several of us were surprised to see the x52 models. Almost any application that can take advantage of SMP will perform better on two 265 cores than on two 254 cores, but AMD apparently has received enough requests for another single core Opteron model that the x54 parts are being made. The price of the 254/854 parts is identical to that of the 265/865 parts, however, so the parts are essentially being rated as equivalent. Users that have applications that depend more on pure clock speed rather than multiple threads are the likely target of the x54 parts. The socket 939 154 part is priced the same as the FX-57, and actually costs more than the 254, so single socket 940 users shouldn't fret too much. The 939 152 is also priced the same as the FX-55, while the 940 152 is substantially cheaper.

Besides the various Opteron models most of you have probably seen before, we also included the HE models. HE stands for "High Efficiency" and the efficiency we're talking about is performance per Watt. Where the standard Opterons have a TDP of 95W, the HE models are only 55W. You basically give up two bins of performance for the lower TDP, however, so the 865 HE costs as much as the 875 and the 850 HE costs as much as the 865. They also max out at two levels below the fastest models, which isn't too surprising.

Looking to the future of Opterons, the current roadmap doesn't contain any information about what's planned. DDR2 is in the works for all the other markets, so it would make sense for Opteron to eventually move that way as well. However, there have been difficulties in the past with getting ECC and registered memory to work with DDR2, so perhaps Opterons will move to FB-DIMMs (Fully Buffered DIMMS) instead. We do know that AMD has something in the works called socket F, a 1207 pin socket for future Opterons, but we don't know what type of RAM is used. That transition is likely more than a year off, but we'll let you know as soon as we get any clearer picture of what AMD has planned.

Final Thoughts

Echoing what we said in our last AMD roadmap, there aren't a whole lot of speed increases showing up on the future roadmaps. FX-59 will add another 200 MHz to the fastest AMD processor in terms of clock speed, and we can guess that the X2 5000+ will do the same. Again there is another low clock X2 coming up that we can't talk about just now, but fortunately you won't have to wait too long for details on that processor to emerge. Rather than focusing on increasing raw clock speed, AMD and Intel have both shifted to improving the features of their various chips. Dual core was the first step in that direction, and quad core (or maybe tri core) is a likely evolution at some point. 65nm processors are also on the horizon, and likely AMD will begin releasing the first such processors just beyond the range of the present roadmap - i.e. in late 2006 or early 2007. The process shrink should bring improved clock speeds as well as more cores per die. The increasing numbers of mobile parts are another indication of the changing goals. We may not be able to get much faster without spending significant effort, but we can try to focus on making the current designs more portable at the very least. We'll take a look at the Intel roadmap in the near future, but the trends are similar to what we see with AMD.