Original Link: http://www.anandtech.com/show/2235
AMD's Next-Generation Mobile Architecture Revealed: Griffinby Anand Lal Shimpi on May 18, 2007 12:10 AM EST
- Posted in
Several years ago Intel discovered surprisingly enough that its NetBurst architecture was not very good for the mobile space. As wonderful as the idea of battery powered space heaters was, Intel quickly discovered that to build the perfect mobile platform you had to start from scratch and design a CPU that was built for the mobile space. By doing so Intel could make tradeoffs that it wouldn't normally make, performance for power reduction, many of which we diagrammed in our first Centrino articles.
Intel also discovered the power of the platform; by bundling a good CPU with a good chipset and wireless controller, three independent Intel products were transformed into a marketing powerhouse. The Centrino brand simplified notebook purchasing and quickly became a mark associated with a notebook you wanted to buy.
It took AMD a bit longer to get on the bandwagon, putting marketing first and worrying about architecture last. We had heard rumors of a mobile-specific AMD microarchitecture, but nothing ever surfaced until now. AMD's design team out of Massachusetts worked on the project, and today we're finally able to tell you about it. The processor is called Griffin, and the platform is called Puma, both are codenames; AMD will undoubtedly come up with a phenomenal name for the final product (sorry we couldn't resist).
When Intel started development on the first Centrino processor, Banias, it had to go back to the P6 for a starting point. The Pentium 4's NetBurst architecture was hardly suitable and the design team was intimately familiar with the P6 core at the time. The end product hardly resembled a P6 and if you look at what the architecture evolved into today, you would be hard pressed to say it was similar at all to a Pentium III.
AMD didn't make the misstep of a Pentium 4, it made a solid evolutionary step to K8 from K7. Griffin's execution core and underlying architecture is based on the current generation 65nm K8 design, not Barcelona/Phenom. You can take everything you are looking forward to being in Phenom and throw it out the window, as AMD is starting from the same K8 core that launched in 2003.
By no means is it a bad starting point, but thankfully AMD did toss in some enhancements. Griffin gets a new North Bridge, a new memory controller, a power optimized Hyper Transport 3 interface and a 1MB L2 cache per core. Griffin will still be built on a 65nm process as AMD will have, at best, only begun its 45nm production by the time Griffin debuts.
Right off the bat you see a disparity between AMD's approach and Intel's approach; while the K8 is arguably a better starting point for a mobile-specific architecture than the P6, the K8 was heavily designed for servers and scaled down. But as we've seen, the K8 is quite power efficient, with 35W TDPs easy achievable for dual core versions, so the race isn't over before it has started.
Griffin will go into production at the end of this year, and AMD is targeting availability in the first half to middle of 2008. Given the launch timeframe, much like Phenom, AMD won't be competing with today's Core 2 processors but rather tomorrow's Penryn based notebooks. Penryn does have some mobile-specific power improvements that even Griffin does not, but the opposite is also true as you will soon see. AMD quoted a maximum TDP of 35W for dual core Griffin CPUs. AMD hopes that notebooks based on Griffin can offer beyond 5 hours of battery life, but do keep in mind that battery life will vary greatly based on OEM implementation.
Truly Independent Power Planes
While architecturally Griffin is no different than today's Athlon X2s, it will draw noticeably less power in normal use. AMD is the first to announce the next step in multi-core power management: independent voltages and frequencies for each core. While Phenom splits the North Bridge and CPU cores into two separate voltage planes, Griffin goes one step further and puts each individual CPU core onto an independent voltage plane. The benefit is that not only can each core run at its own frequency, but it can also run at lower voltages giving you significant reductions in power consumption.
Dynamic power of a CPU can be determined by the following equation:
Power = ∝ * C * V^2 * F
Simply reducing the frequency of a processor (F) will result in a linear reduction in power consumption, but as you can see voltage (V) has an exponential relationship to power. Reducing both is ideal, and that's exactly what Griffin does.
Each core can run at one of 8 frequency steps and five voltage levels, independently of one another. Deep and deeper sleep states are supported, however AMD is currently looking into the possibility of implementing a C6 sleep state similar to what Intel announced for mobile Penryn. AMD wouldn't commit to whether or not we'd see a C6 state in Griffin, leading us to believe that it simply wasn't implemented at the time of Intel's Penryn announcements and there may not be enough time to add it in before launch.
New Memory Controller
Although the underlying architecture of Griffin is K8 based, the memory controller takes a lot of cues from Barcelona/Phenom. There's a new DRAM prefetcher, similar but not identical to what will be in Phenom, but many of the efficiency improvements in the new desktop core will make their way to Griffin as well. Taken from our Barcelona architecture article:
"One strength of Intel's FB-DIMM architecture used in Xeon servers is that you can execute read and write requests to the AMB simultaneously. With standard DDR2 memory, you can do one or the other, and there's a penalty for switching between the two types of operations. If you have a fairly random mixture of reads and writes you can waste a lot of time switching between the two rather than performing all of your reads sequentially then switching over to writes. The K8's memory controller made some allowances for preferring reads over writes since they take less time, but in Barcelona the memory controller is far more intelligent.
Now, instead of executing writes as soon as they show up, writes are stored in a buffer and once the buffer reaches a preset threshold the controller bursts the writes sequentially. What this avoids is the costly read/write switch penalty, helping improve bandwidth efficiency and reduce latency."
AMD did not make it clear whether Griffin also featured two independent 64-bit DDR2 memory controllers or a single 128-bit one. And, of course, as the memory controller is a part of the North Bridge it operates at a separate, lower voltage than the rest of the CPU cores.
Mobile Specific HT3
Much like Phenom, Griffin will support HyperTransport 3, offering more bandwidth between the CPU and the outside world. A major change to Griffin's HT interface however is that it is highly power optimized.
By default, a Griffin CPU has two x16 HT3.0 links (one inbound and one outbound); depending on power and bandwidth requirements, those links can be dynamically scaled down to x8, x4, x2 or completely turned off. The inbound and outbound lanes can dynamically change independent of one another (e.g. inbound could be scaled to x8 while the outbound could be turned off). Each link width change requires a HT disconnect, meaning it can't be done as frequently as a CPU frequency change, but the power savings should be substantial.
Better Thermal Control
Griffin adds better thermal control than its predecessor. The mobile K8 cores simply had a single on-die analog thermal diode that would report CPU temperature, while Griffin features two thermal sensors per core for more accurate thermal monitoring.
The current generation Turion CPUs communicate processor thermal data over the SMBUS, however an external thermal monitoring circuit is used requiring additional board real estate. Griffin supports an integrated SMBUS interface directly to the chipset, so there's no extra chip required.
Griffin also allows monitoring of DRAM temperature (through an external temperature sensor placed near the DRAM modules) and based on pre-configured thermal limits, it can now throttle memory frequency if the modules get too hot.
Just like Intel bundles its mobile CPUs and chipsets as platforms, AMD is doing the same with Griffin. Puma is the name of Griffin's launch platform, but unlike Intel's approach the platform only consists of a Griffin CPU and an AMD chipset; wireless networking is provided by any number of third party vendors. The CPU we already know about, but the chipset is AMD's upcoming RS780.
The chipset supports Hyper Transport 3 (obviously), PCI Express 2.0 (faster link speeds/higher bandwidth) and AMD's answer to Intel's Turbo Memory, Hyper Flash.
The RS780 features an integrated graphics core based on the R600 architecture, however AMD would not provide us any details at this point in time other the following two points. The integrated GPU will feature H.264 decode acceleration, which we're assuming is the same full pipeline UVD from the desktop R600.
Even more interesting is the platform's support for a feature called PowerXPress. The idea is that all Puma notebooks will have integrated graphics, but they can also have optional discrete graphics for better gaming performance. AMD is confident that its integrated graphics will be lower power than any discrete graphics solution, so when you're running on battery power the external graphics core would be disabled and your display would run off of the integrated GPU. On AC power, in the max performance mode, the internal graphics gets disabled and the external GPU is operational.
According to AMD, this switchover happens seamlessly; there's no reboot required and there should be hardly any interruption in use. We have yet to see it in action and are understandably skeptical of how smooth the transition between GPUs would be, but AMD claims that it works and very well at that. AMD also mentioned that this functionality could be overridden by a control panel if need be.
Griffin will use the same S1 CPU socket as current Turion CPUs, however the pinout will be different meaning that these new chips won't be backwards compatible with current Turion notebooks.
As we mentioned above, Hyper Flash will be supported by the RS780 chipset but not over the PCI Express bus. Instead, AMD is using a proprietary interface between the South Bridge and its Hyper Flash controller. According to AMD, PCI Express is not the most power efficient interface to NAND flash so it is using an alternative. AMD expects performance of Hyper Flash to be noticeably better than Intel's Turbo Memory, but without numbers to put to that claim we can't think much of it (it's rare that a manufacturer makes claims about how slow its technology is going to be).
AMD itself stated that it sees the future of microprocessor design as taking two approaches: one for performance and scalability, largely dictated by the server market, and one coming out of the mobile space; the desktop CPUs would scale up and down from the two design points.
We all know very well that the K8 was designed for performance and scalability, so is a new north bridge, memory controller and better power management enough to make it a mobile core? Or would Griffin have benefited from a truly ground up design, similar to what Intel did when it revamped the P6 core? It's tough to say how much more efficiency could be extracted from the K8 if AMD tried.
That being said, there are some impressive improvements to the chip that should yield significant power savings. The revised memory controller and HT interface from Barcelona are welcome additions, but the support for independent voltages per CPU core is particularly impressive.
Platform features like PowerXPress sound gimmicky today but could be significant if well implemented; unfortunately it looks like we've got at least a year to wait before we find out for sure.
As a side note, you may remember that AMD seemed particularly cautious about sharing too much about its plans for Barcelona/Phenom. AMD stated that it didn't want to clue Intel in on what it had up its proverbial sleeve. However, with Griffin not scheduled for release as late as the middle of 2008, we can't help but wonder if all arms of AMD believe the same things, or if AMD is simply giving us excuses for why we don't get more information today.
At least we know what's coming, and Griffin/Puma look to be AMD's best attempt at mobile computing yet. Whether or not it will be enough to dent Centrino is another matter entirely.