Sign in with
Sign up | Sign in

CPU + GPU = APU: East Meets West

The AMD A8-3500M APU Review: Llano Is Unleashed
By

We can see that the CPU and GPU hardware aren’t much different from what AMD already has in the market. But what makes the Fusion initiative unique is the combination of these two strengths in a single package. Company representatives stress that the challenge of Llano was to get everything working together in the best and most power efficient way. That comes down to the plumbing.

There are five main components of Llano that have to communicate with each other: the CPU complex, the GPU complex, the northbridge, the traditional I/O block, and the DDR memory I/O block. The CPU-to-northbridge link isn’t anything new, so let’s talk about what we haven’t seen yet: namely, the GPU-to-northbridge links. Because there are two GPU usage scenarios, it has two dedicated links.

The first link is called the Radeon memory bus. It’s no secret that poor bandwidth equals poor graphics performance, so the GPU can’t rely on the same memory interface as a typical CPU. The Radeon memory bus is a direct path from the GPU through the northbridge to the DDR memory, and it provides priority access to RAM for high-bandwidth operations. AMD’s chief engineer Mike Goddard made it clear that this link is absolutely essential to extract discrete-class graphics performance from the Llano APU.

But there is a second link that the GPU needs, and this is one of the components that really makes Llano an APU instead of just a CPU and a GPU on the same die: the Fusion compute link. Traditionally, I/O devices have been able to go through the PCI Express interface to access the CPU’s cache. This is a bottleneck for GPU compute operations, so AMD augmented this path and gave the GPU better access to share data with the CPU or to extract memory contents from that shared pool. The Fusion compute link is the piece that allows Llano to extract compute performance at a very efficient power level.

Power Management

One of the three main features of Llano is what the chip company brands AMD AllDay Power. AMD considers the mobile market an ideal place for the APU, so it makes sense that power efficiency is high on its priority list. So, how is this achieved on the A-series?

There are two main voltage rails responsible for Llano’s power usage: the VDD rail (shared across the CPU cores) and the VDDNB plane (shared across the GPU, UVD block, graphics memory controller, and northbridge).

At first you might assume that sharing the single VDD voltage rail across four CPU cores might not be the most efficient way to control power when the operating system may often only require cycles from one. But this single supply has two modes: Core C6 (CC6 mode) and Package C6 (PC6 mode). CC6 mode is capable of powering down individual CPU cores, while PC6 mode is able to lower the power on the entire rail. This provides the granularity needed to best handle a variety of situations.

The VDDNB plane is a bit more complex, as it’s responsible for a number of components. From a pure efficiency standpoint this isn’t an ideal situation because the whole rail has to run at the highest common voltage—that is to say, the highest voltage required by any one of these components at a time. This design decision was chosen because breaking it out of the package and adding regulators would increase cost. As a result, the voltage and frequency P-states vary based on a number of factors: the northbridge P-state, the GPU P-state (something usually controlled by the driver, but that also can be altered through hardware), the PCI Express speed (if a second-gen PCIe device is present it drives a higher voltage requirement), and the UVD workload.

But while voltage is shared, there remains significant granularity in which components can be gated. The GPU portion can be powered down if idle time goes past a programmable threshold, or if the driver detects and responds to usage. The graphics memory controller is capable of using significant amount of power, so it can be turned on and off as required. The UVD block can also be used or gated as needed.

There are some display power optimizations in play here, too. Standard frame buffer compression is used, of course. But AMD enhanced the adaptive backlight modulation (ABM) component. This can analyse the image that is to be shown and tweak the backlight and brightness levels to display an image as close to the original as possible using the least amount of power.

The following images show how specific portions of the APU can be bypassed: blue is ambient and green is active. These are only a few examples, as other components like the CPU cores can also be controlled.

Turbo Core

Turbo Core is nothing new. AMD included this frequency-boosting feature as far back as its Phenom II X6 introduction. What’s different is that Llano balances GPU and CPU resources to fit within the chip’s TDP.

AMD’s Mike Goddard stressed that the company’s approach to this feature is unique in that it is primarily based on digital performance measurements, not analogue thermal readings. The APU’s digital APM (Advanced Power Management) module measures activity to estimate power requirements and dithers the P-state to stay within the chip’s TDP. The benefit of this approach should be consistent and repeatable performance compared to one primarily based on temperature measurements. Of course, the downside would be that performance headroom might not be fully utilized if the APM underestimates headroom versus actual temperature limitations.

In any case, when the APM decides that headroom is available, it invokes an OS-invisible state called “P boost” that increases clocks on the CPU-side. Keep in mind that Llano’s GPU cannot be accelerated beyond its shipping speed (it can, however, be throttled down to cut power and minimize heat)—only the CPU clock can be boosted. It’s also important to note that the GPU always takes precedence, so whenever there’s a graphics load, the CPU has less opportunity to see the benefit of Turbo Core.

But there’s a final piece to the puzzle. While temperature is not the primary determining factor of P-state, it remains an important failsafe. There is one scenario where both the CPU and GPU can be stressed beyond Llano’s TDP: when it comes to load-balancing an OpenCL compute application. Because the GPU is more effective at this type of task, it is once again given priority. But the CPU also pushes against its limit, and if the temperature exceeds that barrier, the CPU can drop to less than P0—for instance, P1 or P2—in order to bring temperatures down. We were told that this is the only scenario where Llano operates its processing cores below their nominal speed, but it can happen.

Unfortunately we’re unable to test any of those scenarios because no tools exist that report the true clock speed of Llano’s CPU cores. Even the monitoring utility AMD gave us incorrectly reports the CPU clock as its nominal level, without ever seeing the effect of Turbo Core when the GPU is idle. Clearly, it’s important to validate the company’s claims, as we’d like to see first-hand what scenarios  can force the APU to throttle back.  You can be certain that we will revisit this issue as soon as we figure out a way to properly measure what is going on inside the Llano APU under different loads. 

As mentioned, though the CPU clocks can scale up and down, the graphics engine won’t exceed its factory speed; it cannot be pushed any further by Turbo Core. According to AMD, this may change in the future. Our interpretation is that the upcoming Trinity architecture may feature the flexibility to accelerate graphics performance when the CPU cores have thermal headroom to give up.

Ask a Category Expert

Create a new thread in the UK Article comments forum about this subject

Example: Notebook, Android, SSD hard drive

Display all 6 comments.
This thread is closed for comments
  • 1 Hide
    Jay_83 , 14 June 2011 12:42
    Hmm. Mixed feelings about this one. One the one hand, it's less impressive than I wished for. On the other, it's obvious AMD has done some good work here, in more than one direction.
    Ultimately, We'll have to wait and see what Trinity has to offer to give the final word. Here's hoping Bulldozer throws a serious punch. I'm sure we'd all like to see AMD back in the game, and not just in notebook space.
  • 1 Hide
    aje21 , 14 June 2011 21:50
    If you want us to look at this as the end of entry level discrete GPUs for light gaming, how about comparing it to the GPU running at its real speed? Once you have a real price for an APU+M/B you can put it up against a similar spec/price CPU+GPU+M/B to see how they compare.
    If the transcoding performance was worth having (and supported by Windows 7 in Media Center) then this could be an HTPC route...
  • 3 Hide
    daglesj , 14 June 2011 22:04
    Chaps, you are wasting your time commenting on the UK version of the article. All the comments are over on the US version.

    Why they cant pull the 'shared' article comments together I dont know.
  • 1 Hide
    Anonymous , 15 June 2011 09:22
    Seems like the Turbo CORE was simply turned off... Perhaps it needs more than enabling it in the BIOS. Like in case of Cool'n'Quiet, that you has had to enable in Windows, as well?
  • 1 Hide
    silverblue , 15 June 2011 19:36
    ajaraSeems like the Turbo CORE was simply turned off... Perhaps it needs more than enabling it in the BIOS. Like in case of Cool'n'Quiet, that you has had to enable in Windows, as well?

    Yeah, by selecting Minimal Power Management in XP or Balanced in Vista/7.
  • 0 Hide
    wild9 , 24 June 2011 23:31
    daglesjChaps, you are wasting your time commenting on the UK version of the article. All the comments are over on the US version. Why they cant pull the 'shared' article comments together I dont know.


    I've often thought the same..I wish all the UK and US comments could be combined. I discovered the duality by accident, and also noticed that the UK version of the article has 4 comments, whereas the US had 4 pages. I am grateful for all this tech news..I just wish the comments section could be combined. Also, if they could send the spammers to Mars..that would be a nice bonus.