Kaveri Is In The House

AMD Launches New A-10 APUs

Developer: Advanced Micro Devices, Inc.

Release Date: January 16, 2014

By Malefico

Ever since AMD rolled out the first generation of its Accelerated Processing Units, codenamed Llano for desktops and Brazos for low-power applications, in January 2011 the company has been working toward a perfect union of CPU and GPU.

With the release of the first Kaveri APUs, AMD has taken a big step toward that synergy. In this article, we’ll take a look at some of the key changes/improvements AMD has introduced with Kaveri, and use this perspective to make some guesses about the future of System on a Chip.

The company has plans to offer Kaveri-based chips for a number of different applications, from desktop models designed around a 95W TDP all the way down to ultra low-power 15W TDP units. According to AMD, Kaveri is optimized to perform right around the mid-point, 45-65W. Although both of their initial offerings occupy the 95W space, with Kaveri the best is yet to come. Low-power mobile devices will benefit from Kaveri technology, and AMD even expects to sell Kaveri chips for embedded uses and servers.

Two new APUs were released just days ago- The flagship A-10 7850K and its slightly less capable brother, the A-10 7700K. Later this year AMD will roll out the A-8 7600, a more conservative chip that may prove to be more capable, dollar for dollar and watt for watt than either A-10.

One minor letdown is that the new chips will require a new socket, FM2+ vs. the FM2 utilized by Trinity and Richland. This is not altogether unexpected, however given the fairly substantial changes in design between even Richland and Kaveri. Anyone who had the foresight to purchase an FM2+ board can use their Trinity or Richland chip until a Kaveri catches their eye, but the new APUs won’t work in FM2 boards.

To begin with, AMD has incorporated its Steamroller cores in place of the Piledriver modules present in Trinity and Richland. Although this does represent a step forward in processing power, the company has simultaneously chosen to, for the first time, emphasize GPU over CPU power on the APU chip.

The new top-of-the-line A-10 7850K incorporates 512 shader units, compared to Richland’s max of 384. This obviously represents a substantial increase in potential graphics power. And, as we’ll see below, AMD has made other improvements to take full advantage of Kaveri’s additional graphics guts.

Beware the marketing jargon associated with Kaveri. AMD say Kaveri offers “up to 12 Compute Cores”. In reality, all the desktop APUs in this lineup, both current and announced for release have quad-core CPUs (two dual-core Steamroller modules. The other “compute cores” (eight and six respectively) are contained within the GPU portion of the die. It’s not exactly misleading; the graphics cores have been tweaked to run more like CPUs than ever before and if developers fully utilize new Application Programming Interfaces the performance of the new APUs will improve over the next several years solely because new and updated software will take advantage of their flexible nature.

They’ve also decreased the scale at which these chips are manufactured, from 32 nm Silicon On Insulator to 28nm bulk silicon. Briefly, SOI manufacturing separates transistors from the underlying cores by a thin, insulating layer, usually glass or silicon dioxide, whereas bulk silicon micro-machining takes slabs of silicon, slices it into wafers, then uses a micro-etching process to carve out homes for the various chip components. At least, this is my understanding and anyone who can come up with a better description, please do so. At any rate, the result is that the new Kaveri chips are able to fit over 2.4 billion transistors on a 245 mm² die. By contrast, Richland used one more square millimeter, but features “only” 1.3 billion transistors. By packing more (albeit slower) transistors on the die, AMD has shifted from CPU to GPU optimization.

What does all this mean? Well, in the real world, as far as raw computing power, not much. In order to increase the graphics power, use a slightly smaller die and still include multiple Steamroller cores, AMD turned down the clock speed on the CPU end of things. This was done both to lower power consumption and prevent Where the best Richland APU, the A-10 6800K was capable of running up to 4.4GHz without overclocking, the Kaveri flagship tops out at 3.7GHz out of the box. So while the new Kaveri units do include more efficient CPU cores, they are scarcely faster than the Richland chips they replace, for now.

To dispel any suspense, very little has changed with respect to AMD’s CPU performance, at least with today’s software. Although the new chips have a slight computing edge over their predecessors in the Trinity and Richland lines, Intel Core processors still handily beat AMD in CPU-intensive tasks like photo editing, video and audio encoding, and office productivity tasks. Not only do the Intel chips work better in an office environment, they use far less power to complete their tasks.

But then, that’s not what the APUs were designed to do well. In entertainment-oriented tasks like gaming and multimedia, AMD has the clear advantage. Although Intel has stepped up a bit with their HD Graphics 4600, the gaming results without a discrete GPU are dismal. AMD APUs are all over Haswell like a cheap suit when it comes to graphics performance. Intel just doesn’t have a desktop processor at any price that works well for gaming unless a video card is added to the mix. Although Intel has steadily increased the GPU to CPU ratio on their processors, AMD has stayed ahead in this capability.

Why? As mentioned above, AMD has dramatically increased the number of shaders on the flagship A-10. Additionally, the new chips utilize GCN (Graphics Core Next) technology. One of the main benefits to the end user is that the new architecture is able to use the GPU portion of the chip to perform tasks traditionally relegated to the CPU- IF the software is written to take advantage of the available GCN cores. OpenCL (Open Computing Language) is a programming language heartily endorsed by AMD, and is able to spread out computing tasks based on available system resources, whether CPU or GPU. Many software developers are embracing this kind of programming as a way to increase performance in many devices, especially those that must also use power conservatively- laptops, notebooks and tablets.

Along with more flexible GPU cores, Kaveri introduces users to HSA and HUMA. Heterogeneous System Architecture means the APU (and board) itself is optimized to distribute tasks to either CPU or GPU as the situation requires. Heterogeneous Unified Memory Architecture allows the APU to dynamically prioritize system RAM in order to achieve the best utilization and efficiency of system resources. While Trinity and Richland limited RAM usage for graphics purposes, in Kaveri it’s wide open. Both CPU and GPU can address RAM up to 32GB. In addition, both are treated equally for task prioritization and utilize the same memory bus to address system memory.

Radeon Dual Graphics finally seems to work as advertised. This nomenclature describes a performance arrangement that combines an APU with a compatible discrete video card in order to make them work together to achieve better graphics performance. Those who have read my previous articles know I’ve brought up glitches in previous iterations of this highly-touted arrangement. Brought into the public eye by the gurus at Tom’s Hardware, the same issues have plagued Dual graphics from its first iteration all the way through Richland. Essentially, although previous combinations have resulted in improved frame rates, many of the additional frames suffered from glitches that made them unusable by the system resulting in graphics performance that ranged from barely better to noticeably worse (when measured with criteria other than raw FPS) than using either the APU or video card by itself. With Kaveri, it seems this technology is finally ready for prime time. Early tests indicate legitimate FPS gains of up to 100% over either the APU or a CPU with video card alone.

Now, I am in no way patting AMD on the back for this. In my opinion, to lure customers with details of an important gaming feature that didn’t perform as advertised in not one, not two, but three successive generations of product is borderline criminal. However, I am glad they stuck with the program to give more diehard gamers a fairly economical upgraded graphics option and the flexibility of the APU platform.

So, what about the future of the APU? Well, for this year at least it’s looking pretty good. I have to say I’m ambivalent about the new A-10s. Yes, they work better than the last generation, but they are substantially pricier. At $170 and $190 for the 7700K and 7850K respectively, I don’t see them as an outstanding value. For the same money you can get a new Athlon X4 and a 7750/7770 GPU and have plenty of gaming power in a more traditional setup.

But later this year, AMD is going to release what I believe will be the next great value in chips, that diamond in the rough that performs much better than its price point. In late 2014, AMD will roll out the A8-7600. Judging from what I’ve read about this little jewel, it will be the APU to have. Slated to launch with a $120 price tag, the A8-7600 finished within a nose of its A-10 cousins in benchmarks, and will cost 30-35% less. With the savings, budget gamers can stretch for a compatible R7 GPU, enable a Dual Graphics feature that does what it’s supposed to, and kick the crap out of either A-10 working alone.

Now, here’s what I think is the coolest thing about the upcoming A-8 chip. It will have two power settings, 45W and 65W TDP accessible and adjustable by the user. Low-power will top the CPU cores at 3.3GHz, while moving to 65W will allow the cores to run up to 3.8GHz. For this writer, it doesn’t get any better than a value-priced chip that’s flexible and frugal on power. I expect the A-8 to become one of my new favorites. Remember at the beginning of the article when I mentioned AMD said Kaveri would perform best in the middle of the 15W-95W scale? The A-8 7600 Note that due to its designation, the A-8 will not be unlocked for overclocking, but I’m hoping AMD rolls out an A-8 later that will allow tuners to bump up the performance. Then again, given the price differential they may well demand that gamers step up to A-10 to gain overclocking capabilities.

Looking at the bigger picture, I can see AMD gaining ground in the low-power device market. With laptops, notebooks and tablet manufacturers placing increased emphasis on gaming and multimedia performance, AMD should have some convincing benefits in these devices. Are we starting to see a market schism where Intel will dominate in high-end gaming rigs, business desktops and laptops and AMD will own, well, pretty much everything else? Stay tuned.

One thing is certain. If the industry keeps moving in this direction, and it most likely will, it will bring builders, hobbyists and specialty computer companies one step closer to obsolescence. With mobile devices occupying an ever increasing segment of the market, and due to the fact that such devices can only be produced economically by large companies with the proper manufacturing tech, the APU, while cool, is just one more nail in that particular coffin.

For a much more detailed and in-depth analysis of the new Kaveri chips, and some great benchmarking info please read the Tom’s Hardware article here.

As always, I hope you found this article informative and I welcome all feedback.