AMD clearly made a long-term investment with APU as a concept. This is, in our opinion, the biggest innovation that AMD made in the recent decade. Simply put, we all love expensive video cards, processors and motherboards, but when the time comes to pull out our wallets, the situation drastically changes. The priorities get changed, and the price is more often than not the main factor during decision-making process. In those terms, for most of people, APU is the most logical choice. Richland has been present on the market for a long time, and even though it was a novelty, it shared much of its characteristics with its predecessors. Now, for the first time since Llano appeared, in front of us we have something completely new, and that's, of course, Kaveri architecture.
Steamroller + GCN 2.0
First of all, the basis of this whole idea is the production process. AMD switched from the old 32 nm to the 28 nm lithography, which allowed for increase in number of transistors. Also, in case you're wondering why CPU didn't have a smaller production process, the answer lies in one, simple fact. AMD wanted homogenous build, and that's the reason why none of the components are made with different technologies. While we're at the full-blooded version, the total area of the most powerful version's chip is 245 mm2 and it consists of 2.41 billion transistors. Let us remind you, the most powerful Richland has an area of 246 mm2 and only 1.3 billion transistors. This is a huge improvement, bearing in mind that AMD didn't dabble in 22 nm production processes, even though a direct comparison isn't viable due to the differences in GPU architecture, but also in other modules also.
The huge difference in amount of transistors suggests there will be an improvement in performances. The first step in platform improvement is the change to Steamroller architecture for CPU. As was to be expected, it's an advanced version of Piledriver, i.e. third generation Bulldozer. There are improvements, but on the older outlines, AMD managed to increase optimization by adding the support for new instructions. Concretely, we're talking about FMA4/3, AVX, AES, XOP, etc. In this area alone, the performances in certain situation will be better than before, regardless of the CPU frequency. The processor unit consists of two modules, where each acts as a separate dual-core x86 units. Bearing in mind that there's also 2 MB of L2 cache memory per module (4 MB in total), it can be said that we can't expect too much differences in CPU tests when compared to Richland. Still, for GPU is nothing to scoff at. AMD performed the long-waited transition to GCN 2.0 architecture for which we've hoped ever since we went to Richland's presentation in Moscow. Basically, the graphics part shares the same genetics with Radeon R9 290X. The structure is the same, only less complex, of course. GPU consists of eight Cus with 64 Stream processors each. Therefore, in its most powerful incarnation, the Radeon located inside will have a total of 512 Stream processors at its disposal. An impressive amount, especially bearing in mind that this specification coincides with Radeon HD 7750. GPU also has its own L2 cache and 32 texture units. We have to admit that we're a bit disappointed because the current configuration and architectural limitations restrict the number of ROP units to 8. This is actually the only weak link when compared to Radeon HD 7750, or rather R7 240. Geometry Processor has been implemented in a bit more basic form when compared to Hawaii, but despite that, Geometry and Vertex Assembler are present, and even the tessellation unit. All in all, an excellent combination.
|AMD A10-6800K||AMD A10-7850K|
|Clock / Turbo Core||4.1 GHz / 4,4 GHz||3.7 GHz / 4 GHz|
|Cores||4 cores / 2 modules|
|L1 / L2 / L3 cache||4x 16 KB Data + 2x 64 KB instructional L1 / 4 MB / none|
|Default / max multiplier||41 / unlocked||37 / unlocked|
|Supported memory||DDR3 2133||DDR3 2400|
|Technology||32 nm||28 nm|
|TDP||100 W||95 W|
|Integrated graphics||Radeon HD 8670D||Radeon R7 Series|
|Number of SP||384||512|
|Price||130 €||170 €|
The biggest novelties, in theoretical sense, are the HAS technologies which consist of hUMA and hQ, where each serves a different purpose. hUMA, or rather, "heterogeneous Unified Memory Access" enables for GPU and CPU to have access to complete RAM and virtual memory. That means, for the first time, system doesn't allocate a certain portion of RAM and reserves it for GPU. Quite contrary, CPU and GPU have equal levels of access to memory. Additional increase in functionality hUMA is gained by implementing hQ technology, or rather, "heterogeneous Queuing" technology. It allows CPU and GPU modules to create processes independently and to communicate with each other. This maximizes the performances and removes the existing limitations when CPU is over encumbered which directly influence the GPU. There's a lot more, such as Mantle and TrueAudio support, but if we were to list every single detail, this would probably take up half of this month's issue. Therefore, let's get to the product at hand, which is at the same time the first new generation APU that we've tested.
AMD A10 7850K
This is an APU which serves as a flagship of the company's portfolio and as such it's the most powerful and the most expensive. It's interesting that its TDP of 95 W is actually lower than its predecessor. You guessed it, not only is Kavery faster, but it's also more efficient than Richland. The APU itself is compatible with FM2+ platform, so the current motherboards with this socket will be completely compatible. Still, you'll have to wait for a few BIOS revisions in order for everything to work as intended. A10 7850K is, as a full-blooded Kaveri APU, complete in every way. There are four CPU cores, or rather two, 256 KB of first-degree and 4 MB of second-degree cache. Processor works at maximum of 4 GHz in turbo mode, or rather 3.7 GHz, when all cores are at their maximum capacity. It's backed up by Radeon R7 with 512 Stream processors, distributed into 8 GPCs. Even though that's 50% more SP than Richland, AMD had to reduce the clock from 844 to 720 MHz.
During use, A10 7850K showed very similar results when we were performing strictly CPU tests. On the other hand, GPU performances have been noticeably improved thanks to the graphic processor's new unit, but also thanks to its raw power. Also, we can expect large increases when AMD performs further polishing of the drivers, because we all know how things work out when technology is fresh. Utilization is at a high level, both due to applications and the drivers, and in the end, due to OS itself. Also, as soon as HSA and Mantle start evolving and support for them increases, we can expect even greater things from Kaveri.
|CPU Single Core [cb]||94||95||1.05|
|GPU Accelerated Final Score||54.4||82.2||33.82|
|Futremark PCMark 8|
|Creativity Suite [pts]||3352||3876||13.52|
|Futuremark 3DMark 11|
|Futuremark 3DMark - Cloud Gate|
|Futuremark 3DMark - Fire Strike|
|x264 HD Benchmark 5.1|
|First Pass [fps]||34.03||35.51||4.17|
|Second Pass [fps]||7.78||8.52||8.69|
|1600x900 / 1920x1080 0xAA 0xAF|
|Unigine Valley (Medium)||28.2 / 16.1||30.9 / 17.2||8.74 / 6.4|
|Crysis 2 (Low)||17.5 / 12.3||21.2 / 15.4||17.45 / 20.13|
|Resident Evil 5 (High)||50.4 / 40.1||53.2 / 42.7||5.26 / 6.09|
|Bioshock Infinite (Medium)||34.7 / 23.7||37.4 / 25.6||7.22 / 7.42|
|Hitman Absolution (Low)||28.7 / 23.7||31.8 / 25.2||9.75 / 5.95|
|Tomb Raider (Low)||49.6 / 33.9||59.2 / 43.2||16.22 / 21.53|
|Company of Heroes 2 (Low)||22.9 / 18.7||26.3 / 21.7||12.93 / 13.82|
If only it was cheaper
Unfortunately, when a new product appears on the market, its initial price can be more than it was to be expected. In accordance with that, as a product in demand, A10 7850K initial price was almost 190 euros, which is a really steep price for one APU. In a few months it has dropped by 20 euros, which have made it an absolutely ideal choice for many users who want to cover all their bases with one product. Ultimately, it supports Dual-Graphics, so you can always do an extremely cost-effective upgrade by buying Radeon R7 250.