AMD Kaveri APU – impressive graphics power detailed

AMD Richland header

AMD has revealed some detailed specifications of Kaveri. It’s AMD’s next APU on socket FM2+ and is the culmination of the work AMD has put into HSA (Heterogeneous System Architecture).

AMD announced that it is on track to ship Kaveri to its channel partners before the end of 2013. More details on the APU will be available at CES 2014, with general availability to the public in February 2014. Pricing and exact availability dates were not announced by the company.

Kaveri is based on AMD’s Steamroller architecture, a tweak of Piledriver which was itself a tweak on Bulldozer. Steamroller retains the building blocks of its predecessors – cores are bundled in pairs and made to share hardware like floating point units and cache. The difference with Kaveri compared to Piledriver is that almost everything has been doubled.

AMD Kaveri specification

AMD Kaveri specification

With Kaveri we get more L2 cache, better thread management, a reworked memory controller supporting up to 32GB of DDR3 memory, better decode units and two FMAC decode units that are used for accelerating AVX code. AVX is a coding scheme that allows for more parallelism and is suitable for scalable compute scenarios like scientific and financial prediction programs.

All in all, CPU performance should be above the level of the AMD Richland A10-6800K and on a similar level to Intel’s Sandy Bridge architecture.

But that’s not the biggest part of it. Kaveri dedicates nearly half of the APU die (47%) to the integrated GPU. It’s based on the GCN 1.1 hardware found in the Radeon R7 260X, HD7990 and R9 290 series and features the same power optimisations found in these GPUs.

Inside Kaveri, AMD has slotted in eight SMX units. At 64 shader cores per unit, that puts it at 512 shader cores, which is up 128 shader cores over the A10-6800K. In terms of discrete GPU performance, it’s about the same as a Radeon HD7750, which was previously the fastest discrete GPU in AMD’s lineup that didn’t require an external power source.

AMD Kaveri projected performance

AMD Kaveri projected performance

Kaveri will weigh in at around 856 GFLOPS of theoretical performance with stock clocks at 3.7GHz for the CPU and 720MHz for the GPU. There will likely be enough room for some overclocking above that, making it theoretically possible to have a Kaveri APU break the 1TFLOPs barrier.

To put that in perspective,Β the A10-6800K sits at 779 GFLOPS, the Radeon HD7750 is at 819 GFLOPS, Kaveri at 856 GFLOPS, and the Xbox One is rated for 1,230 GFLOPS.

We’re fast approaching console territory with these APUs and that is good news for cash-strapped gamers.

AMD even showed off a demo running Battlefield 4 on a system with the top-end Kaveri chip and a system with an Intel Core i7-4770K and a Nvidia Geforce GT630. The demo itself wasn’t particularly taxing, but the Intel system was stuttering along at 12-14fps while the Kaveri system easily posted an average above 25fps, and both systems were at 1080p with Medium settings.

AMD TrueAudio initial developer support

AMD TrueAudio initial developer support

Lastly, AMD is also rolling in new technologies like AMD TrueAudio, Mantle, the complete HSA architecture, and PowerTune into the mix. Many people were sceptical of any advantage TrueAudio could give to AMD, but if all of their low-end chips possess the technology it makes it easier for the tech to gain critical mass and attract support from developers.

AMD continues their APU13 conference for the remainder of the week and will have a first look at Mantle, presented by DICE Engineer Johan Andersson, on Thursday 14 November.

Source: Anandtech, PC Perspective

More Hardware news:

Nvidia Geforce GTX780 Ti review roundup

AMD socket FM2+ will unify processor families

Valve reveals Steam Machine prototype hardware

AMD Radeon R9 290, R9 290X SA price revealed

Forum discussion

Join the conversation

  • NicoR

    Haha Wesley, its funny how you bold the important and convincing details. πŸ™‚

    This really hope this Kaveri is at least better than your average i5. I get sad when seeing i3s’ match up against the 6800K…

    I’d like to see the performance of the new DDR3 controller. Are their serious differences between GNC 1.0 and 1.1?

    I was going through an AM3+ upgrade path, but it seems silly now. The local FM2+ selection of motherboards is poor though, but the best overclocking bang for buck mobo I’ve seen so far is the Asrock board at Wootware.

  • I don’t bold them, my editor does πŸ™‚

    I’m also interested in what they end up doing with the DDR3 controller. A 512-core GPU would be hamstrung by memory bandwidth and not everyone buys DDR3-1866 memory with their APU systems.

    Between GCN 1.0 and 1.1, there’s mostly just power improvements and changes to how the GPU steps between different power phases. There are some clock speed improvements and tighter packing of the hardware units, but overall it’s pretty much a standard HD7750 shoved onto there. And TrueAudio, that’s also added in.

    I’m also disappointed in FM2+ selection, but that will hopefully improve later on. Gigabyte, MSI and ASUS all have boards waiting in the wings, ASRock is the only brand that ships anything here for now. Once reviews and performance indications for Kaveri come in, then we’ll start seeing some movement.

  • NicoR

    Haha at your editor then πŸ˜‰

    Have you perhaps seen the blog entry on Corsair’s website? It’s regarding the huge difference in FPS in BF4 when using higher clocked memory.

    The HD7750 as a built on solution is fantastic. Its a good start at purchasing a Kaveri gaming setup, without the initial high cost for a discreet graphics card.

    Check out this motherboard on wootware, its a really good deal.

    After reading up some more on GCN and HSA, it can be a little confusing. HSA sounds like it benefits from software that takes advantage of the more efficient memory allocation model between CPU and GPU, where GCN sounds like a collection of technologies that improves overall GPU speed and efficiency.

    Is HSA part of GCN? Are 77xx-7900x / R7 – R9 graphics cards taking full advantage of GCN right now? Is a Kaveri APU required to enable the HSA component?

  • I did see that a while back, but I’m waiting for a proper test with the final game instead of the beta. Aside from the APUs, I’m intrigued that Intel systems appear to benefit from it as well.

    That motherboard’s nice, but it’s not FM2+! When I upgrade next year, I’m probably aiming for something similar.

    HSA is a collection of software and hardware tech that blurs the line between the GPU and CPU, along with their split memory pools. It pretty much allows developers to take better advantage of the hardware in our rigs and give us better performance.

    GCN is the architecture for AMD’s GPUs and it’s designed to run code in parallel, more so than their previous designs. In fact, GCN and Kepler are incredibly similar as both Nvidia and AMD can see that we’re finally moving into the era where software is far more paralell in nature. Instead of fighting Moore’s law constantly, we’re now attacking Amdahl’s law.

    GCN is required to take advantage of HSA on AMD’s systems, although there are many partners in the HSA foundation that offer their own hardware (Samsung, Apple, MediaTek, AMD to name a few). So HSA will apply to HD7000/R-200 GPUs as well as Kaveri and any hardware based on GCN and its future derivatives.

  • David Stanley


AMD Kaveri APU – impressive graphics power detailed

Related posts