We seen first APU Llano two and half years ago. This APU was AMD 12h family, very similar to 10h AMD family as all Phenoms II/Athlons II stars CPUs. Performance of Llano APUs was not bad, specially iGPU was great, but Phenoms II were better alternative. The FM1 socket had one year longevity.

After this AMD presents 15h family do desktop. New Bulldozer chips with good multithread performance, but not good single thread performance and high power comsumption. Zambezi BD was full of bugs, so AMD worked at new Piledriver cores and edhanced Zambezi in same time. At October 2012 coming new APU, to FM2 socket, AMD Trinity. This APU has similar performance as Llano, but with better iGPU part and with better power consumption with good overclocking ability. Next 10 months and to FM2 coming light refresh with better P-state and higher clocks - AMD Richland. I must to say, I liked A10 Richlands. Awesome OC, good scaling with voltage, still low power consumption after OC. These two 15h APUs are FM2 chips, but working also in FM2+.

Today is the D day from AMD, second generation of 15h APU is here, AMD Kaveri with Steamroller B. Chips are manufactured at 4Q 2013 and today are ready for sale and for reviews of
course :). Whats new here? Kaveri is AMD most advanced APU ever, with very good IPC (instruction per clock). We can talking about new GCN iGPU architecture inside, about new 28nm SHP process, HUMA controller, HSA support, AMD Trueadio, PCIe 3.0 and configurable TDP.

APU die size is 245mm, almost the same as in Richland/Trinity. But there is 2.41 bilion tranzisotrs! This is almost double more than before! This is because SHP process, higher
density at die size. SHP is "mixed" bulk and SOI. So good for iGPU, good for CPU tranzistors, but worse for the highest clock. Thats why Kaveri clocks are -400 MHz lower than with Richland.


Configurable TDP is one interesting part for notebooks segment. But working very well in desktop (we can know it from Haswell). There is slide from 65W and 45W part and performance in 3D Mark Firestrike with iGPU. Richland 65w is A8-6500, 45W is A8-6700T. The drop in score is around 250 points. But what Kaveri? Kaveri 65W is A8-7600. We can change TDP of CPU in BIOS for 45W at the same model. And performance? Wow, only few points lower than before!

Diagram of Steamroller B Kaveri. CPU cores with shared L2, IO part, iGPU part with up to 8 units and HUMA controller at left.

New chipset A88X is almost the same as A85X for FM2. One difference is PCIe 3.0, if we using Kaveri chip, and AHCI update. I have only A85x screenshot for representation...A78 and A55 are next chipsets of FM2+.

Kaveri including up to 12 compute cores. Why? There is (in 7850K) 4 CPU cores and 8 GPU compute cores of GCN. So programmable for parallel compute. Simply pure CPU performance clock to clock is up to 20% better than Piledriver core! Wow. How about improving seems AMD! But of course, the clocks of Kaveri are -400 MHz lower and after all is Kaveri slighty better than Richland A10-6800K (around 5% average at CPU part).
GPU part is very good. The gain is similar as from Llano to Trinity. But there is GCN, DX11.2+Mantle and new TrueAudio. The top R7 iGPU has 8 gpu units with 512 stream processors, 32 texture units and 16 ROPs.

The main stone of Kaveri is again one module. This module has two cores with shared L2 cache 2MB, few pipelines and some specialized units (128 FMAC and MMX). L1D cache is now higher - 96KB and 3-way. Dont forget about Kaveri isntruction set - FMA4/3, AVX, AES, XOP.

iGPU part is 47% of all die size. Again GCN concept as Hawai example working with DDR3 2133 MHz native (but it working with 2666 MHz OK at kaveri). iGPU part support Mantle, DX11.2, AMD Trueaudio, Eyfinity or acceleration UVD and working VCE. GPU part and CPU part can working together for compute - heterogeneous. So GPU and CPU have equal flexibility to create work. HSA compute are writing in opencl or C++ or few next language. This languages named as HSAL. HSAL is supporting now by few programs, example Photoshop, Libre
Office, Gimp etc. HSA are there great for saved time here. There is few examples of HSA.




Now we are back to pure iGPU performance in gaming. Finally some modern games at FullHD without lagging!

Or popular LOL game

Kaveri seems strong with dedicated card too, Im sure, better than Piledriver.

iGPU features as Mantle and TrueAudio are next benifits.
Whats Mantle exactly? Its new alternative to DX. But with benefits from GCN architecture for higher FPS (up to 40% boost!). Simply we can play games with worse cards than before. Next good idea is, you dont need Windows 8 for latest gaming effects! This will be working in theory with Windows XP too! At least with Windows 7 at 100%.AMD Trueaudio is less stress for CPU and better sound quality in games. Sound was limit of modern games. Now not must be. In iGPU parts are few audio coprocessors. And again programeable :).
Who knows me, knows me ;)....AMD 3000+, AMD x2 4600+ EE, AMD X4 955 BE C2,2x AMD X4 965 BE C3, AMD X4 970 BE C3, AMD x4 975 BE, AMD x4 980 BE, AMD X6 1090T BE, AMD x6 1100T BE, 2x AMD FX-8120, 2x AMD FX-8150, FX-6300, FX-8300, FX-8320E, FX-8320, FX-8350, FX-8370, FX-8370E, FX-9370, FX-9590, AMD A8-3850, AMD A8-3870K, A8-5600K, A10-5800K, A10-6800K, A10-7850K, A10-7870K, A 5150, Athlon x4 860K, Intel i7-5960X, i7-6700K, Intel i7-4770K, Intel i7-980x, Intel i7 2600k, Intel i7-3770K, i7-3930K.