AMD Kaveri Review: A8-7600 and A10-7850K Tested
by Ian Cutress & Rahul Garg on January 14, 2014 8:00 AM ESTLlano, Trinity and Kaveri Die: Compared
AMD sent along a high res shot of Kaveri's die. Armed with the same from the previous two generations, we can get a decent idea of the progression of AMD's APUs:
Llano, K10 Quad Core
Trinity and Richland Die, with two Piledriver modules and processor graphics
Kaveri, two modules and processor graphics
Moving from Llano to Trinity, we have the reduction from a fully-fledged quad core system to the dual module layout AMD is keeping with its APU range. Moving from Richland to Kaveri is actually a bigger step than one might imagine:
| AMD APU Details | ||||
| Core Name | Llano | Trinity | Richland | Kaveri |
| Microarch | K10 | Piledriver | Piledriver | Steamroller |
| CPU Example | A8-3850 | A10-5800K | A10-6800K | A10-7850K |
| Threads | 4 | 4 | 4 | 4 |
| Cores | 4 | 2 | 2 | 2 |
| GPU | HD 6550 | HD 7660D | HD 8670D | R7 |
| GPU Arch | VLIW5 | VLIW4 | VLIW4 | GCN 1.1 |
| GPU Cores | 400 | 384 | 384 | 512 |
| Die size / mm2 | 228 | 246 | 246 | 245 |
| Transistors | 1.178 B | 1.303 B | 1.303 B | 2.41 B |
| Power | 100W | 100W | 100W | 95W |
| CPU MHz | 2900 | 3800 | 4100 | 3700 |
| CPU Turbo | N/A | 4200 | 4400 | 4000 |
| L1 Cache |
256KB C$ 256KB D$ |
128KB C$ 64KB D$ |
128KB C$ 64KB D$ |
192KB C$ 64KB D$ |
| L2 Cache | 4 x 1MB | 2 x 2 MB | 2 x 2 MB | 2 x 2 MB |
| Node | 32nm SOI | 32nm SOI | 32nm SOI | 28nm SHP |
| Memory | DDR-1866 | DDR-1866 | DDR-2133 | DDR-2133 |
Looking back at Llano and Trinity/Richland, it's very clear that AMD's APUs on GF's 32nm SOI process had a real issue with transistor density. The table below attempts to put everything in perspective but keep in mind that, outside of Intel, no one does a good job of documenting how they are counting (estimating) transistors. My only hope is AMD's transistor counting methods are consistent across CPU and GPU, although that alone may be wishful thinking:
| Transistor Density Comparison | ||||||||
| Manufacturing Process | Transistor Count | Die Size | Transistors per mm2 | |||||
| AMD Kaveri | GF 28nm SHP | 2.41B | 245 mm2 | 9.837M | ||||
| AMD Richland | GF 32nm SOI | 1.30B | 246 mm2 | 5.285M | ||||
| AMD Llano | GF 32nm SOI | 1.178B | 228 mm2 | 5.166M | ||||
| AMD Bonaire (R7 260X) | TSMC 28nm | 2.08B | 160 mm2 | 13.000M | ||||
| AMD Pitcairn (R7 270/270X) | TSMC 28nm | 2.80B | 212 mm2 | 13.209M | ||||
| AMD Vishera (FX-8350) | GF 32nm SOI | 1.2B | 315 mm2 | 3.810M | ||||
| Intel Haswell 4C (GT2) | Intel 22nm | 1.40B | 177 mm2 | 7.910M | ||||
| NVIDIA GK106 (GTX 660) | TSMC 28nm | 2.54B | 214 mm2 | 11.869M | ||||
If AMD is indeed counting the same way across APUs/GPUs, the move to Kaveri doesn't look all that extreme but rather a good point in between previous APUs and other AMD GCN GPUs. Compared to standalone CPU architectures from AMD, it's clear that the APUs are far more dense thanks to big portions of their die being occupied by a GPU.
Facebook
Twitter
RSS
380 Comments
View All Comments
Nagorak - Wednesday, January 15, 2014 - link
The fact is these "APUs" are really pretty worthless. If you're going to be seriously gaming then you want to get a discrete graphics card, even a low end one. The problem with AMD's strategy here is that the integrated GPU performance is still so anemic that it's only useful for people who want a budget PC and who don't actually play games. Therefore the "benefits" of having both CPU and GPU on one core are highly questionable.ImSpartacus - Thursday, January 16, 2014 - link
That could be solved if we had a much larger GPU on the APU, right?I want to see at least 12 CUs on the top-end APU, preferably more like 16-20 CUs.
lmcd - Monday, January 20, 2014 - link
Did you miss the whole conversation about memory bandwidth?nader_21007 - Saturday, January 18, 2014 - link
below the charts showing the games performanc, there are two tabs, one for showing avarage frame rate and the other, if you just click on it will show you the suddeframe rates for the apus compared. Minimum frame rate of the iris pro is the lowest and worst among all Apu's compared. It means sudden hangings all over the game. Kaveri and other amd apu's have no such weaknesses because of high performing IGP.Now tell me which is Outperformed?themeinme75 - Saturday, January 18, 2014 - link
the good news is APU-7850/SOC-5200pro just caught up to 4850..cryptik - Monday, March 30, 2015 - link
The R7 consistently beats the Iris Pro. You must have read a different article than everyone else.frozentundra123456 - Tuesday, January 14, 2014 - link
Good review, except he doesn't really address the elephant in the room that even for gaming, a low end cpu like the athlon X4 with a HD7750 will be considerably faster than any APU. So in this regard, I disagree with the conclusions that for low end gaming kaveri is the best solution. It is disappointing actually, that he did not use a HD7750 GDDR5 as the discrete gpu comparison, because that would have given a more direct comparison of how the bandwidth restrictions are affecting Kaveri.I will say though that the low TDP parts seem to get a nice improvement in performance. They actually seem more attractive that the high end, since there is little gaming improvement in the top end vs Richland.
thevoiceofreason - Tuesday, January 14, 2014 - link
Or something like Pentium G2120 (90$ newegg) and HD7770 (90$ ASUS or MSI on newegg after rebate). The difference between such a setup and a top of the line kaveri chip is not even funny. Kind of tragic, really.andrewaggb - Tuesday, January 14, 2014 - link
I agree. I am extremely dissappointed. Cpu performance and GPU performance barely changed from last year's model. HSA is definitely the future, and Mantle may deliver a healthy fps improvement, but I feel like this product has almost no value at launch. It's all depending on future software.I was actually considering a kaveri for a kids pc... but I'm definitely way better off getting an intel cpu and amd gpu.
andrewaggb - Tuesday, January 14, 2014 - link
I guess they did make big gains on TDP. So it's sorta their version of a haswell refresh. But haswell had a larger improvement in both cpu and gpu performance relative to their previous chip.