Tuesday, January 14th 2014

AMD Announces 4th Generation A-Series "Kaveri" Desktop APUs

AMD announced its 2014 A-Series APU for the desktop platform, code-named "Kaveri," after the southern-Indian river. Built in the new FM2+ package, the APUs run only on socket FM2+ motherboards based on the AMD A88X, A78, and A55 chipsets; while the socket itself can seat older FM2 APU families, "Trinity" and "Richland." In many ways, the socket transition is similar to that of socket AM3+. "Kaveri" sees AMD integrate two of its newest CPU and GPU micro-architectures, "Steamroller" for CPU, and Graphics CoreNext 2.0 for the GPU. "Kaveri" is also built on newer generation 28 nm silicon fab process.

"Steamroller" is an evolution of the same modular CPU core design as its predecessors, "Piledriver" and "Bulldozer." AMD promises a 10 percent improvement in performance clock-by-clock, per core, which falls in line with AMD's normal scheme of annual incremental performance updates on its CPU micro-architectures. A "Steamroller" module is a combination of two 64-bit x86 cores, which feature dedicated and shared components. "Kaveri" has two such modules, and so physically, it features a quad-core CPU.

The GPU inside A-Series "Kaveri" APUs is the fastest AMD ever crammed into a PC APU, although it doesn't come close to the APUs that drive Xbox One and PlayStation 4. Based on the Graphics CoreNext 2.0 micro-architecture, "Kaveri" features eight GCN compute units (CUs), which make up 512 stream processors, and 32 TMUs. The GPU supports DirectX 11.2, OpenGL 4.3, and Mantle. The GPU component also accelerates AMD's TrueAudio technology, which debuted with the Radeon R9 290 series. The GPU's display I/O lets you drive Ultra HD (3840 x 2160 pixels) displays with 60 Hz refresh rates, over DisplayPort 1.2. It supports up to four display outputs.

Moving on to the uncore portion of "Kaveri," AMD deployed a new-generation integrated memory controller (IMC). It supports up to 64 GB of dual-channel DDR3-2400 MHz memory. Its biggest feature id hUMA (heterogeneous unified memory access), which not only chucks out fixed memory partitions between the system and shared graphics memory area, but also makes the CPU and GPU access the same memory simultaneously. This should translate to greater system memory available to the OS (as "hardware reserved" GPU memory is eliminated), and better GPGPU performance. The second big highlight in the uncore department is the PCI-Express root complex, which now gives out up to 24 lanes of PCI-Express gen 3.0. This should translate into better CrossFire performance, in which a pair of graphics cards are given an 8-lane PCI-Express link, each.

AMD launched its 2014 A-Series APU family with three models, the A10-7850K, the A10-7700K, and the A8-7600. Specifications of the three parts are tabled below. The A10-7850K and A10-7700K will include Origin keys to Battlefield 4, and the two chips apparently meet the minimum system requirements of the game.
Add your own comment

50 Comments on AMD Announces 4th Generation A-Series "Kaveri" Desktop APUs

#1
ZetZet
Can't find any reviews yet :(
Posted on Reply
#2
RCoon
Gaming Moderator
by: ZetZet
Can't find any reviews yet :(
Well it hasn't been released here yet, OCUK are usually the first to get AMD gear in stock. Paper launch? Also If it's anything like Richland's release, the NDA will be absolutely and completely retarded. I wouldn't expect reviews for a few days/a week. I also wouldn't get your hopes up! :laugh:
Posted on Reply
#3
john_
7600 is a really nice choice for anyone looking for a cheap desktop. 7850 and 7700 are a little expensive and in Intel territory, but I am sure they will get in many pre built PCs and especially slim All in One all PCs or tiny boxes with no room for a discrete card.
Now if only we could get a Steamroller CPU(no graphics) on FM2+ and why not, maybe with more than 2 modules.
Posted on Reply
#5
NC37
So same weak dual core acting as a quad core design and now with GPU cores posing as CPU cores? Well, I'll be interested in seeing how it does with handbrake encodes. But I guess if you can't pack them in, you gotta figure out another way to add more. Still no L3 though.

Funny to think we've come full circle. From the GPU makers developing the 3D drivers/etc to just the OS devs handling it. Now AMD is back to it with Mantle. Wonder how long till nVidia does the same. Course if it means M$ won't hold a monopoly over drivers and force people to update to new Windows to get new DX capabilities then that is good. Time will tell.
Posted on Reply
#6
ZetZet
I just don't understand why they don't shrink old phenom to 28nm and go with it, phenoms already performed quite well, better than these module type designs atleast.
Posted on Reply
#7
john_
by: NC37
Wonder how long till nVidia does the same.
If Maxwell is a strong chip, Nvidia doesn't have to follow AMD in creating it's own API, only give a better DirectX performance at the same price points. If for example in a year from now AMD is winning with 20% in some titles because of Mantle and Nvidia is winning by 10% in all other titles that wouldn't support Mantle, things wouldn't be much different than today, just much more complicated(or simple depending how you see it) when you are buying a new graphics card.
Posted on Reply
#8
RCoon
Gaming Moderator
by: ZetZet
I just don't understand why they don't shrink old phenom to 28nm and go with it, phenoms already performed quite well, better than these module type designs atleast.
I wholeheartedly agree, but perhaps Deneb and Thuban aren't that scalable. It all sounds good to us, but it depends on how effectively the Phenom architectures scale down. I'd take a genuine dual, quad or hex core from AMD derived from the Phenom days. I still have a 1055T laying around here somewhere. I also have an FM2 750K processor sat on the shelf, which will never be used for anything ever again.
Posted on Reply
#9
buildzoid
by: ZetZet
I just don't understand why they don't shrink old phenom to 28nm and go with it, phenoms already performed quite well, better than these module type designs atleast.
Only in multi core scenarios because what AMD bulldozer design does wrong is that it doesn't have per core scaling. If only one core is used the core will get 1.23 in Cinbench but the moment you have 2 cores in one module working they only got 1.7 points because of them sharing cache and other resources. If AMD made a quad core that had one core in each module disabled it would give 1.23x4 but they make quad cores by disabling entire modules so you end up with 1.23x2.5. So a Phenom II is basically the same in single core performance and has an advantage when in come to multi threading scaling but this is solved by AMD simply having more cores than before.
Posted on Reply
#10
Kärlekstrollet
by: ZetZet
I just don't understand why they don't shrink old phenom to 28nm and go with it, phenoms already performed quite well, better than these module type designs atleast.
Phenom reached its peak after many years of improving the same core(K* architecture) so if AMD kept using K* today the gap between Intel would have just been bigger.

AMD should've started the Bulldozer project 2 years earlier but there is no time machine in sight yet.
Posted on Reply
#11
bogami
New but only 28nm. GPU standard for current GPU units, it seems that better is not in AMD. Regardless of the forecast do not expect anything better than to before .3 gene. GPU support is not on the current motherboards. only a single ASUS model has the support and here I do not know if it is aligned with the new processor and the ability of 856 GFLOP we have the best model is only less than 1/6 of Haiti GPU 5.9TFLOP so play crysis 3 with pleasure is off. In order to obtain a solid results from the CPU was required frequency as 3.7-4 Gh. What will AMD do when the 14 nm Hasvell in 6 months appeared. Can just pack their suitcases or is some ace under the sleeve !
Posted on Reply
#13
ZetZet
by: buildzoid
Only in multi core scenarios because what AMD bulldozer design does wrong is that it doesn't have per core scaling. If only one core is used the core will get 1.23 in Cinbench but the moment you have 2 cores in one module working they only got 1.7 points because of them sharing cache and other resources. If AMD made a quad core that had one core in each module disabled it would give 1.23x4 but they make quad cores by disabling entire modules so you end up with 1.23x2.5. So a Phenom II is basically the same in single core performance and has an advantage when in come to multi threading scaling but this is solved by AMD simply having more cores than before.
But that's the thing, they don't have more cores, only fx-83xx have more cores and are viable, while fx-6300 is used for budget gaming and Fx-4xxx and apu's are trash.
Phenom's had 6 real cores and did fine with that.
Posted on Reply
#14
Mathragh
I'm wondering why the rest of the internet is totally quiet regarding this launch. Anyone with a link to another site posting similar news/reviews?
Or is it still too early.
Posted on Reply
#15
john_
by: RCoon
I also have an FM2 750K processor sat on the shelf, which will never be used for anything ever again.
That bad?
Posted on Reply
#16
ZetZet
by: john_
That bad?
Cheaper i3 and I don't know why would anyone buy something less than i5.
Posted on Reply
#17
john_
by: ZetZet
Cheaper i3 and I don't know why would anyone buy something less than i5.
750K is way cheaper than i3. Add to that that you can have a motherboard with about everything on it at much better prices than any Intel board and you have your answer.
Posted on Reply
#18
kn00tcn
by: Kärlekstrollet
AMD should've started the Bulldozer project 2 years earlier but there is no time machine in sight yet.
wasnt it 2 years late already? :laugh: they probably started on time, just didnt finish

has gloflo even been ready for mass large high end 28nm?
Posted on Reply
#19
buildzoid
by: ZetZet
But that's the thing, they don't have more cores, only fx-83xx have more cores and are viable, while fx-6300 is used for budget gaming and Fx-4xxx and apu's are trash.
Phenom's had 6 real cores and did fine with that.
Yes but a Phenom IIs single threaded performance at 4.1Ghz is barely on par with the single core performance of a stock A10 6800K and the A10 7850K is even faster and if they improved the resource sharing system the A10 7850K and it's Athlon spinoff will be absolutely perfect for budget gaming.
Posted on Reply
#20
RCoon
Gaming Moderator
by: john_
That bad?
Not at all, I had a 750K overclocked to 4.4Ghz on my current rig with a 780. I could play any game I wanted with no trouble on 1440p. But I do a great deal of parsing, RARing and encoding on some evenings, so when the A85X motherboard broke, I switched to an i5.

I laugh at people hating on these little things, because they've never used them. Played BF3 and all those games on a 750K, and when paired with a high end GPU, I could eat games alive on 1440p. Guess people just don't know what they're talking about...

EDIT: Granted I don't recommend these to anyone playing Total War though!
Posted on Reply
#22
RCoon
Gaming Moderator
by: darkangel0504
Power consumption








What's with the comparison without Richland? That was a minor improvement upon the old 5800K, so that benchmark seems a little poorly thought out.
Nice to see improvements in power consumption though...
Posted on Reply
#23
west7
i think this new apu is great deal for everyone dont wont to spend a 1000$ on pc especially those ho lives outside the US like me with prices of pc parts is to expensive i rather buy an apu then spending alot more for cpu+gfx crd combo for the same money
Posted on Reply
#24
Roph
Now delete the weak GPU from the chip, fill the empty space with more steamroller cores, and package it for AM3+ please.
Posted on Reply
Add your own comment