News
 

Bookmark and Share

(97) 

At the Consumer Electronics Show event Advanced Micro Drives has publicized some of the details behind the code-named “Kaveri” accelerated processing unit (APU). The firm revealed that it had to make rather tough design decisions when developing the new hybrid chip and had to choose a manufacturing technology that had been optimized for highly-integrated chips.

For previous generations of its high-performance APUs (i.e., Llano, Trinity and Richland), AMD chose 32nm silicon-on-insulator process technology that was developed by the company itself and GlobalFoundries with high-performance microprocessor units in mind. Those process technologies allowed AMD to ensure high clock-speeds for its x86 cores, such as Bulldozer and Piledriver. Unfortunately, that manufacturing technology was not exactly tailored for high transistor density of graphics processing units. As a consequence, AMD had to make certain compromises, which did not allow it to integrate all the hardware it wanted.

With Kaveri, AMD chose to use GlobalFoundries’ 28SHP (28nm super high performance) process technology that was designed with various types of chips in mind. Improved transistor density and thinner elements of 28nm process tech allowed the company to integrate whopping 2.41 billion of transistors into a 245mm2 Kaveri die, which is 85% higher transistor count compared to Trinity/Richland design (246mm2, 1.303 billion transistors, 32nm SOI). Now the company has four Steamroller cores, hUMA [heterogeneous unified memory architecture] memory controller, Radeon graphics processing unit with 512 stream processors (AMD calls them eight GPU compute units) and a lot of various special-purpose hardware inside its flagship APU.

The high-density “universal” process technology has its drawbacks too: it lacks thick metal stacks that make possible to run microprocessors at high frequencies. As a result, the company sacrificed extreme clock-rates for x86 general-purpose processing cores in favour or forward-looking APU design with a strong focus on heterogeneous computing.

While overclockers and performance enthusiasts may not appreciate AMD’s choice, the company claims in its documents that thanks to more advanced Steamroller micro-architecture (with up to 20% higher instruction per clock [IPC] performance), the Kaveri will not be slower than currently-available Richland APUs and will even beat them in many cases.

AMD believes that thanks to scalability of its Kaveri design (which is supposed to be available in 15W – 95W thermal design power envelopes), it will be able to successfully address various personal computers, including desktops, notebooks and servers, with its new A-series accelerated processing units.

Tags: AMD, Kaveri, Radeon, GCN, Steamroller, Fusion, 28nm, Globalfoundries

Discussion

Comments currently: 97
Discussion started: 01/09/14 12:00:05 AM
Latest comment: 01/14/14 11:24:49 PM
Expand all threads | Collapse all threads

[1-12]

1. 
One question from an old fart is it 2 pairs or 4 separate cores in the CPU side and is the same true for the graphic side.,so does this mean 2 pairs plus 4 pairs giving a total of 8 pairs? or 4 x 8 giving a total of 32 core or is 4+8 giving it 12 cores in total. I get confused easily now with the different way co. count cores, just like the way they size they have different sizing in their architecture ( Viz. Intel v Global).
2 0 [Posted by: tedstoy  | Date: 01/09/14 12:00:05 AM]
Reply
- collapse thread

 
Its 4CPU Cores + 8 GPU Core = 12 Compute Cores.

12 Compute Cores i think thats the new Marketing Name Strategy by AMD.
3 0 [Posted by: xentar  | Date: 01/09/14 12:18:26 AM]
Reply
 
Thanks
3 0 [Posted by: tedstoy  | Date: 01/09/14 12:26:07 AM]
Reply
 
show the post
2 12 [Posted by: amdzorz  | Date: 01/09/14 10:29:39 AM]
Reply
 
Yes and no.
There are 2 modules with 4 cores (on the left side of the schematic overview on the slide).
The 8 GPU "cores" the basic GCN compute units containing 4 wide SIMD units and one scalar unit.
2 1 [Posted by: Dresdenboy  | Date: 01/09/14 02:02:43 PM]
Reply

2. 
That would explain why AMD got off the FX wagon. If Steamroller is designed to be manufactured on a graphics process, creating a CPU with no graphics with it would make less sense.
8 2 [Posted by: ET3D  | Date: 01/09/14 02:50:58 AM]
Reply
- collapse thread

 
Intel have 2011 socket CPUs without graphic, and this make lot sense, thanks to superior CPU. But AMD... With reduced clock-speed and Richland level perf, you have zero improvements, and no any sense to build next FX. There is not exist any 'graphics process'
5 1 [Posted by: Tristan  | Date: 01/09/14 09:06:19 AM]
Reply
 
GPUs are fundamentally more efficient at some types of compute processing. So it is not about how much you can perfect your CPU it is about how well you can balance your allocation of transistors for different types of compute processing. The purpose of the huma initiative is to eventually reach a point where the APU can utilise its compute resources effectively with less or little or even no burden, on software developers.
AMD, ARM, et al, are all rolling in the same direction, because it is ultimately more compute and more power efficient. Intel currently have a CPU and process advantage which they leverage very effectively. But even Intel will have to jump on the APU/HSA/HUMA train eventually.
0 0 [Posted by: mjv1121  | Date: 01/10/14 04:51:55 AM]
Reply
 
Additionally if every chip they make is an APU, it creates a larger need for software developers to optimize for them. I would love to replace my 8350 with an 8 steamroller core non-apu chip though.
1 1 [Posted by: KeyBoardG  | Date: 01/09/14 10:21:50 AM]
Reply
 
show the post
2 10 [Posted by: amdzorz  | Date: 01/09/14 10:31:10 AM]
Reply
 
I didn't ask for an 8core apu. I said just a newer 8350 with steamroller cores. Although, that isn't likely to happen either.
0 1 [Posted by: KeyBoardG  | Date: 01/09/14 12:07:02 PM]
Reply
 
Of course there will be an 8 core APU once AMD gets down to 14nm. By 7-10nm, they will have a 12 core APU.
1 2 [Posted by: BestJinjo  | Date: 01/09/14 01:01:02 PM]
Reply
 
show the post
2 9 [Posted by: amdzorz  | Date: 01/09/14 01:34:51 PM]
Reply
 
Did you not read the article?

"28nm process tech allowed the company to integrate whopping 2.41 billion of transistors into a 245mm2 Kaveri die, which is 85% higher transistor count compared to Trinity/Richland design (246mm2, 1.303 billion transistors, 32nm SOI)."

They nearly doubled the transistor count going from 32nm to 28nm. By the time AMD is on 14nm, they can easily make an 8-core APU with real cores.
2 3 [Posted by: BestJinjo  | Date: 01/10/14 12:15:03 AM]
Reply
 
Perhaps AMD will have an 8 core APU at some point, but that won't do much good for people who want to upgrade their AM3+ rig. I like AMD, but I'm not a die-hard, and if I have to replace my motherboard I might as well go for Intel.

Besides, it will be a while before AMD gets down to 14nm, and so far it looks like the strategy (which is quite reasonable) is to dedicate the extra transistors to GPU power.
0 0 [Posted by: ET3D  | Date: 01/12/14 04:18:45 AM]
Reply

3. 
sounds to me AMD preparing for bad reviews regarding CPU performance.
7 6 [Posted by: medo  | Date: 01/09/14 05:40:52 AM]
Reply
- collapse thread

 
No, AMD is preparing OEMs for good news... that making motherboards for AMD processors is easier and cheaper because they integrate more functionality than Intel. These APUs are destined for the biggest markets in Asia where AMD holds 40% market share and are 5 times bigger in dollar terms (currently) than North America. That's right!
8 6 [Posted by: linuxlowdown  | Date: 01/09/14 07:02:17 AM]
Reply
 
show the post
2 15 [Posted by: amdzorz  | Date: 01/09/14 10:34:16 AM]
Reply
 
3 2 [Posted by: linuxlowdown  | Date: 01/10/14 02:09:21 AM]
Reply

4. 
not about cpu performance alone anymore it's about apu's and power consumption. this is where the market is now this is why intel has slowed down on the cpu performance in recent years and have focused more on the entire structure of apu's and their power consumption. all in one chips with low power consumption is all the rage now.
5 4 [Posted by: SteelCity1981  | Date: 01/09/14 06:02:52 AM]
Reply

5. 
show the post
4 9 [Posted by: Tristan  | Date: 01/09/14 06:16:07 AM]
Reply
- collapse thread

 
You didn't read the article now, did you boy. That makes you the class clown for flunking the comprehension test. Your mark is FF for Fanboy Fail.
9 5 [Posted by: linuxlowdown  | Date: 01/09/14 06:41:13 AM]
Reply
 
I fully understand your desperation. You cannot point any serious AMD advantage over Intel or NV, so last possibility is to express angry behaviour on anybody who is talking about sad facts rerlated to AMD.
Better return to AMD slideware, and study another 100 times their bright future. Hope this help.
5 3 [Posted by: Tristan  | Date: 01/09/14 09:11:56 AM]
Reply
 
1 serious advantage over NV? AMD's cards pay for themselves far quicker with litecoin mining and for the nearly half a decade paid for themselves fully with bitcoin mining.
2 2 [Posted by: BestJinjo  | Date: 01/10/14 12:22:08 AM]
Reply
 
show the post
5 8 [Posted by: amdzorz  | Date: 01/09/14 10:36:04 AM]
Reply
 
show the post
2 8 [Posted by: amdzorz  | Date: 01/09/14 04:14:30 PM]
Reply
 
Nope. 4.1ghz base clock vs. 3.7ghz is not enough to offset IPC gains of up to 20%. Hence SR is still faster despite slower clocks:

http://www.anandtech.com/...eri-prelaunch-information
1 1 [Posted by: BestJinjo  | Date: 01/11/14 07:29:14 AM]
Reply
 
show the post
2 5 [Posted by: Eldi Kaiser  | Date: 01/09/14 05:45:58 PM]
Reply

6. 
The chip is a pretty good alternative to the 4th gen Core i3 CPU for sure. AMD Kaveri got more power from having a quad core CPU and an incomparable level performance from the GPU.

Gone are the days of users of computer with a Pentium or Core i3 + GT640, Radeon HD 6670 graphics cards.......
2 2 [Posted by: tks  | Date: 01/09/14 07:54:57 AM]
Reply

7. 
Kaveri is a significant step forward in the x86 computing world and good for consumers. AMD decided several years ago to switch to bulk production for flexible manufacturing of new products to prevent some of the Fab issues they have experienced in the past with GloFo.

GloFo has also upped their game and has 20nm almost ready to roll out. It's worth noting however smaller nodes primarily benefit lower power consumption, not increased clockspeed or IPC, so rushing to a smaller node doesn't bring much for the financial investment.

Steamroller brings up to 20% increased IPC over Richland so it's all good because with hUMA now the CPU and GPU do what they are best at without having to wait on the other processor. This is a really big deal for the x86 environment.
5 6 [Posted by: beenthere  | Date: 01/09/14 08:55:59 AM]
Reply
- collapse thread

 
show the post
1 4 [Posted by: linuxlowdown  | Date: 01/09/14 09:01:01 AM]
Reply
 
Its a good upgrade for AMD Llano APU
6 2 [Posted by: xentar  | Date: 01/09/14 09:57:04 AM]
Reply
 
Steamroller is just correction of few big mistakes they pushed into Bulldozer. AMD in general delays progress in computing word, by releasing products technologically delayed few yeras behind Intel and NV. Thankfully, Intel and NV have creative designers, and they push innovations forward, without need AMD competitve products. Rather, their own previous generations and user benefit is enough to make these innovations.

While I agree that GloFo (and other fabs) is delayed by some year behind Intel, this is really unimportant for AMD marketshare. AMD cannot sell everything ther ordered from GloFo, and they must perform inventory write-off to push it out. Now, when AMD bulked their CPU, then can produce it at TSMC or other fab with better small-quantity orders. But the realiity is that small-quantity translates to higher cost per chip, so AMD wont get many benefits from this 'bulked freedom'.
2 3 [Posted by: Tristan  | Date: 01/09/14 03:21:11 PM]
Reply

8. 
show the post
3 12 [Posted by: amdzorz  | Date: 01/09/14 10:29:06 AM]
Reply
- collapse thread

 
Ok, I'll bite. Why is it that the 8320 is a "fake" 8 core again?
3 3 [Posted by: Zshazz  | Date: 01/09/14 11:07:20 AM]
Reply
 
People don't understand how the AMD design works and confuse it with the hyperthreading design from intel.
This combined with some negative attitude toward AMD result into the "fake cores".

I don't blame the first, I though the same when they first introduced that design.
4 3 [Posted by: nitro912gr  | Date: 01/09/14 12:40:42 PM]
Reply
 
show the post
3 12 [Posted by: amdzorz  | Date: 01/09/14 01:41:17 PM]
Reply
 
Intriguing. Well, it would be intriguing if your facts were correct. It doesn't share a decoder, actually (I did look it up and there's clearly 2 decoders per "module". It shares an FPU, but judging from your post history, "floating point math does nothing for integer operation which is 95% of all non-game code. finding pi~ in my network stack is not useful...", so that's hardly a relevant observation.

Unless, of course, FP only matters when it's a disadvantage. And FP doesn't matter if it's for an advantage.

The facts are this: if you're running FP intensive code, then indeed it will not be able to match the speed of something that has 8 FPUs. But, as you've state, most code isn't that way, so it's hardly fair to call it a "fake" 8 core when it's a real 8 core when it matters, according to you.
4 2 [Posted by: Zshazz  | Date: 01/10/14 03:27:32 AM]
Reply

9. 
I don't care if it's Nvidia saying Cuda "cores", or AMD saying "compute cores", because that is all just marketing core obfuscation on both parties part. Just tell me the number of GPU cores, and the number of CPU cores, and in a intelligent manner explain to me how the AMD APU, or the Nvidia "APU" equivalent, utilizes any custom software/drivers/HAL, or Internal to the CPU microcode/hardware, TO: utilize the GPUs graphics cores for general purpose compute. This constant renaming and rebranding of core CPU/GPU functionality, without a properly written primer to explain the HSA abilities, that both AMD GPUs, and Nvidia GPUs, have via OpenCL, proprietary/simi-proprietary drivers(CUDA,Mantle), or other HALs, just confuses rather than informs the potential customer, and any non technical members of the so called technical press. This is just the expected marketing double speak, in its quest to sell to the "unwashed masses", and regurgitating or paraphrasing this Marketing materal in unannotated, or unexplaned form, is just bad journalism.

“APU-Optimized” is just another marketing quackspeak way of saying the process yealds for this APU where not very good at the previously advertised roadmap performence(expected GigaFlops/whatever Flops metric) stated in the previous marketing estimates, so the clocks have been lowered on the CPU side, and more GPGPU performence was added on the GPU side, to make up for the crappy initial projections. This often happens, when marketing monkeys are given to much Sway, over the engineering folks, or CEOs want to look good or pitch little white lies to potential investors!
3 1 [Posted by: BigChiefRunAmok  | Date: 01/09/14 04:39:51 PM]
Reply

10. 
"APU-optimized" = lower peformance, cheaper.

Being able to fit more transistors to same space just means they can make the chip smaller=cheaper, they could fit all the things they want on other process id they just made the chip bigger and more expensive to manufacture. Or they could put slightly smaller amount of those GPU cores with chip that's about the same size.

So they are practically trading single-thread performance(very important thing, where they are already about 1.5 times behind intel) to either
1) making the chip cheaper to manufacture.
2) making the GPU side where they are already leading, even better(even though they cannot get full benefit from making the GPU side bigger because of lack of memory bandwidth)

So intel's lead in single-thread cpu performance will only grow.


2 1 [Posted by: hkultala  | Date: 01/10/14 03:07:26 AM]
Reply
- collapse thread

 
And what makes you think that Intel, who has the lead in single threaded performence, will have any motivation to improve its single threaded performence, now that AMD is no longer competiting for the x86 single threaded performence crown? And any Intel CPU, is for gaming of the high performence kind, no good by itself without a Nvidia, or AMD descerte GPU. AMD's APUs are not being designed exclusively for the desktop environment any more than Intel's CPU are being designed for the same market. Most of the engineering capitol of both Intel and AMD is being used towards competing in the mobile markets, and the low power laptops/tablets market, and the desk top SKUs are just being derived from stripped down server chip technology. The real threat to Intel, from AMD, comes in the form of the console Gaming APU, and its merging of a more powerfull x86 CPU, and with its shareing a Big Bus with a more powerfull AMD GPU, and being connected, via that Fat GPU style BUS, directly to GDDR5 memory. This extra bandith and a unified memory controller for the CPU/GPU combo on the AMD gaming APUs negates any single theading issues that AMDs x86 CPU may have had in the past. If AMD were to ever take its gaming APUs to the next level and pair it with AMDs high end desktop SKU type GPUs with larger GPU executition core counts, and do this in an APU style gaming form factor, and create an entire gaming APU system on a PCI card form factor, with the entire gaming console running on a PCI card, APU, GDDR5 MEMORY, Large on DIE RAM, gaming OS, and gaming engine/s, well the desktop's CPU May no longer be relevant in the gaming equation! And Just to add to this, Nvidia's Denver cores could be paired with Nvidia's high end GPU tech, and will become Nvidia's answer to AMD's threat in having a complete gaming based "APU" type of system, all on a PCI card from factor. These types of APU style, complete gaming systems on a PCI card, will be the next logical step in the arms race for the, new CPU/GPU mashup complete gaming "APU" type systems on a PCI card, competition between Nvidia and AMD, and any maker of only Motherboard CPUs, that does not possess High end gaming "APU" Type technology, will be out of the gaming loop, for all platforms, if they do not quickly come up to high end graphics speed.
2 1 [Posted by: BigChiefRunAmok  | Date: 01/10/14 11:38:24 AM]
Reply

11. 
Hum. Okay then: AMD has used 3D-high density libraries in Kaveri, hence the astounding transitor/density [about 1/3 higher then ivy on intel's 0.22tg.] I guess the terrible transitor count explains why high server [hence high desktop] chips were all cancelled. AMD seems to have fattened up by maybe 1/2 billion in the uncore over trinity, yikes. What are they all doing I wonder? Yes there's: beefed up imc/vce, ecc physical layers, true audio, dvfs micro controller and so on..but still 1/2 billion plus? Are we going to have to wait for 2016 and glofo's 0.14xm node for 16 core+ apus? Times aticking AMD. Sheesh.
0 2 [Posted by: TopHatKiller  | Date: 01/11/14 11:42:22 AM]
Reply

12. 
So basically AMD is betting the farm on devs adopting GPGPU compute stuff as mainstream. These are the same devs that still haven't really bothered for the most part to support multi-core cpu's. By the time any of this happens, Intel will probably have moved onto Skylake and have used their massive process advantage to brute force add tons of graphics capability on-die.
1 0 [Posted by: AnonymousGuy  | Date: 01/11/14 04:26:27 PM]
Reply
- collapse thread

 
Developers, the ones that write most Applications, and gaming engines(non System) developers, do not have to worry, and have not had to worry, about coding/developing for multitreading/multicore processors for years now, as all the Multiprocessor/MultiThreading grunt work is done by the compilers/APIs/frameworks/librarys, and the SDKs(System Development Kits). The need for genral programmers to worry abouting coding for multicore, has been mostly automated, and AMD, ARM, Intel, and Nvidia, others, offer SDK kits, and have allways had SDKs and API frameworks available, for programmers to eaisly code and use multicore CPUs and GPUs, and these SDKs are always avaiable before the consumer sees the new CPU/GPU hardware in any of the finished OEM products. This has been standard practice for years, and GPGPU has been available to any one with Intel's Ivy bridge processors support of OpenCL, and the OpenCL support that AMD's and Nivida GPUs(integrated, and Descrete) have had for quite a few years now, long before Intel got around to supporting OpenCL GPGPU with IVY Bridge. And quite a few software applications support OpenCL GPGPU acceleration, and have been for a few years now. AMD HAS APUs, and the Nvidia Tegra is an "APU" (the Nvidia Tegra k1, is a great new APU type system with full desktop version support for OpenGL, and OpenCL, on a mobile CPU/GPU, A first for a mobile "APU" )! So Nivida's and AMD's GPGPU farm[Bets] have been safe with GPGPU support, way longer than Intel with Intel's late to the party, support of OpenCL drivers, and OpenCL is a Very HSA sort of Open standard, that is open to all.
0 2 [Posted by: BigChiefRunAmok  | Date: 01/11/14 08:20:20 PM]
Reply

[1-12]

Add your Comment




Related news

Latest News

Monday, April 14, 2014

8:23 am | Microsoft Vows to Release Xbox 360 Emulator for Xbox One. Microsoft Xbox One May Gain Compatibility with Xbox 360 Games

Tuesday, April 1, 2014

10:39 am | Microsoft Reveals Kinect for Windows v2 Hardware. Launch of New Kinect for Windows Approaches

Tuesday, March 25, 2014

1:57 pm | Facebook to Acquire Virtual Reality Pioneer, Oculus VR. Facebook Considers Virtual Reality as Next-Gen Social Platform

1:35 pm | Intel Acquires Maker of Wearable Computing Devices. Basis Science Becomes Fully-Owned Subsidiary of Intel

Monday, March 24, 2014

10:53 pm | Global UHD TV Shipments Total 1.6 Million Units in 2013 – Analysts. China Ahead of the Whole World with 4K TV Adoption

10:40 pm | Crytek to Adopt AMD Mantle Mantle API for CryEngine. Leading Game Developer Adopts AMD Mantle

9:08 pm | Microsoft Unleashes DirectX 12: One API for PCs, Mobile Gadgets and Xbox One. Microsoft Promises Increased Performance, New Features with DirectX 12

3:33 pm | PowerVR Wizard: Imagination Reveals World’s First Ray-Tracing GPU IP for Mobile Devices. Imagination Technologies Brings Ray-Tracing, Hybrid Rendering Modes to Smartphones and Tablets

2:00 pm | Nokia Now Expects to Close Deal with Microsoft in Q2. Sale of Nokia’s Division to Close Next Month