Friday, November 1st 2013
![AMD](https://tpucdn.com/images/news/amd-v1723816949774.png)
AMD to Unveil Next-Generation APUs on November 11
As a follow-up to our older article on how December-January will play out for AMD's next-generation APU lineup, we have news that the company will unveil, or at least tease its next-generation desktop APU, codename "Kaveri," on November 11, 2013. It's when the company will host its APU'13 event, modeled along the lines of GPU'13, held in Hawaii this September, where it unveiled its Radeon R9 200 and R7 200 GPU families. On its backdrop, the company will also hold its 2013 AMD Developer Summit, which brings together developers making software that take advantage of both CPU and OpenCL-accelerated GPUs. APU'13 will be held in San Jose, USA, and like GPU'13, will be live-streamed to the web. In addition to new APUs, the company is expected to make some big announcements with its HSA (heterogeneous system architecture) initiative that brought some big names in the industry on board.The agenda for APU'13 follows.
- 4:00 - 5:00 p.m. (PST), Monday, November 11:
o Lisa Su, senior vice president & general manager, Global Business Units, AMD: "Developers: The Heart of AMD Innovation"
o Phil Rogers, corporate fellow, AMD: "The Programmers Guide to Reaching for the Cloud" - 8:30 - 9:30 a.m. (PST), Tuesday, November 12:
o Mike Muller, CTO, ARM: "Is There Anything New in Heterogeneous Computing?"
o Nandini Ramani, vice president, Java Platform, Oracle Solutions: "The Role of Java in Heterogeneous Computing, and How You Can Help" - 1:15 - 2:15 p.m. (PST) Tuesday, November 12:
o Dr. Chien-Ping Lu, senior director, Mediatek USA: "How Many Cores Will We Need?"
o Tony King-Smith, executive vice president, Marketing, Imagination Technologies: "Silicon? Check. HSA? Check. All done? Wrong!" - 8:30 - 9:30 a.m. (PST), Wednesday, November 13:
o Dominic Mallinson, senior vice president, Software, Sony: "Inside PlayStation 4: Building the Best Place to Play"
o Brendan Iribe, CEO, Oculus VR: "Virtual Reality - A New Frontier in Computing" - 1:15 - 2:15 p.m. (PST) Wednesday, November 13:
o Johan Andersson, technical director, DICE: "Rendering Battlefield 4 with Mantle"
o Mark Papermaster, CTO, AMD: "Powering the Next Generation Surround Computing Experience"
44 Comments on AMD to Unveil Next-Generation APUs on November 11
www.planet3dnow.de/cms/?p=3279&preview=true
Benchmark results of AMD on the Internet " Kaveri " chips are being discussed, which have appeared on the Bench portal of Sisofts Sandra . The test results are less interesting than the output hardware configuration :
If the information is correct , it would " Kaveri " remarkable about 13 Compute Units ( CUs ) - which corresponds to 832 shader - can have and would thus view of computing power between a Radeon HD 7770 (10 CUs/640 shader) and a Radeon HD 7790 ( 14 CUs / 896 shaders) .
Overview of graphics performance makes this because of the much lower memory bandwidth (128-bit DDR3 on socket FM2 + vs. 128 -bit GDDR5 on the graphics card ) , which is about a quarter , not much sense. However, the pure computing power with GPGPU beats by eg OpenCL benchmarks. In SiSoft achieved a " Kaveri " in an encryption test thus more than twice the power of a "rich country " A10- 6800K with the Radeon HD 8670D GPU called internal :
The GPU of the " Kaveri " overclocked but with only 600 MHz and accesses DDR3 -1600, while the A10- 6800K system is clocked at 844 MHz and DDR3 -2133 features . The strong growth in terms of GPU later versions with DDR4 would also be declared .
This system scored only 5.8 GB / s, which is exactly the ratio 8/13 to 9.4 GB / s of the 13- CU system. Thus, one would have a strong indication for the 13 CUs . A 100% security but there is not natural. Maybe it was only in May of any performance robbing bug is now fixed , but the CUs are now read wrong. Up to the Windows version ( Windows 7 / Windows 8) and a slightly different version of Sandra ( 19:35 / 19:44 ) the configuration are identical. The same user also has an A10- 5800K system, in which also the graphics part only runs at 600 MHz. This system achieved the benchmark only 2.6 GB / s, which is even less than a third of the putative 13 -CU- Kaveri .
A closer look at the Sandra results , we noticed that when the system information 1 device is shown with 2 threads. This is strange because usually far - even for large cards - only one thread per device is specified . 2 threads there is only Crossfire setups then 2 devices . However, we can not say exactly how a possible linkage of " Kaveri " and " Hainan" GPU ( 5 CUs have the chance ) would be in the system at the moment. In previous hybrid CrossFire setups are indeed always called both crisps and displayed by Sandra , but this could change due to technical reasons (HSA ) or simple marketing considerations ( AMD's new naming scheme) . Coupling of 8 " Kaveri " CUs and 5 " Hainan" CUs would naturally explain the increased benchmark .
The other benchmarks are unspectacular , the measured 15 GB / s memory bandwidth with expectations higher than its predecessor . According to the leaked inside information Kaveri will indeed have an internal 256 -bit bus per module , so that one could expect a higher single-thread memory throughput. Interesting perhaps is that the graphics unit is described as "AMD Radeon R5 M200 Series" :
Because of the sloping number of 13 CUs of course also raises the question to what extent it might not be a read error . Conveniently, the same encryption result a user has already uploaded but in May , which ran on a " Kaveri " with the expected 8 CUs/512 shader according to system info: But to be fair in some apps "beat" translates into a small boost.
www.techpowerup.com/reviews/AMD/FX8150/11.html
Sure the 8350 is faster, but it's a slight disgrace at how effectively the Phenom X6's keep up with AMD's current gen of Bulldozer CPU's.
there is little "stress" in the stress testing.
I don't think the modules were the only problem. Because I made an observation in another thread:
The FX 8150 8 core is actually slower than the FX 6300 6 core in a lot of applications too. In Dawn of War II and Dragon Age Origins 12 FPS and 15 FPS separate the two. The FX 6300 seems faster in some of the non-gaming apps too, despite the 2 core, 2MB L2 and 100Mhz handicap. I can only put this down to the Piledriver refinements outweighing the Bulldozer architecture and both are module based.
www.anandtech.com/bench/product/434?vs=699
EPIC :roll:
Also 8 modules and 4 FPU's means those jazzy doubled up modules mean jack to a great deal of programs.
Until there is a significant CPU performance improvement or I feel the need my X6 will stay.
First step is to reducing floating point while reserving two integer units.A first batch gimped design,they perform underpar with previous Phenom II X6.After little tweaking,by giving two Decoder to feed Integer Unit and ALU within FPU they finally catch performance of single threading.A traditional concepts were 1+1,so major folks claims AMD didn't gave'em a proper core.With such a modular design,AMD can further tweak like creating more complex Fetch (divider,multiplexer,brancher,shifter),maximizing Decoder with two microCode each,giving more data path for Load-Store I/O within Integer and FP unit,maintain L2 cache coherency for Local Data and implementing eDRAM for Global L3 cache with advanced cache granularity shared with graphics compute unit.Notice this is why AMD "need" HSA to taking further serial and parallel task in one die.
Pretty much i'm delirious,taking nonsense without any proof or link...so take this with a huge grain of salt :p
Those new APUs are going to be excellent for gaming, especially if developers decide to utilize the GCN logic on 'em through Mantle or open-CL. The high-end APU I believe has 13 GCN vector cores, each with 64 ALUs.
GCN like Vector processors are the natural place to compute physics calculations because, like graphics, physics processing is driven by thousands of parallel and mutually independent vector workloads.
Shared memory access and heterogeneous queuing make those APUs the best gaming CPUs, if utilized the way they should be.
With Haswell things are not much better.. +18% clock costs over 60% more power (3900 to 4600mhz).. and that last 400Mhz is going to cost another 50%+ in power consumption to reach 5Ghz.
John Peddie