Then why are Intel releasing them in opposite order for Broadwell? Mobile first, then desktop.
Because they use a smaller node that will indeed reduce
power comsumption even if not optimaly in respect of process
maturation and eventual better binning since their high volumes
will allow to get good chips from the first batches , non qualified
dies will be spared for later use in DT SKUs.
When sites dubbed serious will stop spreading viral marketing,
and innacuracies , like in this quote :
http://www.anandtech.com/show/7711/...f-kaveri-and-other-recent-amd-and-intel-chips
Kaveri is also on a smaller node, moving from 32 nm (Richland) -> Kaveri (28nm).
AMD's issues are never their own fault, are they?
Only a half node shrink and to a bulk process that
is clearly better only for density.
Well, Intel have never released mobile before desktop until Broadwell. That is despite the stuff you mentioned, which has been true for several other node shrinks earlier too.
They are competitive in the DT space but not in mobile,
AMD said that they lost APU market in the mobile market,
certainly due to harsh competition from the ARM based
devices that are , wrong or right, estimated as being good
enough by the younger consumers.
They could do nothing if a guy who is not very competent
think that four times the throughput in a few % of a code
will yield four times the throughput for the whole code.
He could as well have used the GPU FP perfs as comparative
parameter but the outcome wouldnt had been the one doctored
in this innacurate, not to say trollish, article.
And if ARM is considered good enough in mobile, that issue should apply to Intel as well.
The article is about floating point peak performance. It was never represented as a look at general purpose performance.
Also -- as far as I can tell it looks like Anandtech is sponsored by AMD, kudos to AT for still remaining objective.
I know how the human brain can mix up things. So I'm here to help:They're not responsible that Kaveri didn't get the 30% performance increase they touted, It's Intel stock pumpers fault, right?
The 30% claim was 30% higher ops/cycle provided by the front end of the core.I know how the human brain can mix up things. So I'm here to help:
http://forums.anandtech.com/showthread.php?t=2355354
The article is about floating point peak performance. It was never represented as a look at general purpose performance.
gFlops and gFlops doesnt give you the same performance always. See for example the NVIDIA GTX580 vs GTX680 in CUDA and compute. GTX680 has double the gFlops over GTX580, BUT it is slower in the majority of the workloads than GTX580.Now we compare Kaveri against Haswell. In applications depending upon fp64 performance, conditions are not generally favorable to Kaveri. Kaveri's fp64 peak including both the CPU and GPU is only about 110 gflops. You will generally be better off first optimizing your code for AVX and FMA instructions and running on Haswell's CPU cores. If you are using Windows 8, you might also want to explore using Iris Pro through C++ AMP in conjunction with the CPU. Overall I doubt we will see Kaveri being used for fp64 workloads.
Exactly. And this doesn't mean 30% more performance.The 30% claim was 30% higher ops/cycle provided by the front end of the core.
Does anybody have any idea why AMD will not be releasing the mobile versions of Kaveri until Q2/Q3? Intel will be doing just the opposite with Broadwell, and release mobile CPUs before desktop CPUs... :hmm:
AMD's issues are never their own fault, are they?
Like a child, they are not responsible for their own situation.
They're not responsible that Kaveri didn't get the 30% performance increase they touted, It's Intel stock pumpers fault, right?
Then what hes trying to say by the following exactly ??
gFlops and gFlops doesnt give you the same performance always. See for example the NVIDIA GTX580 vs GTX680 in CUDA and compute. GTX680 has double the gFlops over GTX580, BUT it is slower in the majority of the workloads than GTX580.
Just because you have lower gFlops doesnt mean you will get lower performance. So, his recommendation to optimize the code for AVX and run it in CPU than optimize it for the GPU is not correct because it is only based on the number of gFlops.
Cuda and compute use a lot of DP. The 580 has ~190 DP gflops. The 680 has ~128 DP Gflops.
Is there a conclusion to be drawn about Kaveri efficiency? I mean, let's forget about absolute performance.
What is the sweet spot for these parts (probably 45W?)?
Is it competitive with intel products of comparable TDP?
I might be interested in an ultraportable with mobile Kaveri if that's the case.
IPC grow it's at least double than from Bulldozer to Steamroller
Citation needed.
A 100% improvement in IPC is the equivalent of saying IPC has doubled. That is, if clockspeed is held constant, instructions per second will be doubled.