AMD Launches The A10-7800, The 65 Watt Kaveri

[Marc Driftmeyer]

4 cores means here 4 modules right? like the fx4300 as example.

I just bought a fx6300, in my opinion the best u can get at the moment from amd price + performence / consumption wise.

I dont really get why amd brings no 6 core kaveri, seems to me the best compromise my fx6300 consumes more or less the same energy then a fx4300 but has in most games nearly the performance of a fx8300.

Of course if u need more performance u have to buy intel stuff.

I had a fx4100 before so I had the board and because I dont have a 400 euro grafics card the fx6300 should not bottleneck me (at the moment hd5850 soon maybe hd7950)

The 4 cores are Two Dual Core CPUs.

http://www.guru3d.com/articles-pages/amd-a10-7800-kaveri-apu-review,2.html

[rikkinho]

no

The 4 cores are Two Dual Core CPUs.

http://www.guru3d.com/articles-pages/amd-a10-7800-kaveri-apu-review,2.html

it's not a true quadr core or true dual core, they share resourses, because of this single core performance is so bad, amd need to give up from this cpu and build something new, single core performance is to crucial yet (in some cases ofc)

[dungeon]

it's not a true quadr core or true dual core, they share resourses, because of this single core performance is so bad, amd need to give up from this cpu and build something new, single core performance is to crucial yet (in some cases ofc)

That is design for HSA , no question that AMD will give up from that . Because all is shared you have 12 Compute Cores in this APU .

People currently not interested in all this, may buy traditonal FX CPUs .

[benmoran]

The market it's shooting for is probably almost non existent because anyone really interested in graphics performance will most likely run a discrete graphics card, and those not interested in graphics are probably more interested in cpu performance, which is Kaveris weak spot.

I run an a10-7850k Kaveri on my main Desktop. The reason? I only have space for small form factor PCs, and I wanted the best graphics performance I could get in a CPU/APU. A TV tuner is in my pcix slot, so no discreet card. I am not a crazy gamer, but I do play a lot of games. This APU is more than sufficient for 98% of the 150 games I have in my Steam library. And the CPU is "good enough" by a long shot.

You're also ignoring the budget market. If you look at who's actually buying these things, it's folks who want to play games on medium at a cheap price.

[Zan Lynx]

it's not a true quadr core or true dual core, they share resourses, because of this single core performance is so bad, amd need to give up from this cpu and build something new, single core performance is to crucial yet (in some cases ofc)

AMD's cores are full speed at integer math. As far as I know they only share the FPU for floating point.

I think AMDs end goal with this architecture is to replace the FPU and all MMX/SSE instructions with their HSA design and use the floating point units on the GPU instead of the CPU.

Why keep inventing newer and wider SSE instructions like Intel's 512 bit ones, when you can throw over a thousand GPU compute units at the data?

[GraysonPeddie]

What is so special about single-threaded performance? Does single-threaded performance beat the snot out of multi-threaded performance?

Why are people making such a big deal out of single-threaded performance?

[Luke]

With the right software this would make one HELL of a video editing chip

generally speaking it has quite somecompute power via its GPU which seems to lack bandwidth and a low performing CPU. Maximum performance is not reached for simple FP operations is it? but for some graphic or OpenCL/HSA stuff. Thats what id call it specialized. At least for the software that exists to this point and is unable to get a hold of the power it could offer. Youd need specially written software. thats why i call it specialised.
(i wouldnt if some Gallium state tracker was able to make its graphic's part's CU's power available for completely general purpose so that (almost) every program that uses the cpu could run on the GPU instead. Or maybe someone will make a framework for such optimisation someday)

long story short: if you wanted a generally good system youd go with an Intel CPU and a discrete Nvidia graphics card (on linux)
AMD is, as a matter of fact, (in) a niche and for lovers
... and budged if you do not rely on fast ram... whats a cheap CPU for when you need f***ing expensive ram withought still exceeding intel

And yes i am very much interested in AMD to succeed and bring some new stuff. They need to really get some kind of feature and make it common and i think HSA is a good approach but with the software lacking so much behind it very much reminds me of the introducion of 64bits when it took years to be somewhat fully used. There are so many people out there craving for some new toys especially in the linux world. They could really be the ones giving hardware to universities for some masters thesis. given they can or do (afford to) open up documentation...
i mean look at raspberry pi. They could do the same if not better. or if you want look at tesla cars. AMD cannot win on a closed market (or maybe with the before mentioned inspiration and a HUGE(!!!) investition as a final trump card... which will not happen due to investors). and the longer they wait the worse it gets. i mean if it goes on like this they may have to ask Intel to manufacture chips for them when GF and TSMC cannot fill their needs...

An HSA backend under Kdenlive, based on the GLSL backend now being played with would make Kdenlive the equal of Adobe's mercury backend for GPU acceleration without a CPU/GPU memory copy penalty, and would make common, cheap APU's easily capable of outperforming Bulldozer 8 core in terms of render time and realtime playback with many effects enabled. Would be seriously nice to see a proc overclocked to about 80W with on-chip GPU defeat an 8 core chip that can draw 200W when heavily overclocked and requires a discrete GPU that uses even more power.

Maximizing HSA performance would in fact REQUIRE that the GPU being used for HSA be on-chip, so the same memory controller and the same RAM are being used. Otherwise you'd have a CPU reading GPU ram over PCI-E, or a GPU reading CPU ram over PCI-E. I don't know if that would even be supported, but it would still impose about half of the memory copy penalty that now slows down video editors using the GPU to render complex floating point effects (movit) that are beyond the CPU's capacity. You could still use another card for games, but for most Linux games you'd never need it and could pocket the savings in both power and hardware cost.

I've always been impressed with hardware video encode/decode, and how a tiny camera can encode a video in realtime that a general purpose CPU using ten times the power (P4 or Intel Atom) cannot play back in realtime.

[ninez]

How are you comparing TSX vs. non-TSX performance?

via linux-3.14 [although, you could do it b4 that release]

https://software.intel.com/en-us/blogs/2013/05/03/intelr-transactional-synchronization-extensions-intelr-tsx-profiling-with-linux-0

I've built glibc with/out lock-elision support on my 4790/Haswell and tested/bench'd heavy locking apps on PREEMPT_RT_FULL. In some cases, there are some fairly substanial gains to be had... Also, i have used xbegin(), xend() and friends [elsewhere] directly, similar to what is described/shown here; https://01.org/blogs/tlcounts/2014/lock-elision-glibc ... I plan on digging more- but this machine is only 2 weeks old or something [and i am fairly busy with work,etc].

[johnc]

via linux-3.14 [although, you could do it b4 that release]

https://software.intel.com/en-us/blogs/2013/05/03/intelr-transactional-synchronization-extensions-intelr-tsx-profiling-with-linux-0

I've built glibc with/out lock-elision support on my 4790/Haswell and tested/bench'd heavy locking apps on PREEMPT_RT_FULL. In some cases, there are some fairly substanial gains to be had... Also, i have used xbegin(), xend() and friends [elsewhere] directly, similar to what is described/shown here; https://01.org/blogs/tlcounts/2014/lock-elision-glibc ... I plan on digging more- but this machine is only 2 weeks old or something [and i am fairly busy with work,etc].

Interesting. When I upgrade my hardware I'll have to look into how I can get the most out of that.

Thanks for the links.

[Brane215]

This is nice, but...

New ARMs from AMD are not that far away. They should offer updateg GPU and up to 8-core A57.

Which should play nicely with freshly compiled code on Linux methinks. And ECC support should fit like hand in glove with new filesystems like BTRFS and like.

So, if I had to make my choice now, I'd go for something like that A10-7800, but will wait for AMD ARM.