The Official Kaveri Review Thread (A10-7850K, etc)

[NTMBK]

It seems there is some reason kaveri is not 20nm. TSMC already has the 20nm process functional since fpga specialist xilinx is offering 20nm products and is on route for 16nm finfet process.
What would be the reason ?
http://www.xilinx.com/products/silicon-devices/fpga/virtex-ultrascale.html#Jan14NLEN

FPGAs are classic pipecleaner products, as they are priced so high that they can absorb terrible yields and high wafer prices. Ready for FPGA != ready for consumer APUs.

Besides, Kaveri was meant to launch a year ago originally.

 

[AtenRa]

TSMC 20nm could easily be 50%+ more expensive than GloFos 28nm currently.

 

[USER8000]

Watt the heck is happening with this thread??

Gets coat\.

 

[Abwx]

Hardware.fr made a separate test for the 7600 variant with a study
of the whole line TDPs , they conclude that the 95W 7850k and 7700K
are not real 95W APUs but rather 65W given their behaviour.

http://www.hardware.fr/articles/915-1/apu-amd-a8-7600-test-ctdp-turbo-retour-kaveri.html

 

[Enigmoid]

Hardware.fr made a separate test for the 7600 variant with a study
of the whole line TDPs , they conclude that the 95W 7850k and 7700K
are not real 95W APUs but rather 65W given their behaviour.

http://www.hardware.fr/articles/915-1/apu-amd-a8-7600-test-ctdp-turbo-retour-kaveri.html

Yep, there appears to be a hard cap on clocks under load to avoid blowing past the tdp.

a10-6800k

Basically can run luxmark on CPU + GPU at almost max boost clocks.

a10-7850k

Loses 900 mhz when CPU and CPU are stressed.

65W a8-7600

Loses even more clocks.

I3/i5 are solid

Goes on to show that its not really that the a10-7850k has lower power consumption at the same clocks, but that it drops clocks under full load.

There is a nice 300 mhz jump in clocks at 45W though the a8-7600 45W uses more power on the individual CPU or GPU tests though is able to put higher clocks when both are stressed.

 

[Abwx]

Yep, there appears to be a hard cap on clocks under load to avoid blowing past the tdp.

Goes on to show that its not really that the a10-7850k has lower power consumption at the same clocks, but that it drops clocks under full load.

There is a nice 300 mhz jump in clocks at 45W though the a8-7600 45W uses more power on the individual CPU or GPU tests though is able to put higher clocks when both are stressed.

I like how you abstained to post the power comsumption graph
just to do the wrong conclusion according to your prejudices,
implying that Kaveri reach TDP as high as a 6800K and then
throttle to not "blow past the TDP".

Here the graph so everyone can make his own conclusion
and have a take at you sense of "accuracy"...

This is power of the system measured at the main.

http://www.hardware.fr/articles/915-3/consommation.html

 

[coercitiv]

Goes on to show that its not really that the a10-7850k has lower power consumption at the same clocks, but that it drops clocks under full load.

So now we have more indirect proof that Kaveri does indeed abide by it's TDP rating.

However, if we know it does that by limiting clocks and we consider that a bad thing, how come it usually beats it's predecessor both in performance and power consumption?

 

[inf64]

However, if we know it does that by limiting clocks and we consider that a bad thing, how come it usually beats it's predecessor both in performance and power consumption?

Must be some magic

On the more serious note, once AMD get pass this crappy node (hopefully with Carrizo?) they should be doing better even at "65W" TDP bracket (vs Kaveri "95W" parts).

 

[Abwx]

Must be some magic

On the more serious note, once AMD get pass this crappy node (hopefully with Carrizo?) they should be doing better even at "65W" TDP bracket (vs Kaveri "95W" parts).

Dont be fooled by assumptions thrown out of thin air by people who do not pay attention to the measured numbers , actualy GFs 28nm process is not that bad at all , it takes a 4.3GHz frequency under Fritzchess MT to reach the 95W TDP mark , and still this is the power including the losses of the
VRMS , likely that 95W is reached when they measure 109.2W.

It makes no doubt that we ll see higher clocked versions this year.

http://www.hardware.fr/articles/913-9/cpu-consommation-overclocking.html

 

[Enigmoid]

I like how you abstained to post the power consumption graph
just to do the wrong conclusion according to your prejudices,
implying that Kaveri reach TDP as high as a 6800K and then
throttle to not "blow past the TDP".

Here the graph so everyone can make his own conclusion
and have a take at you sense of "accuracy"...

This is power of the system measured at the main.

http://www.hardware.fr/articles/915-3/consommation.html

I didn't post it because it was irrelevant to what I was talking about which was kaveri dropping clocks under load on the top SKU compared to richland.

Don't get me wrong there is a nice increase on the 45W sku which manages higher clocks (but then on that power graph you provided is using more than richland 45W).

If you want to look at efficiency I see the a10-7850k and the a10-5800k neck for neck on 1 thread fritz with the a10-7850k ahead on 4 threads (less MT penalty). The a10-7850k barely matches phenom x6 (on 45nm for crying out loud) in fritz MT efficiency.

I have no doubt that there will be a refresh on the same physical chip, much like AMD did with trinity -> richland, turbo on kaveri needs improvement (gpu turbo) and excavator will probably be delayed due to budget issues.

So now we have more indirect proof that Kaveri does indeed abide by it's TDP rating.

However, if we know it does that by limiting clocks and we consider that a bad thing, how come it usually beats it's predecessor both in performance and power consumption?

You cannot look at peak power (furmark and prime) and compare it to application performance. If you are looking at efficiency then the power consumption must be for the work the CPU is doing, not the max power of the CPU.

 

[coercitiv]

You cannot look at peak power (furmark and prime) and compare it to application performance. If you are looking at efficiency then the power consumption must be for the work the CPU is doing, not the max power of the CPU.

No, i'm looking at application performance and comparing it to application performance.

I'm also looking at Tomshardware and Techreport and comparing energy consumption for a given task. For example, in the tomshardware bench, the A10 7850 finishes work a bit sooner than A10 6800K while consuming about 20% less energy to do so.

 

[AtenRa]

Kaveri A10-7850K vs Core i3 4330 + GT630 in 5 Games at 1080p

Well, you dont need 2133MHz with the Core i3 but even at the same price, the Kaveri is the better option not only in price/performance but also in power efficiency and it can be used in a slim SSF case.

 

[Enigmoid]

No, i'm looking at application performance and comparing it to application performance.

I'm also looking at Tomshardware and Techreport and comparing energy consumption for a given task. For example, in the tomshardware bench, the A10 7850 finishes work a bit sooner than A10 6800K while consuming about 20% less energy to do so.

A10-6800k is on the far side of the performance power curve. Compare to the a10-6700 and they are roughly the same.

 

[know of fence]

Kaveri A10-7850K vs Core i3 4330 + GT630 in 5 Games at 1080p

As comical and desperate as this comparison is, at least PCper left a remaining 30$ bucks budget for the RAM premium, and they didn't use the ridiculously priced i3 4340 (20$ extra for 0.1 GHz).

Not sure where these bottom of the barrel DDR3 dGPUs are in terms of performance per dollar, though the GT 640 is already significantly worse in FPS/$ than a GT 650. I imagine that trend continues for the GT 630 used in this comparison.

 

[Dresdenboy]

I didn't post it because it was irrelevant to what I was talking about which was kaveri dropping clocks under load on the top SKU compared to richland.

Is that 100W value ACP? Could explain the high usage, avoiding he need to drop clocks.

It also looks like Luxmark GPU uses nearly as much on 45W Kaveri as on the higher models. -> less room for CPU clocks, since it uses TDP

 

[ams23]

Well, you dont need 2133MHz with the Core i3 but even at the same price, the Kaveri is the better option not only in price/performance but also in power efficiency and it can be used in a slim SSF case.

That is incorrect. The [DDR3] GT 630 that PCPer used is based on the Fermi architecture which is many years old, fabricated on a 40nm fab. process, and has 96 CUDA cores. The more recent [DDR3] GT 630 (which was launched seven months ago) fabricated on a 28nm fab. process and based on the Kepler architecture [GK208] has 4x more CUDA cores, more than 2x the GFLOPS throughput, and less than one half the TDP in comparison to the ancient Fermi variant that they used.

 

[AtenRa]

That is incorrect. The [DDR3] GT 630 that PCPer used is based on the Fermi architecture which is many years old, fabricated on a 40nm fab. process, and has 96 CUDA cores.

So ?? it is NVIDIAs product in the ~$60 segment.

The more recent [DDR3] GT 630 (which was launched seven months ago) fabricated on a 28nm fab. process and based on the Kepler architecture [GK208] has 4x more CUDA cores, more than 2x the GFLOPS throughput, and less than one half the TDP in comparison to the ancient Fermi variant that they used.

The Kepler GT630 also has 64-bit memory interface making it even slower than Fermi GT630 . If you want to use the Kepler card do it

http://www.geforce.com/hardware/desktop-gpus/geforce-gt-630/specifications

The only faster GT630 than the one they used is the Fermi based GDDR-5 variant but that is even more expensive.

 

[AtenRa]

As comical and desperate as this comparison is, at least PCper left a remaining 30$ bucks budget for the RAM premium, and they didn't use the ridiculously priced i3 4340 (20$ extra for 0.1 GHz).

Why is that comical and desperate ???

Yes they could use the Core i3 4130 that is a little cheaper. Still even at the same price, the Core i3 + GT630 will loose to the Kaveri A10-7850K in almost every game.

 

[ams23]

So ?? it is NVIDIAs product in the ~$60 segment.

Incorrect. One can get a Kepler GT 640 right now for $59 US$: http://www.amazon.com/Zotac-NVIDIA-G...eywords=gt+640 (and there are various other variants that are $69 US$).

making it even slower than Fermi GT630

Incorrect again. The [DDR3] Kepler GT 630 variant has 2x more ROP's than the [DDR3] Fermi GT 630 variant. So in comparison, it has > 2x more pixel fillrate [ie. ROP throughput], > 1.28x more texture fillrate, and > 2x more shader throughput, all at half the TDP. In fact, Asus claims that their $55 [DDR3] Kepler GT 630 variant has up to 40% higher performance than the Fermi GT 630 variant: http://www.newegg.com/Product/Produc...82E16814121836

Last but not least, if one wants a Kepler GPU with a 128-bit memory bus interface, there is Kepler GT 640, which has far superior performance and comparable (if not lower) power consumption compared to [DDR3] Fermi GT 630. The Fermi GT 630 card is based on the ancient Fermi GT 440 that was launched three years ago.

Note that one can now find [GDDR5] GTX 650 for $79 US$, which is 100-200% faster (ie. 2-3x faster) than [DDR3] Fermi GT 630, with only slightly higher power consumption: http://www.newegg.com/Product/Produc...82E16814134163

 

[know of fence]

Why is that comical and desperate ???

Already hinted at it in my reply, but it boils down to those low bottom cards offering terrible performance per dollar. Saying we are better than trash, is pretty comical. Maybe, just maybe, people who are buying cards like the HD 7770/650 TI are not poor and stupid, it could be that they prefer to get the most value for their money at a reasonable thermal envelope, something top tier product don't offer basically.

This whole 1080p@30 narrative is desperate, because it's the setting at which Kaveri is bound to be GPU bottlenecked, best case scenario along with meaningless FPS averages not remotely representative of the actual experience. There is no use contesting discrete cards and i3s (intel IPC lead and DDR5 can't be be beat, sad but true), the whole point of Kaveri is not to have a dGPU.

I like Kaveri, it eliminates redundancy and a fan and thus a source of noise, it has low idle consumption and the low graphics clock means its graphics is reasonably power efficient. The RAM limitations, even sort of make Kaveri interesting for tinkering. Most of all it removes the need to dedicate one's build to something as stupid and soul destroying as hardcore gaming, still leaving the option to experience your Portal 2s and Stanley Parables casually, knowing that drivers and architecture will support it.

 

[SPBHM]

Kaveri A10-7850K vs Core i3 4330 + GT630 in 5 Games at 1080p

Well, you dont need 2133MHz with the Core i3 but even at the same price, the Kaveri is the better option not only in price/performance but also in power efficiency and it can be used in a slim SSF case.

"great" gaming selection, they tried hard to make the 7850K look good I guess,

630 is crap, but even the 640 is enough

Athlon X4 750K $79.99
ECS GTX650-2GR5-GFM GeForce GTX 650 2GB 128-Bit GDDR5 PCI Express 3.0 Video Card
$79.99

why not include this on the video and actually make it informative, relevant?

cheaper CPU + VGA is far better for gaming.

 

[Abwx]

"great" gaming selection, they tried hard to make the 7850K look good I guess,

630 is crap, but even the 640 is enough

Athlon X4 750K $79.99
ECS GTX650-2GR5-GFM GeForce GTX 650 2GB 128-Bit GDDR5 PCI Express 3.0 Video Card
$79.99

why not include this on the video and actually make it informative, relevant?

cheaper CPU + VGA is far better for gaming.

There s barely 10% between the 7850k and a gt640 with whatever CPU,
not a difference worthy of a dgpu , indeed you have to upgrade your
recomendation up to a gt650 for the dgpu to really make sense....

 

[Arkaign]

"great" gaming selection, they tried hard to make the 7850K look good I guess,

630 is crap, but even the 640 is enough

Athlon X4 750K $79.99
ECS GTX650-2GR5-GFM GeForce GTX 650 2GB 128-Bit GDDR5 PCI Express 3.0 Video Card
$79.99

why not include this on the video and actually make it informative, relevant?

cheaper CPU + VGA is far better for gaming.

This times a billion. Once the new nodes roll through and the $69 and $99 cards start to get that arch, it's going to get much, much, much worse for current IGP/APU as well.

Q4 2014 we'll probably see something like a GT840 2GB for $89ish, 20nm, performance in the 660 vanilla range or a little better.

APU is for people who really really want a slim case. That's it. Performance will always be worse than a CPU+GPU combo of even approximately equal price. AMD 750k + 7790 is somewhere near 7850K pricing, and it absolutely destroys the 7850K in gaming. It's not even in the same universe.

 

[Erenhardt]

So, first people complain 7850K CPU part is too slow. Now people recommend a whole bunch slower trinity athlon instead?
I'm not keeping up, give me a break!

 

[SPBHM]

There s barely 10% between the 7850k and a gt640 with whatever CPU,
not a difference worthy of a dgpu , indeed you have to upgrade your
recomendation up to a gt650 for the dgpu to really make sense....

yes, I would not recommend buying a GT 640 for gaming, and I have to say the same for the 7850K, when even the GT 640 is a little faster it's clear that pushing the 7850K as a good gaming solution is to much.

with the 650 GDDR5 and 7750 GDDR5 for $79-89 not to mention 7770s for around $100 it really makes 0 sense to buy a 630-640-r7250-7850K,

if you absolutely need to use IGP in your gaming desktop for some strange reason (probably small case), I also don't see the point of the 7850K when the 7600 is better balanced part.

if you don't need gaming and want a cheap solution it's hard to beat Haswell Pentium/Celeron, if you need a $120-$180 CPU for discrete graphics gaming the 4130 looks like the best choice at the moment, with IGP gaming the A8 7600 (when you can find it)

hey, if you use mantle maybe Celeron+r7 260x will be great for BF4 lol

So, first people complain 7850K CPU part is too slow. Now people recommend a whole bunch slower trinity athlon instead?
I'm not keeping up, give me a break!

Kaveri is pretty slow for $170
Trinity/Richland is good for $80