The Kaveri Pre-Launch Thread (A10-7800 and A10-6800k @3,5 Ghz)

[mikk]

http://translate.google.de/translat...u=http://tieba.baidu.com/p/2795083406&act=url


Not the greatest CPU benchmark selection to be honest. In gaming benchmarks Kaveri is all over the place. From +17% to -5%. Maybe a TDP issue (65W).

Closed. The reviews thread is up here
-ViRGE

 

[SiliconWars]

Whoah that's a big improvement in Sandra and WPrime 32 (at the same clocks). Are they claiming TSMC for the process?

 

[inf64]

Not the best GPU drivers but mem. BW bottleneck is pretty obvious.
Also Kaveri in ASrock supposedly performs better than in MSI boards.

 

[Homeles]

"Home at last! The first thing fat "Computer" on kaveri apu full version evaluation"

Lol. Google translate is funny.

If you're having issues with viewing the pictures because of Google Translate, here's a direct link: http://tieba.baidu.com/p/2795083406

Woo! Look at that SuperPi score! AMD's best yet

In all seriousness, is that because of the 3 FPU stages, down from 4?

Gaming results look mixed. Even if Kaveri's bandwidth starved, we should be seeing a clear improvement. That definitely has to be throttling.

 

[ancapdev]

Imgur album of all the benchmarks here (was posted on SemiAccurate, not mine): http://imgur.com/a/OYWL0#0

For those who don't have a baidu account, it's all discussion past page 1 anyway, no more data.

 

[monstercameron]

this is quite disappointing!

 

[Homeles]

We're seeing that regression in Cinebench again... so those leaked numbers seem like they're legit (as inf64 was saying).

But yeah, I think we're seeing throttling here. The 95W part should look much better. All things considered, this 65W part doesn't look bad next to a 100W part.

 

[inf64]

Well per clock it is faster, in C11 around 11%.
I see an issue in their results tho, it seems their sample was throttling in C11.5 when at stock(with Turbo enabled):

With fixed clock of 3.5Ghz their sample actually performed better:

Maybe this is the culprit in some other benchmarks. The iGPU regressions are driver and BW related. Some things one can fix and the others he can't.

 

[mikk]

Throttling seems unlikey in CPU only tests with a fixed 3,5 Ghz frequency.

 

[SPBHM]

not a bad gain on CB, but same score on 3dmark11?

 

[inf64]

Throttling seems unlikey in CPU only tests with a fixed 3,5 Ghz frequency.

Not in the run with fixed clock... Fixed clock has HIGHER benchmark score.
It throttled in this run (stock with Turbo Enabled) :

There are other benchmarks apart from the ones with fixed clock (pure x86 benchmarks). I was referring to those (like Musemage and Fritz).

 

[SiliconWars]

But yeah, I think we're seeing throttling here. The 95W part should look much better. All things considered, this 65W part doesn't look bad next to a 100W part.

The problem is they won't get much extra performance from the extra 30W they have on the 95W Kaveri's. Less than 10% in all probability.

They've done ok to beat Richland at all I think but I'm pretty sure they would have been expecting more on the CPU side. They can't do anything about the iGP, that's just completely bottlenecked by bandwidth.

The good news is that the process should have a little bit of room to grow and we might see similar clocks to the 6800K in around 6-9 months time.

 

[monstercameron]

it seems that there are no surprises, another meh release from AMD. Lets just hope mantle and hsa can compensate.

 

[Sweepr]

this is quite disappointing!

Indeed it is. This kind of perf/clock improvement probably wont be able to touch 6-8 core Vishera performance in recent/future games (with a discrete GPU), especially if hits lower clocks @ OC.

There was a lot of buzz about 30% performance improvement @ ST workloads over the past few months from the usual suspects, that estimate is quicky changing to ''I''m fine if it beats Richland at lower clocks at all'' since the first results came out.

Anyway, better wait for the actual reviews.

 

[Enigmoid]

Can someone paste the standalone link?

This is all I see.

 

[inf64]

Here you go:
http://tieba.baidu.com/p/2795083406

Note this part in the review is 65W A10 7800 (3.5-3.9Ghz). It apparently throttled in C11.5 (that we know) when it ran at stock with Turbo enabled.

edit:
OR Turbo doesn't kick in at all and it ran it at stock . Difference is negligible between these two runs.

 

[mrmt]

Lets just hope mantle and hsa can compensate.

I think you need more than hope when talking about HSA. AMD didn't even fix issues with their OpenCL SDK, let alone deliver a working HSA SDK.

 

[ElFenix]

so, when might we expect actual reviews from reputable sites?

 

[mikk]

3.49 points for the A10-6800k in Cinebench 11.5 is a bit too slow by the way. In other reviews it scored ~3.6.

 

[Ventanni]

Good benchmarks, but kinda hard to trust a source that states Kaveri is made at TSMC.

 

[inf64]

3.49 points for the A10-6800k in Cinebench 11.5 is a bit too slow by the way. In other reviews it scored ~3.6.

Difference between those two is ~3%, margin of error.

 

[Homeles]

so, when might we expect actual reviews from reputable sites?

Next week is CES, from the 7th through the 10th. So sometime then for "legit" numbers to pour out. The product launches on the 14th, so full blown reviews may not be out until then.

Difference between those two is ~3%, margin of error.

I forgot that this review has version 11.5 of Cinebench... the previous leak was version R15.

 

[mikk]

Difference between those two is ~3%, margin of error.

In Cinebench this margin is too big.

 

[inf64]

It's variable Turbo core, just a slight clock drop can cause it (higher temps due to worse cooling and/or higher ambient temps) or some background task. Also we can argue the same might be true for Kaveri part. Don't pay attention to small things like that, it's pointless.

 

[avtek21]

Whoah that's a big improvement in Sandra and WPrime 32 (at the same clocks). Are they claiming TSMC for the process?

From the translation and results, the Sandra financial options pricing overall performance of 190 kopt/s would seem to put the A10 Kaveri GPU compute speed as fast as a NVidia Titan!

Also, wouldn't testing under Win 8.1 add another 5-10%.