AMD CPU speculation... and expert conjecture

[jdwii]

When(if) a new forum is created their needs to be a chart on the first post that allows people to fully understand BS.
1. up to statements
2. Who posts the benchmarks or info
3. the type of benchmark

 

[gamerk316]

^^ Frankly, not impressed. Looking at the first chart, a combined CPU/GPU upgrade is only buying a few FPS at low settings. Granted, AMD still crushes intel in the iGPU deparment.

Second chart shows Intel still winning on CPU performance head to head by a decent margin. Not shocked in the least.

Third one shows Kaveri starts to reach max potential at 1866MHz RAM. Some gains going faster, but for most of the tests, <3 FPS gains from that point onward. More likely secondary improvements at that point, rather then the GPU still being starved. Clearly no issues with Kaveri getting starved on DDR3.

I think most here are/were expecting a bigger improvement then those charts indicate.

Also FYI: Source?

But...But 30% more IGPU power lol if anything my estimates where to much WTF Amd. 512 GCN cores really? look at the memory test as predicted the memory bandwidth is bottlenecking the GPU to much.

Not really. Benchmark showed some gains going up from 1833, but for the most part, you don't see large increases after that point. So at 1833, you aren't significantly RAM bottlenecked.

You can say 30% more GPU power, but if the CPU that drives it is the bottleneck, you won't get to use it. Which could very well be what we're seeing. On the game side tests:

(Note: Review is down, but image links are still up. Someone may want to mirror before they get taken down. Just change the id="xxxxx" up and down to go through the images; they're sequential)

Unigene Heaven Pro 4.0:
Richland: 51
Kaveri: 53.6
% increase: 4.85%

Alien Vs Predator
Richland: 51.2
Kaveri: 56.3
% increase: 9.06%

Batman: Arkham City
Richland: 56
Kaveri: 60 [Possible game limit?]
% increase: 6.67%

Just Cause 2 - Concrete Jungle
Richland: 51.5
Kaveri: 59.4
% increase: 13.3%

Hitman: Absolution
Richland: 40.8
Kaveri: 47.7
% increase: 14.47%

Street Fighter IV Arcade Edition
Richland: 55
Kaveri: 62
% increase: 11.29%

Average %Increase: 9.94%

So yeah, 15% looks mighty generous, based on actual benchmarks. Nevermind the fact AMD told Pudget Systems to take down the review, which is a giant warning signal that they might be spot on.

EDIT

I stress these are combined CPU/GPU scores, not just CPU improvements. Two of the tests (Batman and STIV) may also be hard locked at 60 and 62 FPS, resulting in understated score [IDK if these FPS are limited for these games or not]. Removing them doesn't change the numbers much though.

 

[de5_Roy]

When(if) a new forum is created their needs to be a chart on the first post that allows people to fully understand BS.
1. up to statements
2. Who posts the benchmarks or info
3. the type of benchmark

no way, jdwii. take a moment to consider the magnanimous efforts of those people who spend their invaluable time and energy to misrepresent data to make false claims i.e. the BSers. and corporations and their p.r. divisions. i guess now you will campaign for adding how to protect oneself from fallacies, red herrings, using of one bottleneck to hide another in benchmarks, over-inflation of performance improvement by comparing against the lowest performing sku, not posting minimum fps/frame time variance, compiler engineering, poorly coded benchmark softwares as well, hmm? how insensitive.

 

[-Fran-]

Given the charts with the 2133 and 1866 memory... Can that little difference be attributed to the IMC instead of the GPU change? I'm going on a limb here, since I do know the IMC is CPU dependant and also AMD doesn't have the best implementation because Intel+reasons.

Also, I'm wondering if the hUMA support is "active" right off the bat there or if we'll have to wait 'till HSA support is made official. Or maybe I have something wrong here with my understanding of things, haha.

Cheers!

 

[logainofhades]

I am basing my comments regarding 750k vs PhII on the fairly recent THG article comparing AMD chips. CPU boss is a poor place to get real results imo. On average, as can be seen in the images I linked from the end of the review, PhII 965 beats the 750k. Trinity and Richland are both Piledriver based just like Vishera FX's. It is pretty safe to say the lack of L3 cache is what holds Trinity and Richland back. If they had L3, realistically, they wouldn't perform any different than a quad core FX Vishera at the same clock speed. Also, if it isn't from THG, Anandtech or another respectable review site, I don't fully trust the results.

http://www.tomshardware.com/reviews/piledriver-k10-cpu-overclocking,3584.html

 

[gamerk316]

Given the charts with the 2133 and 1866 memory... Can that little difference be attributed to the IMC instead of the GPU change? I'm going on a limb here, since I do know the IMC is CPU dependant and also AMD doesn't have the best implementation because Intel+reasons.

You see a huge jump from 1600 to 1866, but maybe 3FPS going from 1866 ro 2133. So while the RAM may be limiting a little, it could also be just noise. Don't think its a major factor above 1866, and certainly not causing the giant loss in performance some people are claiming.

 

[-Fran-]

You see a huge jump from 1600 to 1866, but maybe 3FPS going from 1866 ro 2133. So while the RAM may be limiting a little, it could also be just noise. Don't think its a major factor above 1866, and certainly not causing the giant loss in performance some people are claiming.

Its just that I've read OCing the IMC is possible and actually helps in that regard fro APUs (I haven't tried with mine TBH). OC'ing the CPU helps as well. If it's not an IMC problem, then OC'ing the GPU should show big improvements with 2133 as well as 1866...

I really hope Toms makes a deep investigation when they get their hands on Kaveri. Are you reading Mr. Angelini? XD

And in that note, they could make an article about "fine tunning" the APUs. Or palladin should, haha.

Cheers!

 

[noob2222]

Gpu scores can be attributed to the core clock. 6800k gpu runs at 844 mhz where the 7850k is only running 720 mhz. Taking that into account, thats a 17% slower clock for 9% more performance. That aspect is pretty impressive on the IGP. If only they could have hit their target speeds kaveri would look a whole lot better, but bulk has its problems.

P.s. I stand corrected on the 750k, but without the "hsa improvements" on the gpu, the 750k and 760k are essentially the same die with the exception that the 760k clocks better. The cpu cores are identical.

 

[logainofhades]

750k and 760k do not have a GPU to improve on. :lol: Had you been talking 5800k to 6800k, your P.S. statement would have made sense. 😛

 

[jdwii]

When(if) a new forum is created their needs to be a chart on the first post that allows people to fully understand BS.
1. up to statements
2. Who posts the benchmarks or info
3. the type of benchmark

no way, jdwii. take a moment to consider the magnanimous efforts of those people who spend their invaluable time and energy to misrepresent data to make false claims i.e. the BSers. and corporations and their p.r. divisions. i guess now you will campaign for adding how to protect oneself from fallacies, red herrings, using of one bottleneck to hide another in benchmarks, over-inflation of performance improvement by comparing against the lowest performing sku, not posting minimum fps/frame time variance, compiler engineering, poorly coded benchmark softwares as well, hmm? how insensitive.

For 1 yes i do agree we should educate tomshardare readers about fallacies, and yes we should include in the reviews how bad some software is coded.

 

[Cazalan]

So now we need some OC numbers. How much does DDR3-2400 help?
Richland seemed to peak at 2133.

 

[esrever]

those tests are fake, they came from a no name site

 

[-Fran-]

I think your referring to this, http://www.tomshardware.com/reviews/a10-5800k-a8-5600k-a6-5400k,3224-5.html

In the SiSoft Sandra tests you’ll see shortly, the Trinity-based APUs yield less memory bandwidth from our 16 GB DDR3-1600 kit than Llano.

But that's RAM scaling using just the RAM speed for it. I'm talking about on the other side of the equation. Man, where are you palladin when we need you? I have a hard time explaining it, lol

Cheers!

 

[juanrga]

I could see the Pudget review before they were obligated to retire it (flagrant NDA violation). It seems to have some configuration and software problems. I asked to someone who has information about Kaveri performance (he is preparing one review) and his response was: "this review is pure bullshit"

I find it weird that each time that some leaked benchmark appears suggesting a good improvement for Steamroller, another benchmark is leaked showing the contrary: The first CB score, the second CB score, the Chinese review, now the Pudget review... It is a ping-pong game. I wonder who are the players ;-)

 

[tracker45]

Let's not forget : bf4 comes bundled with Steamroller, which is a good bonus.

 

[Cazalan]

I could see the Pudget review before they were obligated to retire it (flagrant NDA violation). It seems to have some configuration and software problems. I asked to someone who has information about Kaveri performance (he is preparing one review) and his response was: "this review is pure bullshit"

I find it weird that each time that some leaked benchmark appears suggesting a good improvement for Steamroller, another benchmark is leaked showing the contrary: The first CB score, the second CB score, the Chinese review, now the Pudget review... It is a ping-pong game. I wonder who are the players ;-)

That PugetSystems company actually sells a lot of AMD products (including water cooled Kaveri config) so it's not really in their interest to make it look bad. They could have old drivers or something.

For those that didn't see the graphs they were quickly copied here: http://wccftech.com/amd-a10-7850k-performance-previewed/

 

[etayorius]

Let's not forget : bf4 comes bundled with Steamroller, which is a good bonus.

Really? Damn! i hate the retarded FPS Wargames... but still, most people seem to love them.


That PugetSystems company actually sells a lot of AMD products (including water cooled Kaveri config) so it's not really in their interest to make it look bad. They could have old drivers or something.

For those that didn't see the graphs they were quickly copied here: http://wccftech.com/amd-a10-7850k-performance-previewed/

[/quotemsg]

They need non released Kaveri Catalyst Drivers in order to really see the true performance from Kaveri, just installing Kaveri on a new Mobo wont cut it, so yeah most surely.

 

[anxiousinfusion]

Let's not forget : bf4 comes bundled with Steamroller, which is a good bonus.

Really? Damn! i hate the retarded FPS Wargames... but still, most people seem to love them.

Yes, Kaveri will come with free Battle Medal Duty of Field War Honor: Mountain Dew Ops Dudebro Edition 3.

I actually hate that genre, please don't hurt me.

 

[etayorius]

Let's not forget : bf4 comes bundled with Steamroller, which is a good bonus.

Really? Damn! i hate the retarded FPS Wargames... but still, most people seem to love them.

Yes, Kaveri will come with free Battle Medal Duty of Field War Honor: Mountain Dew Ops Dudebro Edition 3.

I actually hate that genre, please don't hurt me.

I dont like such games... but that does not means no one else does, in fact... i could easily get 30-40 bucks by selling the damn Origin Key.

 

[noob2222]

@logain

Its a technicality. The 750k and 760k do have an IGP. Its disabled due to not passing functionality tests. The overall untested die of the 750k = 5800k and 760k = 6800k as well as all the other "bins". In that sense its richland vs trinity and without the IGP, its pd vs pd.

Microscopic details is what makes the difference.

 

[-Fran-]

Yes, Kaveri will come with free Battle Medal Duty of Field War Honor: Mountain Dew Ops Dudebro Edition 3.

I actually hate that genre, please don't hurt me.

But the Bromance is the best in those games. You should've seen how I was crying when Soap died. Manly tears all over the place while killing for revenge.

And in regards to the release. I really hope Toms has a deep analysis on the software part of Kaveri, since it's the biggest selling point for it. The little there is, must be analyzed IMO. And of course tweaking.

Cheers!

 

[griptwister]

Let's not forget : bf4 comes bundled with Steamroller, which is a good bonus.

Really? Damn! i hate the retarded FPS Wargames... but still, most people seem to love them.

Yes, Kaveri will come with free Battle Medal Duty of Field War Honor: Mountain Dew Ops Dudebro Edition 3.

I actually hate that genre, please don't hurt me.

Hahaha! This is fantastic! I look forward to selling BF4. I got it for $10 off a friend.

 

[con635]

Which software or type of software will suit hsa? Will any be available at kaveri launch? I thought hsa/huma was the real improvement here? I also thought the whole point of kaveri was the igp wasnt sitting idle if using a discrete card as it could 'help' with cpu tasks that suit. All I've seen is face recognition and office software that I wont use, I fear wee wont see the true benefit of this cpu until late 2014, but the hsa enabled benches will be funny if the marketing is even half right!
Fwiw the iris pro is only for choice oems, right? If thats right only fair to compare it to an oem only amd apu, like say the one they did for sony 😛
Also in passmark 3d bench my hd7560d @ 844mhz scores near the same as average iris pro score, 927 vs 950, I'm on win7 as well....
edit: Don't diss fps games, they're actually good for your brain apparently, I love them

 

[palladin9479]

So basically kaveri will finally allow an APU to reach the phenom II x4 level of performance.

I think your referring to this, http://www.tomshardware.com/reviews/a10-5800k-a8-5600k-a6-5400k,3224-5.html

In the SiSoft Sandra tests you’ll see shortly, the Trinity-based APUs yield less memory bandwidth from our 16 GB DDR3-1600 kit than Llano.

But that's RAM scaling using just the RAM speed for it. I'm talking about on the other side of the equation. Man, where are you palladin when we need you? I have a hard time explaining it, lol

Cheers!

Been on vacation in Thailand, still here but leaving soon.

Yes OCing the CPU_NB gives a pretty nice boost to memory subsystem performance while having negligible impact on the integer / scalar processing speed. OCing the CHIP_NB will yield next to no actual performance difference, neither will OCing the HTT link. Currently AMD's iGPU's are largely limited by memory bandwidth and the speed of the IMC. Due to the IMC, CPU and iGPU all using the same TDP there is a large amount of room for custom CPU settings based on what you want to do with it. Also remember memory speed is important when it's being used, so resolution and memory intensive settings (AA/AF) are where you'll see most of your improvements.

Now about PH II vs APU's. Llano was a Phenom II, that uArch hit a large wall when it came to scalability. L3 is a bogey monster that's frequently used to justify performance differences. L3 is only useful if your dataset is larger then L2 cache size. PHII had 512KB of L2 per core, BD has 2MB of L2 per module with 1MB per core. L2 has a much greater impact on performance then L3, typically by an entire order of magnitude. When comparing the two design's, the presence of lack of L3 would have a very little impact. The real difference between K10 and BD is in the ALU's and L2 cache system. K10 had 512KB of dedicated L2 cache per core and 3 ALU's per core. BD has 2MB of shared L2 per module and 2 ALU's per core. So your looking at 66% of the integer power and a less efficient L2 cache system as the primary culprits for the speed difference. The proper comparison for the PH2 would be the 6x series CPU. People are placing WAY to much important on a sluggish L3 cache, seriously AMD's caching system isn't good enough to make L3 important. Intel on the other hand use's a radically different cache architecture, it's L3 acts far more like an extended L2 cache.

 

[juanrga]

I could see the Pudget review before they were obligated to retire it (flagrant NDA violation). It seems to have some configuration and software problems. I asked to someone who has information about Kaveri performance (he is preparing one review) and his response was: "this review is pure bullshit"

I find it weird that each time that some leaked benchmark appears suggesting a good improvement for Steamroller, another benchmark is leaked showing the contrary: The first CB score, the second CB score, the Chinese review, now the Pudget review... It is a ping-pong game. I wonder who are the players ;-)

That PugetSystems company actually sells a lot of AMD products (including water cooled Kaveri config) so it's not really in their interest to make it look bad. They could have old drivers or something.

For those that didn't see the graphs they were quickly copied here: http://wccftech.com/amd-a10-7850k-performance-previewed/

Well, the chinese review also came from a source closely aligned to AMD. AMD sell lots of products in that zone and even make customs gifts for them...