Separate names with a comma.
Discussion in 'CPUs and Overclocking' started by Tralalak, Aug 13, 2013.
my A8-5550M only scored 1.31 (seems it throttle at 64C) tested on windows 8.1 tech preview
Want to know power usage of this versus a A4-4000 platform in idle
Just for one more data point I ran cinebench on my alienware m11x r1. It has a dual core Core 2 duo Su7300 clocked at 1.73ghz. It scored identically to the 2ghz kabini in single core coming in at .49 with multi coming in at .93.
Whoever was saying this is so slow that people would be complaining to him about them I would say is very incorrect. I have used my m11x as my htpc for years since the hinge broke on it and it performs very well for multi-tab browsing, hd video, and light gaming. With the a6-5200 you are getting single-thread equivelant to that and twice the cores for your multi-thread.
As for my results on cinebench with the a6-5200 I did fear that some windows 8 processes may have been eating cycles as I had some weird microsoft thing that kept popping up 30% cpu usage randomly. Its possible that negatively impacted the cinebench on kabini. I just ran the tests on my su7300 and kept an eye on cpu usage, other than a few 1-2% processes all cpu was used by cinebench.
In other words Kabini is just below or equivelant to original core architecture in ipc(atleast in cinebench). In my opinion, very impressive.
So I looked up the specs, it is 64bit ddr3 memory. 64bit adapters have been around for 20 years. :thumbsdown:
But don't you think AMD dropped the ball on the GPU part of Kabini? they should have given it more shaders (384 at least) and dual channel memory. At full throttle power consumption would be increased but at similar performance, probably less due to lower GPU clock rates. A single channel memory setup also impacts CPU performance (10% back in the day).
Whether the OEM's fail to SKU dual channel is irrelevant - its up to sales and marketing to find competent partners.
Yes but its not native res and then you have the scaling 1024 x 768 to a 1368 x 768 screen resulting in additional blur and on low settings too. Visual mess.
No, Kabini is a low power low cost chip. AMD's financial state means that building more than one die wouldn't be very cost effective (they currently have three dies in production, FX, Trinity/Richland, and Kabini disregarding console) if they were going to have to cut it down so much for lower end chips.
Dual channel should have been implemented (atom z2760 has a dual channel controller and its a 2W chip). Small die size increase but massive igp gain (probably take it up about 20-40%)
That would require a massive die and would force AMD to raise prices (currently the a6-5200 costs ~$50-60: could be higher but I can't seem to find the source for pricing). 384 shaders for the most part would be CPU limited. TDP would massively increase. (384 GCN shaders is more powerful than the 384 VLIW4 shaders in richland/trinity making product canabalization a major factor given than the a10-5750m isn't that much faster than the a6-5200)
I dont think so. Its a cheap solution that really uses the single memory channel to the max and scales fine to ultraportables.
Anyway giving kabini that size of gpu would make the big core even more useless than it is and there would be no job for gf...
But 16nm version will get to the conflict here. 8 core and 512 on dual ddr4. What is left for piledriver derivatives then? And is this still produced at tsmc?
I agree with everything you have said except massive gain on GPU side.
Much more powerful IGP of Richland (A10-5750M) dropping about 50 percent of its performance with single memory channel.
It's hard to believe me that HD 8400 (A6-5200) which is just 1/4 of HD 7750 at significant lower core clock could gain 40 percent with dual MC (15-20 is possible IMO).
Only if they planned to completely cancel their mainstream APU line. Die salvaged Richland's would need to be trashed. Low end Kaveri will probably have those kinds of GPU specs (256-384 GCN). I do think they should have gone for dual channel memory though, less for consumer devices (OEMs put out plenty of Trinity systems with 1 stick of RAM) but for their embedded and server SKUs.
Although AMD has basic competitive performance problems, within that context they tend to put out well balanced APUs. Temash is just on the edge of thick tablet use, 192 GCN cores would push it completely out. Based on the salvaged E1-2xxx chips AMD is not interested in the cost and logistics of maintaining multiple jaguar die layouts.
Oems would not care about dual channel for they are only going to use 1 4 GB stick of ram in such a cheap system, thus it would be wasted die space and thus less profit for AMD. AMD is not going to add features that increase cost unless doing so increases average selling prices.
Atom z2760 has two 32 bit memory controllers which should give it similar memory bandwidth as a single 64 bit memory controller kabini uses
I wonder if adding 4 or 8 MB of eSRAM as a cache for the GPU would help any.
Note that even Richland mobile (with like 600-700 mhz clocks under gaming load probably lower given that desktop is much faster) loses massive amounts of performance with single channel RAM and show major gains with increasing RAM speeds (desktop increases near linearly with RAM speeds - note that trinity/richland's memory controller efficiency drops with faster speeds as seen in Sandra Tests- this is the main reason why scaling starts tapering at faster speeds, not because bandwidth is no longer a limiting factor) up to 2133 mhz we can see that trinity/richland is massively memory bound.
Clearly can see intel's memory controller scales better.
You can clearly see what happens when the 6670 (about the same performance as an a10 igp) has memory restrictions removed. 26% gain
Also remember that a 50% reduction from dual to single channel means a 100% gain from single to dual.
So in conclusion trinity/richland is very memory bound even with dual channel RAM at high speeds.
Kabini will also be memory bound especially considering that
1) Kabini systems will be more 'el cheapo' than piledriver systems for the most part. Single channel 'slow' RAM (probably one stick of 1066 or 1333 mhz RAM vs 1333 or 1600 mhz ram for piledriver). Dual channel would help.
2) Kabini, being a low power design may have an even crappier memory controller than piledriver. (Especially considering it supports relatively low speeds compared to piledriver).
3) The CPU is also using that bandwidth which given less and less bandwidth becomes a larger and larger relative chunk of bandwidth.
40% is a little high but in a lot of cases I'm thinking 20%+ easy.
Just realized this
If we take the 7730 GDDR5 (no memory restrictions) which has 384 gcn cores at 800 mhz and divide it by 3 you get about 50% of an a10-6800k (about equal to a 6670 w/ DDR3) reduce clock to 600 mhz (75%) and you get ~ 1/2 of a mobile richland chip. Considering that kabini (top model) is barely equal to if at all ULV i3 HD 4000 (a 25% gain in clockspeed from the a4-5000 toms/AT/techreport tested to the a6-5200 would probably get 15-20% fps gain) and mobile trinity easily beats ULV i3 by more than 50% you would probably see good gains with dual channel RAM
Personally I've yet to see a Kabini laptop with lower than DDR3-1600.
the stretching and scaling aren't as bad as you are picturing.
this pricing might only be for the g-series soc and not standard skus[in relatively large quantities] but it seems to be around ~$100
Kabini is ideal for sub USD 400 laptops. 11.6 inch and 13.3 inch are the ideal form factors of Kabini. Kabini can provide enough GPU power for running games like Skyrim at low settings.
mind you thats just a A4-5000. A6-5200 will be faster
also not everyone wants to play Battlefield 3 on a sub USD 400 laptop. ideal for light and causal gaming. excellent battery life provided the OEM fits a decent capacity battery. some like Acer screw it up by fitting a 30wh battery in the Acer aspire V5-122P. seriously you can't screw AMD worse than that. a tablet like iPad 4 has a larger 43wh battery.
AMD said the pricing of the G-series SoC was between $49-72 when it launched:
And I'd assume that the retail chips will be lower than that, considering the G-series are embedded parts.
The pricing difference is still pretty huge. For example, HP Pavilion 14 Sleekbook w/ A6-5200 @ $550 vs. Pavilion 14 Ultrabook w/ i3-3227U (recent IVB) @ $680. With everything else remaining the same, you pay $130 more for the i3 equipped model. Not worth it IMO.
I think HP's laptops are pretty shit, but I like that they're one of the few companies that actively include AMD parts in their lineup.
Btw that is excellent testing methology for games reflecting user experience. We could use that at AT. How many ms in 90sek does it go below eg. 30ms between frames.
Unfortunately, the manufacturers have not decided to make the A4-5000 available in anything smaller than a 15.6 inch laptop.
<- Still waiting for a 11.6 inch laptop with an A4 5000 because the A4 1250 is just too anemic.
I'm not sure where AT sits in this evolution of graphics testing, but personally I feel like this is where things need to go.
I play games on my laptop, not demanding games mind you, but the hitching and stuttering is what bothers me when it happen. A slow framerate is one thing, but I can't train my mind to ignore when the fps momentarily drop to near zero for a single frame.
I hope graphics reviews continue to evolve and embrace this new area of evaluating the smoothness of graphics technology.
When you are averaging numbers, as an FPS does, the average value itself is only meaningful if the standard deviation is quite small. What we all know, anyone who plays games, is that the standard deviation for our hardware is not small.
And that makes the averages rather meaningless.
99th percentile is the better metric IMO because it accounts for both extended frame rate drops and frame jitter. It's a shame few sites except TechReport use it.
edit: Maybe an easier metric to relate to is 'effective FPS' - ie, the lowest "frame rate feel" you can expect to encounter for extended periods of time. You could calculate this using the 99th percentile times.
eg, Battlefield 3:
Average FPS: 20 fps
99th percentile: 71.7 ms
Effective FPS: 13.9 fps
Average FPS: 12 fps
99th percentile: 155 ms
Effective FPS: 6.4 fps
Average FPS: 24 fps
99th percentile: 109 ms
Effective FPS: 9.2 fps
Kind of amazed Skyrim is playable at all even if at 720p at lowest settings. Looks pretty consistent to me in the 99th percentile+graph of milliseconds.
For what its worth, I know i am just some random guy on a tech forum, but I did not find gameplay smooth on my beefier a6-5200 laptop. Could have been a simple issue with the laptop I had, a personal increased sensitivity on my part, or any number of other things.
With that said if I were in your shoes I would feel inclined to believe techreport over myself- but it doesn't change the fact that in my real-world use Skyrim was too jerky for me to comfortably play. To each their own, just wanted to throw my 2 cents in.
In my opinion they need a turbo-boost option on all the jaguar based cpu's. Perhaps that would be enough to push it over the edge. Still badass that a $329 laptop can damn near play skyrim fluidly... Its kickass living in the future!
30fps average does mean it'll spend some time below 20fps, at which point it will be borderline unplayable. Skyrim is a terribly optimized game though. The A6-5200 should be averaging 35-38 fps I guess, with lows approaching 20.
That makes sense to me but i have no personal experience interpretating the results. Subjective test could be done to select the best method imho.
With time each person will know what his personal preference would be here. Do i need 15 or 45 of those seldom minimum fps.
Another diffucult aspect is where to test in the game. Take eg. Bf3. There is times on big maps with 64 players that cpu will be the limiting factor, but on far most maps and situation it tends to be gpu limited. Its quite important the games is tested under different situations or at least where many people play the game: eg. For bf3 Metro 64 players i guess is typical. And ofcource with plenty combat scenes.
Man this testing is difficult because testing consistently in multiplayer is next to impossible and will take tons of time.
I think anandtech is constantly evaluating the cost benefit of such meassures and as of today judges average is the best solution here. There is much more to it than we can know.
Hmm..., as we all have experienced average fps is difficult number to use in your evaluation. The easy and cheap solution is not there.
A huge number of SKU's don't even support 1600 mhz RAM.
(These are $400 laptops, no one is going to spend more money for faster RAM).
The problem with that approach is that it is using absolute numbers on relative levels of performance.
The second graph is absolutely meaningless.
Because the time spent beyond 33.3 ms is inversely proportional to fps. Higher fps means less time spent beyond 33.3 ms. It's not just a measure of consistency but also of fps (to measure consistency fps would have to be identical). Furthermore who is going to play the game at 22-26 fps? Settings will be turned down, fps will increase and time spent beyond whatever threshold will decrease.
What you ideally want is a measure of stutter per fps; a normalized measure of smoothness.