"Mysterious" AMD launch (Mobile Kaveri APUs under FX)

[Enigmoid]

Your correct that Intel has more efficient CPU, and AMD more efficient iGPU.

But even taking that into consideration, I think the AMD iGPUs are usually larger than the Intel ones. I.e. they devote more of the transistors and power consumption to the iGPU. The end result is higher iGPU performance on AMD APUs.

Also, the graph you showed does not display power usage, but energy usage to complete a certain task. So CPU A can have lower energy usage than CPU B, but CPU A can still have higher power consumption than CPU B. I.e. that would mean that CPU B takes longer time to complete the task than CPU A.

Finally, you only showed energy consumption for one type of workload. What does it look like for other types of tasks? And when comparing mobile CPUs (perhaps including AMD Beema/Mullins), instead of desktop CPUs as you did?

Igp performance is much better on AMD APUs on the mobile space though intel has significantly caught up. In a thermally limited environement the HD 4600 on a 35W SV DC part is for the most part between the a10-4600m and the a10-5750m, ie the two are largely equal.

Here is actual power consumption.

Right away you can see that for CPU tasks intel is using way less power comparatively per TDP.

I'm using efficiency because both ULV chips are going to be running 15/17W tdp so efficiency is what it is going to come down to. Don't forget HW ULV has the integrated PCH while ULV kaveri was simply binned.

A lot of real world power tests show something similar.

x264 is a pretty standard MT test. Here are some other tests.

CPU + igp load

http://www.bit-tech.net/hardware/cpus/2014/01/30/amd-a10-7850k-and-a10-7700k-kaveri-review/10

CB

http://cdn.pcper.com/files/imagecache/article_max_width/review/2014-01-13/power-load.png

Number of CPU, igp, and CPU + igp tests.

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

Again TDP != power consumption and from many reviews you can't compare intel TDP vs. AMD TDP directly.

 

[Homeles]

I like it
If it delivers or not, we will see soon.

I don't doubt it'll be a great mobile processor. The issue is whether or not it'll have any real availability... would really like to see those "ultrathins" come into existence.

 

[CrystalBay]

What is the IPC ? thats all that matters to me for desktop ?

 

[bunnyfubbles]

who says a 3ghz puma+ core can't "pleasurably" play games now?

and secondly have you seen mantle? I don't see how you can conclude that games will run subpar on such a theoretical apu? Mantle -and by extension dx12- will alleviate many bottelnecks in the graphics pipeline and allow for the possibility of better multithreading, which the MOAR COARS strategy can take advantage of.

so yeah I would sacrifice single threaded perf for more cores and better multitasking and multithreading!

you're just making all sorts of assumptions, like what kind of clocks it could run, or how many cores they could pack in, or that mantle/dx12 will truly be that efficient (spoiler alert, they won't - at best they'll reduce CPU overhead for the GPU, but if the game is a CPU demanding game to begin with, you still need a powerful CPU)

I go on what I know. If your solution was the obvious magic bullet, it would have been realized already. Otherwise I know that the steamroller cores are better than jaguar cores for a desktop PC situation, and that the vast majority of games (if not virtually all) will be better off with a few fast cores vs. many slower cores.

 

[Homeles]

What is the IPC ? thats all that matters to me for desktop ?

Why does IPC matter? IPS is all that matters.

 

[Enigmoid]

you're just making all sorts of assumptions, like what kind of clocks it could run, or how many cores they could pack in, or that mantle/dx12 will truly be that efficient (spoiler alert, they won't - at best they'll reduce CPU overhead for the GPU, but if the game is a CPU demanding game to begin with, you still need a powerful CPU)

I go on what I know. If your solution was the obvious magic bullet, it would have been realized already. Otherwise I know that the steamroller cores are better than jaguar cores for a desktop PC situation, and that the vast majority of games (if not virtually all) will be better off with a few fast cores vs. many slower cores.

+1

No dev is going to want to have to code the game logic for 32 "slow" cores vs. 8 "fast" cores. That would be a huge nightmare.

 

[bunnyfubbles]

Why does IPC matter? IPS is all that matters.

yeah, no kidding

if AMD could put out a 10GHz CPU with reasonable power levels that was 10% faster than the 3-4GHz intel CPUs we have now, that would be a wildly successful product despite having horrible IPC

 

[Enigmoid]

yeah, no kidding

if AMD could put out a 10GHz CPU with reasonable power levels that was 10% faster than the 3-4GHz intel CPUs we have now, that would be a wildly successful product despite having horrible IPC

Very true but the problem will be that the frequency increases will increase power consumption in a non-linear manner.

 

[Homeles]

Very true but the problem will be that the frequency increases will increase power consumption in a non-linear manner.

Well, he was exaggerating to make a point.

 

[Enigmoid]

Well, he was exaggerating to make a point.

Yep but high frequency designs haven't worked well for AMD (BD) or Intel (P4). IPC is important because it allows for lower clockspeeds at the same performance level which tends to increase efficiency. Its fairly easy to increase frequency but much harder to increase IPC.

IPC tends to decrease as frequency increases (ie clock a CPU really high and retrieving data from memory takes more cycles causing the pipeline to stall). If you have a choice, increasing IPC at the same frequency or increasing frequency at the same IPC increasing the IPC is almost always the better choice.

That said there are clearly exceptions. Haswell for example seems to be slightly less efficient than IVB despite power improvements. The increase in IPC didn't really result in an increase in efficiency.

 

[Homeles]

Yep but high frequency designs haven't worked well for AMD (BD) or Intel (P4). IPC is important because it allows for lower clockspeeds at the same performance level which tends to increase efficiency. Its fairly easy to increase frequency but much harder to increase IPC.

IPC tends to decrease as frequency increases (ie clock a CPU really high and retrieving data from memory takes more cycles causing the pipeline to stall). If you have a choice, increasing IPC at the same frequency or increasing frequency at the same IPC increasing the IPC is almost always the better choice.

That said there are clearly exceptions. Haswell for example seems to be slightly less efficient than IVB despite power improvements. The increase in IPC didn't really result in an increase in efficiency.

I disagree about Haswell. A lot of people see its slightly higher power consumption at slightly higher performance as a regression... and I just don't see it that way. If you were to keep the performance level flat compared to Ivy Bridge, power consumption would decrease.

That said, it's not remarkably more efficient for desktop workloads. A bit of that has to do with the regression in L3 latency. For mobile though, battery life is tremendously improved.

All said, it's a bit silly to bit silly to put an emphasis on either IPC or clock speed, when neither constitutes the full performance picture, regardless of difficulty in increasing one or the other. You need both.

 

[AtenRa]

Who cares about that? go and tell OEMs that, not me, E-350 was present on 11.6" netbooks, hell, before that they even used Neos that was CPU+NB+SB on +11.6".

C-XX where suppoused to be used on 10", but also offered on some 11.6" as a cheaper option.

When Jaguar came out, A4-1250 launched on 11.6", reemplacing E-350, A4-1200 launched on 10" reemplacing C-XX, like it or not, dont bother me with TDP numbers.

A6-1450 was the one that was good, but also very rare i think only Acer launched 1 model, and there was maybe one or two 11.6" with a A4-5000. You are trying to compare that with the widely used E-350 on 11.6"

I dont want to continue the Off topic but E-350 (18W TDP) was replaced by E-450 (18W TDP), this one was replaced by E2-1800 (18W TDP).

C60/70 were LOW POWER SKUs and as i have said earlier they were replaced by A4-1250 and all the other low power Kabini SKUs.
E-350 and later Brazos 2 18W TDP SKUs where replaced by 15W and 20W TDP Kabini SKUs.

Now, just because a low power A4-1250 Laptop cost the same as 18W TDP Laptop doesnt mean that OEMs replaced E-350 with A4-1250.

end of the OT

 

[Shivansps]

I dont want to continue the Off topic but E-350 (18W TDP) was replaced by E-450 (18W TDP), this one was replaced by E2-1800 (18W TDP).

C60/70 were LOW POWER SKUs and as i have said earlier they were replaced by A4-1250 and all the other low power Kabini SKUs.
E-350 and later Brazos 2 18W TDP SKUs where replaced by 15W and 20W TDP Kabini SKUs.

Now, just because a low power A4-1250 Laptop cost the same as 18W TDP Laptop doesnt mean that OEMs replaced E-350 with A4-1250.

end of the OT

Thats looks very nice, but it did not was what did happen, im telling you what DID happen, why? i have no idea, it could be price, availability, dont know, Kabini on 11.6" did not happen, there is 1 MSI and thats about all, there is not even Temash quad either, just 1 Acer. The only thing i know it was not about tdp, that was never the problem on 11.6".

And im not so sure about A4-1250 being any better than a C-70, if you look for results they give very similar results, both on igp and cpu.

BTW, i said E-350 because if was the one most widely used.

 

[Idontcare]

Yep but high frequency designs haven't worked well for AMD (BD) or Intel (P4). IPC is important because it allows for lower clockspeeds at the same performance level which tends to increase efficiency. Its fairly easy to increase frequency but much harder to increase IPC.

IPC tends to decrease as frequency increases (ie clock a CPU really high and retrieving data from memory takes more cycles causing the pipeline to stall). If you have a choice, increasing IPC at the same frequency or increasing frequency at the same IPC increasing the IPC is almost always the better choice.

That said there are clearly exceptions. Haswell for example seems to be slightly less efficient than IVB despite power improvements. The increase in IPC didn't really result in an increase in efficiency.

Higher power at the socket level because they moved a fair amount of power delivery from the mobo onto the package, but higher power for the system at the wall?

 

[Homeles]

Higher power at the socket level because they moved a fair amount of power delivery from the mobo onto the package, but higher power for the system at the wall?

It can be a bit higher at the wall, IIRC.

 

[monstercameron]

interesting theory

cduchesne
Quote:
Originally Posted by yuri View Post
Another Indian river.

It might be Pennar - a Jaguar APU which was mentioned by an oldish roadmap (along with Samara and Krishna).
The images are pictures of Kerala:

http://www.istockphoto.com/photo/kerala-3505758
http://www.istockphoto.com/photo/sun...-canoe-3579727
http://www.istockphoto.com/photo/hog...-falls-2769422


Kerala would signify AMD's ultrabook platform:
http://www.tomshardware.com/news/ult...ita,13800.html

http://semiaccurate.com/forums/showpost.php?p=214264&postcount=149

 

[TreVader]

A quad channel 8 core APU is what AMD needs. PS4 clocked to 3ghz with quad channel ddr3

 

[NostaSeronx]

interesting theory

https://ad.moontoast.com/amd/GULFSHORES-104/amd_kaveri_share.jpg
https://ad.moontoast.com/amd/GULFSHORES-104/amd_kaveri_share2.jpg
https://ad.moontoast.com/amd/GULFSHORES-104/screensaver/kv2-0.jpg
https://ad.moontoast.com/amd/GULFSHORES-104/screensaver/kv2-3.jpg

Read the image names the big reveal is Kaveri.

 

[Homeles]

A quad channel 8 core APU is what AMD needs. PS4 clocked to 3ghz with quad channel ddr3

Unfortunately, quad channel is completely useless for mobile.

 

[NostaSeronx]

Unfortunately, quad channel is completely useless for mobile.

It makes sense the faster the RAM you need.

DDR4 for example;
256-bit: DDR4-2132 1.05v or DDR4-2666 1.2v
vs
128-bit: DDR4-4264 1.2v or GDDR5M-5000 1.35v

 

[Homeles]

It makes sense the faster the RAM you need.

DDR4 for example;
256-bit: DDR4-2132 1.05v or DDR4-2666 1.2v
vs
128-bit: DDR4-4264 1.2v or GDDR5M-5000 1.35v

The huge system board area penalty is a tough pill for OEMs to swallow.

Thankfully, it's one they won't have to. DDR4 in a dual channel configuration will provide enough bandwidth for AMD to get by until stacked memory is cheap enough, and Intel has their eDRAM.

 

[Blitzvogel]

A quad channel 8 core APU is what AMD needs. PS4 clocked to 3ghz with quad channel ddr3

Not a good idea at the moment. I think AMD is right to focus on small scale 4 core APUs like Beema and Mullins, especially that it can be scaled to fit tablets, up through notebooks and small desktops with ease. My only qualm so far is the single channel memory and the lack of more graphics CUs in the upcoming Beema and Mullins APUs.

The high end APUs like Kaveri I think are at a good niche to go head to head with the notebook i3 and i5 lines and desktop i3 processors. The problem is that AMD hasn't hit the power efficiency and IPC targets they need to hit to be evenly competitive with Intel. Graphics are the only thing going for them and most consumers don't care.

As for the truly high end, AMD can't get their new core out soon enough. I think CMT turned out to be an awful direction for AMD. It ties with your 8 core desire in that it's high amount of multithreading didn't mean diddly-squat for the most part when pure Intel quad cores clocked less than AMD quad-modules still took home the bacon, at much lower power consumption.

 

[NTMBK]

The huge system board area penalty is a tough pill for OEMs to swallow.

Thankfully, it's one they won't have to. DDR4 in a dual channel configuration will provide enough bandwidth for AMD to get by until stacked memory is cheap enough, and Intel has their eDRAM.

Is the board area really that massive, if you solder down your memory chips? The XBox One motherboard looks like this:

Yes, a lot of area is used in memory traces, but that was area which is mostly hidden under the heatsink anyway.

 

[Homeles]

The CPU+RAM+power delivery portion of that board is essentially the size of a main logic board in a MacBook Pro.

Anyway, even if I'm wrong on the space penalty thing, the cost penalty is very real.

 

[sm625]

As for the truly high end, AMD can't get their new core out soon enough. I think CMT turned out to be an awful direction for AMD.

CMT was only bad in the sense that it diverted resources away from improving power management and improving IPC (profiling, etc). AMD's implementation of CMT really isnt that bad, since two threads run almost as fast as one. The problem is that a single thread just doesnt run all that fast.