AMD A10-4600M Review: Mobile Trinity Gets Tested

[Wisecracker]

AID64/Everest might work as a backstop to Sandra.

Add 256-512 MB of GDDR5: No competition for Trinty in GPU performance.

... A big part of the IGP here is for it to manage in GPGPU workloads to make up for the inferior CPU performance compared to it's competition. Having 1GB instead of 512MB would help that...

Maybe... However, I'm not sure. If AMD support it in Trinity's successor, then it's a recipe for diminishing the memory bandwidth bottleneck greatly. If the next APUs are 28nm with GCN cores, then it could be pretty important to have more memory bandwidth, even more so than it is today...

It's not part of the progression/transition, doesn't fit, or is being left behind.

Just be patient Baby Steps

There is no frame buffer or VRAM. The Turks graphic cores represent the 3rd step in direct memory addressing - primarily for OpenCL, of course. HD Graphics (I think) are hung off the ring bus and share 'last level cache' prior to the memory controller. Turks has (as presumably Cape Verde GCN cores) DDR3 128-bit memory interface to a (2x64-bit) 128-bit memory controller.

Singles, not home runs!

If yah wanna go all (F)X Files on the memory, bandwidth, unified controller/address space, etc., if you want Cape Verdi massive clock 1125/4500MHz GDDR5-type 72GB/s GCN-cores bandwidth, you first sandwich 2xIMC into a 256-bit (4x64-bit) quad-channel memory controller. Like Interlagos, as a matter of fact (!). Even more interesting when looking at the Bulldozer FPU and how it will handle SIMD AVX instructions.

Then, it gets really interesting. A single stick of DDR4 2250 (a double-pumped 1125MHz) should handle 18 GB/s. Being that you have a quad-channel IMC and point-to-point DDR4 . . .

Yah got 72GB/s [:lutfij:4]

not interested in what AMD writes in their press breifs. I want to see the numbers. I read the article and saw all the charts and I want more.

And? You want it, now? Since you asked nicely, here's some Fun With Slide 42 - Dual Graphics. . .

If OEMs don't build mobiles with Trinity G2 Dual Graphics, they can't be tested.

 

[blazorthon]

AID64/Everest might work as a backstop to Sandra.



It's not part of the progression/transition, doesn't fit, or is being left behind.

Just be patient Baby Steps

There is no frame buffer or VRAM. The Turks graphic cores represent the 3rd step in direct memory addressing - primarily for OpenCL, of course. HD Graphics (I think) are hung off the ring bus and share 'last level cache' prior to the memory controller. Turks has (as presumably Cape Verde GCN cores) DDR3 128-bit memory interface to a (2x64-bit) 128-bit memory controller.

Singles, not home runs!

If yah wanna go all (F)X Files on the memory, bandwidth, unified controller/address space, etc., if you want Cape Verdi massive clock 1125/4500MHz GDDR5-type 72GB/s GCN-cores bandwidth, you first sandwich 2xIMC into a 256-bit (4x64-bit) quad-channel memory controller. Like Interlagos, as a matter of fact (!). Even more interesting when looking at the Bulldozer FPU and how it will handle SIMD AVX instructions.

Then, it gets really interesting. A single stick of DDR4 2250 (a double-pumped 1125MHz) should handle 18 GB/s. Being that you have a quad-channel IMC and point-to-point DDR4 . . .

Yah got 72GB/s [:lutfij:4]



And? You want it, now? Since you asked nicely, here's some Fun With Slide 42 - Dual Graphics. . .

http://i716.photobucket.com/albums/.../Trinity_Dual_Graphics_projected-FarCry_0.jpg

http://i716.photobucket.com/albums/...y/Trinity_Dual_Graphics_projected-Dirt3_0.jpg

http://i716.photobucket.com/albums/...ty/Trinity_Dual_Graphics_projected-Civ5_0.jpg

http://i716.photobucket.com/albums/...y/Trinity_Dual_Graphics_projected-Metro_0.jpg

If OEMs don't build mobiles with Trinity G2 Dual Graphics, they can't be tested.

Trinity doesn't have a Turks GPU, it has a custom VLIW4 GPU. Also, we weren't saying that it had it's own frame buffer, but that having it's own frame buffer would help it a lot before DDR4 comes out.

 

[assassin123]

hmmm. . . . . . . Nice

 

[ojas]

[citation][nom]blazorthon[/nom]Intel's graphics gets weaker on lower end models... For example, the HD 3000 on the i7s is FAR faster than the HD 3000 on the i3s and is considerably faster than the HD 3000 on the i5s (although even within each family, there can be differences, all of this is because although they have the same graphics hardware, the clock frequency of the IGP differs). The same is probably true for the HD 4000. The cheaper i5s and i3s will probably have weaker graphics performance than the top i5s and the i7s do.[/citation]
I'm somehow not sure Intel can afford to do that anymore...

 

[cleeve]

It has to do being a journalist and being impartial.

Your logic is horribly flawed.

Just because because we see the value in testing in Dx9, and you do not, that doesn't make us impartial.
It simply means you don't agree with our methods.

DirectX 9 remains the most used code path. It's usually the least demanding, too, and with IGPs that's the name of the game.
I'm not going to bench unplayable settings that no-one will use in the real world. That's worthless.

 

[blazorthon]

[citation][nom]ojas[/nom]I'm somehow not sure Intel can afford to do that anymore...[/citation]

How could Intel have trouble affording to lower the IGP frequency of lower end processors? That's ridiculous.

 

[alphaalphaalpha]

[citation][nom]Avro Arrow[/nom]You know, I just realised something. Looking at the above chart, Trinity is about 50% better than Llano but it's also clocked 50% higher. Is it just me or is Trinity's IPC roughly the same as Llano?[/citation]

Piledriver seems excellent at scaling up frequencies without increasing power usage a whole lot. Bulldozer is pretty good at this too. For example, the FX-4100 is at 3.6GHz and it uses less power than most of the considerably lower clocked Phenom II processors, even if it doesn't out-perform them. Piledriver's performance per Hz wasn't supposed to beat Llano, it was just supposed to beat Bulldozer and that it did. It's higher clock frequencies seem to do the job of improving performance here very well without increasing power usage, despite Trinity and Llano both being on the same 32nm process node.

 

[palladin9479]

How could Intel have trouble affording to lower the IGP frequency of lower end processors? That's ridiculous.

He's talking about competition wise. As the competition heats up in the $700 and below segment Intel will have up their GPU performance. Right now people are comparing HD4000 on an i7 from $1K+ laptops to the APU inside a $700 laptop, and ATI still typically wins. This is all without even touching driver support, something that is harder to get right then just making hardware. Intel's new to the GPU scene and hasn't had time to stock pile the sheer amount of information that NVidia / ATI have for optimizing their drivers / Capability profiles / ect. They can catch up eventually, it's gonna be one bumpy road though.

 

[blazorthon]

He's talking about competition wise. As the competition heats up in the $700 and below segment Intel will have up their GPU performance. Right now people are comparing HD4000 on an i7 from $1K+ laptops to the APU inside a $700 laptop, and ATI still typically wins. This is all without even touching driver support, something that is harder to get right then just making hardware. Intel's new to the GPU scene and hasn't had time to stock pile the sheer amount of information that NVidia / ATI have for optimizing their drivers / Capability profiles / ect. They can catch up eventually, it's gonna be one bumpy road though.

Intel hasn't cared about their IGPs being the greatest before and even now, they still don't. If their IGP is insufficient, then they'll just throw in a Nvidia card.

 

[ojas]

[citation][nom]palladin9479[/nom]He's talking about competition wise. As the competition heats up in the $700 and below segment Intel will have up their GPU performance. Right now people are comparing HD4000 on an i7 from $1K+ laptops to the APU inside a $700 laptop, and ATI still typically wins. This is all without even touching driver support, something that is harder to get right then just making hardware. Intel's new to the GPU scene and hasn't had time to stock pile the sheer amount of information that NVidia / ATI have for optimizing their drivers / Capability profiles / ect. They can catch up eventually, it's gonna be one bumpy road though.[/citation]
Thank you!

[citation][nom]blazorthon[/nom]Intel hasn't cared about their IGPs being the greatest before and even now, they still don't. If their IGP is insufficient, then they'll just throw in a Nvidia card.[/citation]
Intel either bins their parts or differentiates based on clock speeds on perfectly good parts. For their ultrabooks, they're looking to push the IGP. They keep going on about how the new notebooks have the most brilliant visuals ever and bla bla bla. Plus the moment you put in an Nvidia GPU, your battery life will take a hit.

Plus all the media hype about how HD4000 is killing the low end GPU segment.

Now you have Trinity. To simply match this stuff Intel will have to put a full, uncut, unrestrained GPU in processors (is it safe to call sandy and ivy bridge processors APUs? They pretty much fit the definition) all the way down to the $700 segment, HD2500 won't cut it. $700 isn't really i7 territory.

i5s will have to get the same IGP (in terms of clocks as well), even the i3s can't get HD2500 and stay competitive.

Same goes for the desktops. i3s will HAVE to get HD4000 if Intel cares about that market at all.

 

[blazorthon]

[citation][nom]ojas[/nom]Thank you! Intel either bins their parts or differentiates based on clock speeds on perfectly good parts. For their ultrabooks, they're looking to push the IGP. They keep going on about how the new notebooks have the most brilliant visuals ever and bla bla bla. Plus the moment you put in an Nvidia GPU, your battery life will take a hit.Plus all the media hype about how HD4000 is killing the low end GPU segment.Now you have Trinity. To simply match this stuff Intel will have to put a full, uncut, unrestrained GPU in processors (is it safe to call sandy and ivy bridge processors APUs? They pretty much fit the definition) all the way down to the $700 segment, HD2500 won't cut it. $700 isn't really i7 territory.i5s will have to get the same IGP, even the i3s can't get HD2500.Same goes for the desktops. i3s will HAVE to get HD4000 if Intel cares about that market at all.[/citation]

They will have HD 4000, but their HD 4000 will be weaker. I'm trying to explain this to you people, but you just aren't getting it. HD 3000 on an i7 is faster than it is on an i5 which is way faster than it is on an i3. The same works for HD 2000, 2500, and 4000. Do I need to pull up examples of it to prove it to you? I will if I have to.

The HD 4000 on the i7s has a higher clock frequency than it does on the lower end CPUs. That is how Intel differentiates the graphics performance of their CPUs so they can say that higher end processors are better in every way. The HD xxxx model number only defines the IGP's hardware, it does not define the performance. The HD 4000 on each model will have the same exact hardware, but differing clock frequencies. Intel does NOT want to beat AMD in IGP performance, Intel just wants their IGPs to be adequate for most usage. They have Nvidia for situations where their IGPs are not adequate.

 

[ojas]

[citation][nom]blazorthon[/nom]They will have HD 4000, but their HD 4000 will be weaker. I'm trying to explain this to you people, but you just aren't getting it. HD 3000 on an i7 is faster than it is on an i5 which is way faster than it is on an i3. The same works for HD 2000, 2500, and 4000. Do I need to pull up examples of it to prove it to you? I will if I have to.The HD 4000 on the i7s has a higher clock frequency than it does on the lower end CPUs. That is how Intel differentiates the graphics performance of their CPUs so they can say that higher end processors are better in every way. The HD xxxx model number only defines the IGP's hardware, it does not define the performance. The HD 4000 on each model will have the same exact hardware, but differing clock frequencies. Intel does NOT want to beat AMD in IGP performance, Intel just wants their IGPs to be adequate for most usage. They have Nvidia for situations where their IGPs are not adequate.[/citation]
We know that, we're saying that Intel may not be able to use that differing clock frequencies strategy (or binning). Intel have a lot to gain by equaling IGP performance at a given price point. Though that's probably Haswell's job.

 

[blazorthon]

We know that, we're saying that Intel may not be able to use that differing clock frequencies strategy (or binning). Intel have a lot to gain by equaling IGP performance at a given price point. Though that's probably Haswell's job.

Intel will use differing clock frequencies to differentiate HD 4000. They already do it on the processors with HD 4000 that are out right now. Intel does not have anything to gain by matching IGP performance at all. If they want to match performance, then they use discrete cards. It's a much better solution for Intel because if they want to win in graphics, but an APU is released after there that wins, then all they have to do to win again is use a better graphics card. If Intel used more fully integrated solutions, then they would be stuck needing to release a whole new chip to win again or deal with losing. It has other advantages as well. Intel makes their IGPs to be good enough and if they want more, they use a discrete card. AMD doesn't have thus option as well as Intel does because AMD needs the advantage that integrating the different chips offers (better power efficiency and performance scaling, among others) to compete with Intel right now.

Intel is beefing up their IGPs to stay with the times just like they have been doing for years. They will probably always just be good enough IGPs, not top of the line. Even when Haswell comes out, AMD will already have Trinity's successor either out or almost out and then take back the crown, that is if Haswell even takes it from Trinity. The granularity and ease of advancement that discrete graphics offers to Intel far exceeds the advantages of more highly integrated solutions.

 

[blazorthon]

Well, that's one example. The difference between the i3s and i5s is generally greater than the difference between the i5s and i7s.

 

[blazorthon]

your welcome ! I thought this was all common knowledge ?

That's what I thought, but these guys weren't listening when I told them.

 

[cleeve]

[citation][nom]deanjo[/nom]Ummm, no it can't. Handbrake is 100% cpu.[/citation]

Um, yupp it will be. Accelerated version in development.

 

[SuperVeloce]

[citation][nom]Avro Arrow[/nom]And just how the hell is anyone supposed to use 16GB of RAM with a 32 bit build? LMAO!!!!I seriously doubt that. No video card that is less powerful than a desktop ATi Radeon HD 4870 is powerful enough to use more than 512MB which is why most of the nVdia GTX 2xx series only had 768MB on board. I remember reading an article where the author only managed to get the HD 4870 to use more than 512MB in a very specific instance of Crysis.[/citation]

lol, try anno 1404, 500+ is no exception, 750mb easily on larger maps. medium resolutions, high settings, 4x anisotropic. With a dx11 gpu at HD4870 performance levels you could play with 16x anisotropic and maybe full-hd, easypeasy 1gb...

[citation][nom]palladin9479[/nom]Not AMD / Intel's fault really, there is nothing they can do in hardware to remedy the issue.The problem is how Windows does scheduling, no thread is given 100% of a CPU's time. Threads are routinely interrupted so that the OS can query the CPU and possibly task switch another thread onto it. When Windows goes to resume execution on the thread it doesn't always put it back where it got it, where it puts it is based on the CPU utilization of each core. If the CPU is powering down / down clocking under-used cores then Windows will see them as under-used cores and try to assign that thread to them. That cause's them to spike back up. Clocking up / down is determined by a CPU utilization threshold vs a time period. Llano's were factory set to only clock up if there was > 60% utilization for 500ms, and only clock down if there is <40% utilization for more then 2000ms. TB only activates if two or more cores are at their lower P states, I haven't had time to figure out exactly where the threshold was on Llano. I do know that if three cores are running at their P0 state then TB will never activate.Realizing the logic behind boosting you can see why we rarely see it activate on anything in Windows, even if it's a single threaded application. Windows is constantly moving it around which prevents the CPU from ever properly down clocking / boosting. Forcing processor affinity is a way to prevent windows from moving it around and thus allows for the conditions where TB can activate. I've done it multiple times on my A8-3550MX.[/citation]

nah, I tried that many times, even if I manually set affinity of every process to only one core (where possible) and got anywhere near 100% utilization, TB just did not kick in. It seems to me that TB is just random and at moderate loads only

 

[scubaaaDan]

What tool/software was used to capture and graph the Power Consumption charts?
I would like to do something similar for a couple of processors I have access to.

 

[palladin9479]

nah, I tried that many times, even if I manually set affinity of every process to only one core (where possible) and got anywhere near 100% utilization, TB just did not kick in. It seems to me that TB is just random and at moderate loads only

It depends on how heavily your CPU is loaded and what your running. I've succeeded in getting TB @3.0Ghz on a 3550MX to kick in every time on heavy single threaded loads. If your on a laptop you need to check the advanced power settings on Windows 7. If you have it on "High Performance" then you need to go in and set the Minimum CPU to 5% and ensure the maximum CPU is at 100%, this will prevent Windows from trying to maintain high clock speeds constantly on all cores.

What I really want is if someone with access to a Trinity can confirm if K10Stat works or not. Even though it's name is "K10" the program actually works with any CPU that can have it's clocks set in software. All it does is set the MSR's inside the CPU, which is all that "power saving" features do in the OS. Also if your running multiple "power saving / management" applications, they all like to dictate to the CPU what to set it's clocks to and if multiple of them are running then they'll conflict with each other. The worst of which is the Windows 7 power service, something I haven't figured out how to completely shutdown yet.

 

[SuperVeloce]

nah, I tried that many times, even if I manually set affinity of every process to only one core (where possible) and got anywhere near 100% utilization, TB just did not kick in. It seems to me that TB is just random and at moderate loads only

It depends on how heavily your CPU is loaded and what your running. I've succeeded in getting TB @3.0Ghz on a 3550MX to kick in every time on heavy single threaded loads. If your on a laptop you need to check the advanced power settings on Windows 7. If you have it on "High Performance" then you need to go in and set the Minimum CPU to 5% and ensure the maximum CPU is at 100%, this will prevent Windows from trying to maintain high clock speeds constantly on all cores.

What I really want is if someone with access to a Trinity can confirm if K10Stat works or not. Even though it's name is "K10" the program actually works with any CPU that can have it's clocks set in software. All it does is set the MSR's inside the CPU, which is all that "power saving" features do in the OS. Also if your running multiple "power saving / management" applications, they all like to dictate to the CPU what to set it's clocks to and if multiple of them are running then they'll conflict with each other. The worst of which is the Windows 7 power service, something I haven't figured out how to completely shutdown yet.

That's exactly how I set it up in power settings. As I said, I also tried manually set affinity to processes. K10 will almost certainly not work. Liano was derivate of K10, Trinity is bulldozer

 

[nforce4max]

You guys should have benched a a8 35x0 MX with ddr3 1600 just for comparisons. As for the ram used in the tested machines is much worse than mine. I got a a8 3530mx with a 8gb ddr3 1600 kit running stock at 9.9.9.21. Handles skyrim pretty decently.

 

[palladin9479]

That's exactly how I set it up in power settings. As I said, I also tried manually set affinity to processes. K10 will almost certainly not work. Liano was derivate of K10, Trinity is bulldozer

Now I know something is not right. You didn't actually read my posts.

K10 will almost certainly not work. Liano was derivate of K10, Trinity is bulldozer

I said

Even though it's name is "K10" the program actually works with any CPU that can have it's clocks set in software. All it does is set the MSR's inside the CPU, which is all that "power saving" features do in the OS.

K10 stat can set the power state of any CPU using "Cool and Quiet".

What K10 does is set the CPU's fid / did / voltage registers in software. Any CPU that supports dynamic clocking, akaTurboBoost, TurboCore, ect.. supports having it's registers set in software. Sometimes those registers are locked or otherwise write-protected, in which case on the manufacturers driver has the key to unlock them. Sometimes those registers are not write-protected and any software with ring-0 (kernel) access can change them.

 

[palladin9479]

You guys should have benched a a8 35x0 MX with ddr3 1600 just for comparisons. As for the ram used in the tested machines is much worse than mine. I got a a8 3530mx with a 8gb ddr3 1600 kit running stock at 9.9.9.21. Handles skyrim pretty decently.

What memory did you put in it? I'm currently in the process of looking for a 2x4GB kit of DDR3-1600 for my DV6. The crap HP sends with them has disappointed me pretty badly.

That and I'm getting really angry at how AMD is refusing to support any setup with switchable graphics. My board is the newer DV6-6C00 model and HP's removed the fixed GPU setting and made it entirely dynamic. Of course HP sucks at making drivers so it doesn't work half the time and its constantly trying to run things on the 6620G that I want run on the 7690M. Hybrid graphics works pretty well, so DX11 stuff is fine, but DX9 / OpenGL is totally messed up right now.

 

[Guest]

What's the quality in Quick Sync vs VCE? On the "expert boards" (doom9, et al) Quick Sync is considered to be quick and crappy and you can hardly tune it, in spite all the "review sites" saying how great it is. So what's the quality of Quick Sync vs VCE vs CUDA vs x86 ? Can VCE be tunned better than QS?

 

[leaderWON]

[citation][nom]Burnit[/nom]What's the quality in Quick Sync vs VCE? On the "expert boards" (doom9, et al) Quick Sync is considered to be quick and crappy and you can hardly tune it, in spite all the "review sites" saying how great it is. So what's the quality of Quick Sync vs VCE vs CUDA vs x86 ? Can VCE be tunned better than QS?[/citation]

so far the existence of quick sync has amounted to no more than a marketing bullet point.