AMD CPU speculation... and expert conjecture

[szatkus]

I think that could help a lot with feeding GCN. My blind shot: +30%?

I doubt it for two reasons. First off, traversal times. L3 is already very inefficient, in terms of space versus performance, and L4 will be even more so. Secondly, it simply isn't big enough to really benefit a GPU. It will help, don't get me wrong, but only about 5-10% or so. And if it takes a significant portion of the die, one needs to wonder if the space could be better utilized by something else (more GPU cores, more CPU cores, specialized logic processors, etc). I do expect the APU as a whole to be ~20-25% faster on average though, simply due to generational gains.

Right, but both L3 and stacked on-die DRAM are more efficent than RAM (it needs a few hundred cycles to get data). Intel used just 40 MB of L4 in Iris. If AMD used the same capacity it could be smaller than L3 (40 MB / 4 layers = 10 MB and DRAM is more dense than SRAM) in FX. Anyway I hope that they know what they're doing.

 

[gamerk316]

I think that could help a lot with feeding GCN. My blind shot: +30%?

I doubt it for two reasons. First off, traversal times. L3 is already very inefficient, in terms of space versus performance, and L4 will be even more so. Secondly, it simply isn't big enough to really benefit a GPU. It will help, don't get me wrong, but only about 5-10% or so. And if it takes a significant portion of the die, one needs to wonder if the space could be better utilized by something else (more GPU cores, more CPU cores, specialized logic processors, etc). I do expect the APU as a whole to be ~20-25% faster on average though, simply due to generational gains.

Right, but both L3 and stacked on-die DRAM are more efficent than RAM (it needs a few hundred cycles to get data). Intel used just 40 MB of L4 in Iris. If AMD used the same capacity it could be smaller than L3 (40 MB / 4 layers = 10 MB and DRAM is more dense than SRAM) in FX. Anyway I hope that they know what they're doing.

At least back in the Pentium 4/DDR days, it took about 80-100 cycles to hit main memory, assuming no pagefaults (which take MILLIONS of cycles to resolve, thanks to the HDD latency coming into play). Can't imagine it's much worst then that 10 years later. Point being, L3 already takes about 30-40 cycles to access, so L4 really isn't THAT substantially faster. I can't help but think you'd be better off using the die space for more GPU cores then for 128MB of stacked DRAM.

 

[szatkus]

I think that could help a lot with feeding GCN. My blind shot: +30%?

I doubt it for two reasons. First off, traversal times. L3 is already very inefficient, in terms of space versus performance, and L4 will be even more so. Secondly, it simply isn't big enough to really benefit a GPU. It will help, don't get me wrong, but only about 5-10% or so. And if it takes a significant portion of the die, one needs to wonder if the space could be better utilized by something else (more GPU cores, more CPU cores, specialized logic processors, etc). I do expect the APU as a whole to be ~20-25% faster on average though, simply due to generational gains.

Right, but both L3 and stacked on-die DRAM are more efficent than RAM (it needs a few hundred cycles to get data). Intel used just 40 MB of L4 in Iris. If AMD used the same capacity it could be smaller than L3 (40 MB / 4 layers = 10 MB and DRAM is more dense than SRAM) in FX. Anyway I hope that they know what they're doing.

At least back in the Pentium 4/DDR days, it took about 80-100 cycles to hit main memory, assuming no pagefaults (which take MILLIONS of cycles to resolve, thanks to the HDD latency coming into play). Can't imagine it's much worst then that 10 years later. Point being, L3 already takes about 30-40 cycles to access, so L4 really isn't THAT substantially faster. I can't help but think you'd be better off using the die space for more GPU cores then for 128MB of stacked DRAM.

GCN in Kaveri was already bottlenecked by brandwidth (384 version is only slightly worse than 512 even in GPU bottlenecked tests), so I don't think that moar cores would help having only DDR3.

 

[gamerk316]

I think that could help a lot with feeding GCN. My blind shot: +30%?

I doubt it for two reasons. First off, traversal times. L3 is already very inefficient, in terms of space versus performance, and L4 will be even more so. Secondly, it simply isn't big enough to really benefit a GPU. It will help, don't get me wrong, but only about 5-10% or so. And if it takes a significant portion of the die, one needs to wonder if the space could be better utilized by something else (more GPU cores, more CPU cores, specialized logic processors, etc). I do expect the APU as a whole to be ~20-25% faster on average though, simply due to generational gains.

Right, but both L3 and stacked on-die DRAM are more efficent than RAM (it needs a few hundred cycles to get data). Intel used just 40 MB of L4 in Iris. If AMD used the same capacity it could be smaller than L3 (40 MB / 4 layers = 10 MB and DRAM is more dense than SRAM) in FX. Anyway I hope that they know what they're doing.

At least back in the Pentium 4/DDR days, it took about 80-100 cycles to hit main memory, assuming no pagefaults (which take MILLIONS of cycles to resolve, thanks to the HDD latency coming into play). Can't imagine it's much worst then that 10 years later. Point being, L3 already takes about 30-40 cycles to access, so L4 really isn't THAT substantially faster. I can't help but think you'd be better off using the die space for more GPU cores then for 128MB of stacked DRAM.

GCN in Kaveri was already bottlenecked by brandwidth (384 version is only slightly worse than 512 even in GPU bottlenecked tests), so I don't think that moar cores would help having only DDR3.

Again, I didn't see many cases where Bandwidth was killing Kaveri. We saw diminishing returns past DDR3-2133, which indicated bandwidth was less of an issue past that point.

 

[szatkus]

<cut>

GCN in Kaveri was already bottlenecked by brandwidth (384 version is only slightly worse than 512 even in GPU bottlenecked tests), so I don't think that moar cores would help having only DDR3.

Again, I didn't see many cases where Bandwidth was killing Kaveri. We saw diminishing returns past DDR3-2133, which indicated bandwidth was less of an issue past that point.

Yeah, I know, it's strange because between 1800 and 2133 difference was quite substantial.
Maybe because "The A10-7850K natively supports up to DDR3-2133"? Maybe just memory controller is too weak?

 

[szatkus]

Ok, I found something to support my theory:

http://www.corsair.com/en/blog/2014/february/understanding-kaveri

No or very little difference between 2133 and 2400, but overclocking BClk (and because of that, also the memory controller) actually improved performance.

 

[gamerk316]

True, but you improve the CPU/GPU too, so it's hard to say what factored into the performance gains. I do agree the Kaveri MC seems like a weak link though.

 

[gamerk316]

Just going to plop this here:

http://www.extremetech.com/computing/185797-forget-moores-law-hot-and-slow-dram-is-a-major-roadblock-to-exascale-and-beyond

 

[jdwii]

The 20% bottleneck should only occur only heavier workflows where the frontend is essentially starving the backends.
Also windows have provided a "hotfix" for the old problem, and should now threat is as a quad core.

When i test my 8 core FX and lock a 4 threaded game to just use cores 1,2,3,4 it lowers performance around 12% under GTA4-Watchdogs really dies and farcry 3. So again 20% in heavy CPU tasks actually under fritzchess its 30%

 

[jdwii]

4 core FX are turds....I couldn't believe just how bad the Fx 4300 is/was....it literally needs to run like 4.5-5.0ghz to be even remotely acceptable and get past the bad slowness/lag. even then its still pretty slow....It runs into some hell with Battlefield 4 haha

Well come on if a game needs to use 4 of amd cores to be performing great (since they have at least a 30% performance per core pentaitly) That means a 4 core fx would only be able to handle 2 cores well and then the other 2 cores lightly which would perform better on a 8 core fx which could use all 4 cores well

Edit, more about my setup i left C6 enabled on my board and it lowered my power consumption on idle to 81 watts vs 106 watts. With windows set to balanced and Nvidia set to adaptive clock speed over max speed all the time i get 140-150 is normal use on chrome and vlc player but it does peak at 170 watts when loading newer tabs now i'm going to enable cool and quiet and see.

Juan is on to one thing high end PC's today use too much power doing normal tasks such as web browsing there is no reason logically for a PC to use more than 20-40 watts when on youtube or chrome-VLC.
Disable cache cores and lower clock speed to improve power consumption and then enable it doing heavy tasks.

 

[jdwii]

Why is it that these sites have to start with a lie?????

http://www.forbes.com/sites/greatspeculations/2014/07/07/amds-expanding-apu-portfolio-to-accelerate-demand-for-its-processors/

Then we have this site
http://www.ign.com/articles/2014/07/14/pc-dominates-gaming-hardware-sales

That can prove they are wrong.

 

[Cazalan]

Why is it that these sites have to start with a lie?????

http://www.forbes.com/sites/greatspeculations/2014/07/07/amds-expanding-apu-portfolio-to-accelerate-demand-for-its-processors/

Then we have this site
http://www.ign.com/articles/2014/07/14/pc-dominates-gaming-hardware-sales

That can prove they are wrong.

What was the lie? The 7800 did just launch.
The PC gaming segment can still grow in an overall declining (growth) PC market. There are still 280 million or so PCs shipped a year. If you averaged 1k for a gaming PC that would only be 20 million of those 280 million.

According to Intel the XP support ending by Microsoft helped slow the decline by about half what they were expecting. We'll find out Thursday (AMD Earnings call) if any of that increase made it to AMD as well.

 

[jdwii]

Why is it that these sites have to start with a lie?????

http://www.forbes.com/sites/greatspeculations/2014/07/07/amds-expanding-apu-portfolio-to-accelerate-demand-for-its-processors/

Then we have this site
http://www.ign.com/articles/2014/07/14/pc-dominates-gaming-hardware-sales

That can prove they are wrong.

What was the lie? The 7800 did just launch.
The PC gaming segment can still grow in an overall declining (growth) PC market. There are still 280 million or so PCs shipped a year. If you averaged 1k for a gaming PC that would only be 20 million of those 280 million.

According to Intel the XP support ending by Microsoft helped slow the decline by about half what they were expecting. We'll find out Thursday (AMD Earnings call) if any of that increase made it to AMD as well.

Its nothing less then a fact that hardware is going up not down from last year as even newegg states anything else is trying to distort from the truth

 

[de5_Roy]

AMD Mantle: A Graphics API Tested In Depth
http://www.tomshardware.com/reviews/amd-mantle-performance-benchmark,3860.html

Juan is on to one thing high end PC's today use too much power doing normal tasks such as web browsing there is no reason logically for a PC to use more than 20-40 watts when on youtube or chrome-VLC.
Disable cache cores and lower clock speed to improve power consumption and then enable it doing heavy tasks.

check your power profiles and bios to power saving options. if windows' power profile is set to high performance or even balanced, it may cause the pc to use more power (e.g. optimizes for max. performance). you can fine tune which parts (as much as windows allows, which is sadly not much) go for performance and which to save power (e.g. wifi or usb ports). some of the bios power saving settings set to "auto" may be disabled to ensure compatibility. you might try setting them to "enabled" as long as your pc operates reliably. even in individual applications such as media players, you can ensure they use less recource e.g. doesn't wake up all of the cpu and uses the gpu for video playback etc. you can go even further with undervolting, underclocking, e.g. save a seperate bios profile for low power use. it depends on how much you fine-tune. it's great that you have a kill a watt meter - you can directly monitor which settings save power or use more power.

edit: the "pc sales rebound" seems temporary to me. after the xp upgrading period is over, we might be seeing same old decline rates. ms will have much harder time getting people to upgrade from windows 7 (to.. what, really? ) gaming and hedt are different segments from overall pc market.

 

[colinp]

Newegg doesn't even represent the whole of the US market, never mind the rest of the world

 

[Embra]

AMD Mantle: A Graphics API Tested In Depth
http://www.tomshardware.com/reviews/amd-mantle-performance-benchmark,3860.html

I wish they added a i5-4670k or 4690k. Interesting read.

 

[jdwii]

AMD Mantle: A Graphics API Tested In Depth
http://www.tomshardware.com/reviews/amd-mantle-performance-benchmark,3860.html

Juan is on to one thing high end PC's today use too much power doing normal tasks such as web browsing there is no reason logically for a PC to use more than 20-40 watts when on youtube or chrome-VLC.
Disable cache cores and lower clock speed to improve power consumption and then enable it doing heavy tasks.

check your power profiles and bios to power saving options. if windows' power profile is set to high performance or even balanced, it may cause the pc to use more power (e.g. optimizes for max. performance). you can fine tune which parts (as much as windows allows, which is sadly not much) go for performance and which to save power (e.g. wifi or usb ports). some of the bios power saving settings set to "auto" may be disabled to ensure compatibility. you might try setting them to "enabled" as long as your pc operates reliably. even in individual applications such as media players, you can ensure they use less recource e.g. doesn't wake up all of the cpu and uses the gpu for video playback etc. you can go even further with undervolting, underclocking, e.g. save a seperate bios profile for low power use. it depends on how much you fine-tune. it's great that you have a kill a watt meter - you can directly monitor which settings save power or use more power.

edit: the "pc sales rebound" seems temporary to me. after the xp upgrading period is over, we might be seeing same old decline rates. ms will have much harder time getting people to upgrade from windows 7 (to.. what, really? ) gaming and hedt are different segments from overall pc market.

I have everything set right and i'm sorry typing to you right now i'm using 95 watts with cool and quiet on that means what juan has to say is some what right you know i'm no fan to him but that doesn't make some of his claims not correct

 

[jdwii]

Newegg doesn't even represent the whole of the US market, never mind the rest of the world

Yes it does :lol:

No but generally amazon cost the same

 

[gamerk316]

The 20% bottleneck should only occur only heavier workflows where the frontend is essentially starving the backends.
Also windows have provided a "hotfix" for the old problem, and should now threat is as a quad core.

When i test my 8 core FX and lock a 4 threaded game to just use cores 1,2,3,4 it lowers performance around 12% under GTA4-Watchdogs really dies and farcry 3. So again 20% in heavy CPU tasks actually under fritzchess its 30%

Makes sense, because you're forcing the 20% performance penalty. To avoid using the second core of a BD module, you have to use every other core (eg, 0,2,4,6 OR 1,3,5,7). I'd be interested to see how that affects your results.

 

[gamerk316]

AMD Mantle: A Graphics API Tested In Depth
http://www.tomshardware.com/reviews/amd-mantle-performance-benchmark,3860.html

I wish they added a i5-4670k or 4690k. Interesting read.

Kinda ironic that BF4 officially sucks with Mantle, given how hard AMD pushed it.

 

[jdwii]

The 20% bottleneck should only occur only heavier workflows where the frontend is essentially starving the backends.
Also windows have provided a "hotfix" for the old problem, and should now threat is as a quad core.

When i test my 8 core FX and lock a 4 threaded game to just use cores 1,2,3,4 it lowers performance around 12% under GTA4-Watchdogs really dies and farcry 3. So again 20% in heavy CPU tasks actually under fritzchess its 30%

Makes sense, because you're forcing the 20% performance penalty. To avoid using the second core of a BD module, you have to use every other core (eg, 0,2,4,6 OR 1,3,5,7). I'd be interested to see how that affects your results.

actually 0-2-4-6-8 comes with the same benchmarks as 1-3-5-7 as it should

 

[gamerk316]

Right, but I'd assume they both would be faster then 0-1-2-3, unless Turbo is still kicking in.

 

[Rum]

AMD Mantle: A Graphics API Tested In Depth
http://www.tomshardware.com/reviews/amd-mantle-performance-benchmark,3860.html

I wish they added a i5-4670k or 4690k. Interesting read.

Kinda ironic that BF4 officially sucks with Mantle, given how hard AMD pushed it.

Considering BF4 sucked to begin with I am not surprised, they tried hard to polish that turd as much as possible in the end it was still a turd!

 

[gamerk316]

AMD Mantle: A Graphics API Tested In Depth
http://www.tomshardware.com/reviews/amd-mantle-performance-benchmark,3860.html

I wish they added a i5-4670k or 4690k. Interesting read.

Kinda ironic that BF4 officially sucks with Mantle, given how hard AMD pushed it.

Considering BF4 sucked to begin with I am not surprised, they tried hard to polish that turd as much as possible in the end it was still a turd!

Which is irrelevant when discussion DX vs Mantle performance metrics.

 

[Cazalan]

Its nothing less then a fact that hardware is going up not down from last year as even newegg states anything else is trying to distort from the truth

Not sure where you're getting the numbers from but even Intel is saying the unit numbers are down from last year, just not down as low as anticipated. Intel saw a sizable increase in Server revenue, more than in PC client.

AMD was expecting a 7-10% decline in 2014 in PC client.