AMD CPU speculation... and expert conjecture

[-Fran-]

Well, a purchase decision is (or should, haha) be based on what your needs are.

In particular, if there are no games or HSA enabled software in your life, you might be better off with Intel, since they have the better CPU (overall). That's how I see it currently.

And the other not less important factor, now that it was included in the conversation, is what OEMs put as goodies to platforms. In Desktop, although AM3+ and FM2 are (in my opinion) better armed, OEMs tend to leave AMD with the design leftovers. I guess that's tangent to what AMD and Intel can do in terms of negotiation, but Intel does get better builds in mobile from what I've seen.

Cheers!

EDIT: Rephrased and idea.

 

[Ags1]

Spoiler

Nope, If I want play games at 1080p at medium settings testing CPUs at low settings and 720p is misleading. Moreover, as commented before, the tendency in games will be towards offloading the CPU generating GPU-bounded games.

Now you talk two different things. You can either:

A: Bench the CPU at the settings where the difference in CPU is the greatest

OR

B: Bench the CPU at settings where everyone actually plays at.

The problem with the second, of course, is any differences in CPU power is suppressed due to the GPU bottleneck in games at high settings. Running at lower settings shows the difference in power between two CPUs when said bottleneck is removed. Frankly, telling me CPU A and CPU B perform the same in game X with GPU Y is meaningless; what if I have a faster GPU, do the two CPU's still perform the same? Woops, don't have a clue. By testing at minimal settings though, you see which one is truly more powerful, even if the FPS numbers are meaningless.

Frankly, both ways should be done; min settings to show the order of CPU's in terms of power on said game, and at max settings to determine typical FPS with a given hardware setup.

Double review is a good idea, and some sites do that, finding that A10-6880k is ~40% behind i5-2500k at low settings and 720p/1024p, but only ~8% behind at high settings and 1080p. All this using a 7970 GHz Ed with Windows 7 and 'older' games. The differences are smaller with modern games under W8.1 or SteamOS.

Which is a somewhat typical result. GPU bottleneck at high settings hiding the CPU difference, which is at least 40% for that given game (possibly more, if you could get settings lower via GPU driver tweaks). Run at low settings where CPU power is FAR more important, then the differences in power (and power reserve) becomes clear.

Which infers that unless AMD increases performance by 40%, Kaveri won't be catching the 2500k, at least in that particular game.

So understand what review sites are benching: They typically bench how a given processor(s) perform at given settings in a given benchmark. They do NOT directly compare the relative processing power of said processors when doing these benchmarks. This leads to false impressions, like a 6880k being only 8% slower then a 2500k.

First, I wrote "games", plural. The above percentages are averages.

Second, this comparison was made under Windows 7 (which is optimized for Intel). AMD APUs are 5-15% faster under W8.1 or SteamOS that under W7, thanks to a new scheduler than understand the CMT architecture of AMD modules, closing the artificial gap with Intel chips.

Third, the comparison used 1600MHz RAM for both, when Richland stock memory speed is 2133MHz. Add another 5% by using stock memory timings for Richland.

Therefore, AMD increasing hardware performance by ~ 20% is enough to caught the i5 in so-called CPU benchmarks. The rest 10-20% comes from using stock memory and an operative system not optimized for Intel .

Fourth, the result of the 6880k being only 8% slower than a 2500k at 1080p at high settings is real and measured. If you play those games at that resolution and settings, that is the performance than you obtain. It is not some imagined score.

Fifth, as explained again and again and again and again in this thread. Future games (i.e. games developed for PS4/Xbox1) will be GPU-bound, and games ported to PC, using MANTLE, will be much more GPU-bound. As shown during APU13, a CPU as that in Kaveri will be able to feed the fastest graphics cards using MANTLE.

Resume:
=======

Old games: Kaveri =< i5-2500k

Modern games: Kaveri >= i5-2500k

MANTLE games: Kaveri ~ i7-4770k

Bonus:
=====

Interview with EIDOS. THIEF will come with improved multi-core support, MANTLE edition for "high performance graphics", and PhysX support. Enjoy!

http://www.dsogaming.com/interviews/eidos-montreal-talks-thief-tech-tessellation-fov-slider-multi-core-cpus-graphical-features/

The orange AMD bar chart showing APU improvements is beyond absurd - the bars represent 100, 106 and 109% but they choose to show the 100 bar half the length of 106, which implies AMD cut off the first 94% of the bars. A correct graph would show three bars all approximately the same length.

 

[juanrga]

Spoiler

Nope, If I want play games at 1080p at medium settings testing CPUs at low settings and 720p is misleading. Moreover, as commented before, the tendency in games will be towards offloading the CPU generating GPU-bounded games.

Now you talk two different things. You can either:

A: Bench the CPU at the settings where the difference in CPU is the greatest

OR

B: Bench the CPU at settings where everyone actually plays at.

The problem with the second, of course, is any differences in CPU power is suppressed due to the GPU bottleneck in games at high settings. Running at lower settings shows the difference in power between two CPUs when said bottleneck is removed. Frankly, telling me CPU A and CPU B perform the same in game X with GPU Y is meaningless; what if I have a faster GPU, do the two CPU's still perform the same? Woops, don't have a clue. By testing at minimal settings though, you see which one is truly more powerful, even if the FPS numbers are meaningless.

Frankly, both ways should be done; min settings to show the order of CPU's in terms of power on said game, and at max settings to determine typical FPS with a given hardware setup.

Double review is a good idea, and some sites do that, finding that A10-6880k is ~40% behind i5-2500k at low settings and 720p/1024p, but only ~8% behind at high settings and 1080p. All this using a 7970 GHz Ed with Windows 7 and 'older' games. The differences are smaller with modern games under W8.1 or SteamOS.

Which is a somewhat typical result. GPU bottleneck at high settings hiding the CPU difference, which is at least 40% for that given game (possibly more, if you could get settings lower via GPU driver tweaks). Run at low settings where CPU power is FAR more important, then the differences in power (and power reserve) becomes clear.

Which infers that unless AMD increases performance by 40%, Kaveri won't be catching the 2500k, at least in that particular game.

So understand what review sites are benching: They typically bench how a given processor(s) perform at given settings in a given benchmark. They do NOT directly compare the relative processing power of said processors when doing these benchmarks. This leads to false impressions, like a 6880k being only 8% slower then a 2500k.

First, I wrote "games", plural. The above percentages are averages.

Second, this comparison was made under Windows 7 (which is optimized for Intel). AMD APUs are 5-15% faster under W8.1 or SteamOS that under W7, thanks to a new scheduler than understand the CMT architecture of AMD modules, closing the artificial gap with Intel chips.

Third, the comparison used 1600MHz RAM for both, when Richland stock memory speed is 2133MHz. Add another 5% by using stock memory timings for Richland.

Therefore, AMD increasing hardware performance by ~ 20% is enough to caught the i5 in so-called CPU benchmarks. The rest 10-20% comes from using stock memory and an operative system not optimized for Intel .

Fourth, the result of the 6880k being only 8% slower than a 2500k at 1080p at high settings is real and measured. If you play those games at that resolution and settings, that is the performance than you obtain. It is not some imagined score.

Fifth, as explained again and again and again and again in this thread. Future games (i.e. games developed for PS4/Xbox1) will be GPU-bound, and games ported to PC, using MANTLE, will be much more GPU-bound. As shown during APU13, a CPU as that in Kaveri will be able to feed the fastest graphics cards using MANTLE.

Resume:
=======

Old games: Kaveri =< i5-2500k

Modern games: Kaveri >= i5-2500k

MANTLE games: Kaveri ~ i7-4770k

Bonus:
=====

Interview with EIDOS. THIEF will come with improved multi-core support, MANTLE edition for "high performance graphics", and PhysX support. Enjoy!

http://www.dsogaming.com/interviews/eidos-montreal-talks-thief-tech-tessellation-fov-slider-multi-core-cpus-graphical-features/

The orange AMD bar chart showing APU improvements is beyond absurd - the bars represent 100, 106 and 109% but they choose to show the 100 bar half the length of 106, which implies AMD cut off the first 94% of the bars. A correct graph would show three bars all approximately the same length.

Putting the baseline above zero is a typical marketing trick, used to visually exaggerate differences, but I took the numerical values, aka the percentages, not how the bars look.

 

[juanrga]

Discussion about the Cinebench benchmark of Kaveri continues

http://wccftech.com/amd-kaveri-a107850k-benchmark-surfaced/

Is it legit or fake?

 

[palladin9479]

From what I can see Kaveri is not destined to be a monster of a CPU ... rather it is targeted at a particular "niche" in the market AMD feel will make them money, appeal to a particular range of devices.

It seems a really nice notebook processor, HP is selling a bunch of cheap ones in the Pavilion 15 and Envy 15 lines, even have some cheap chromebooks. It's the desktop version I'm suspicious of due to the lower clocks eating into it's performance enhancements. It'll be a solid upgrade in all places that currently appeal to APU's, namely HTPC and SFF / Mini-ITX box's. Places where power and space limit the use of dGPU's and favor quiet solutions.

 

[FALC0N]

That cinebench number is very realistic. It represents a 10% to 20% IPC increase and we already know bulldozer and piledriver could hit 4.9 ghz. I don't know if its real, but its certainly realistic.

 

[jdwii]

The score seems like it could be possible until you realize that Cincebench focuses more on the FPU which Amd did not really improve much under steamroller which means i doubt these are real scores. Another reason why i doubt this is the next gen APU is being built on 28nm bulk and is already having clock speed target issues. Meaning if this was real don't expect the majority to clock to that limit even richland has problems doing that with stability.

 

[FALC0N]

Jdwii, I find that to be an extremely weak and unconvincing argument.

For starters, your overstating the impact of the FPU on the test. Not to mention the FPU is fed in part by the decoders, which have doubled in steamroller. That point was just not well thought out.

Your clock speed argument is non-relevant. I'm quite certain they cherry picked a part to get to 4.9 ghz. Overclockers are not the target market and growing pains are not unexpected. If it can't do 5 ghz at launch, it will in the near future.

 

[jdwii]

^ we shall see, i still doubt the scores myself but then again i already stated why.

 

[FALC0N]

I would agree if it was a the score at stock. But at 4.9 ghz? That only a modest improvement over piledriver at the same clock.

 

[Cazalan]

I would agree if it was a the score at stock. But at 4.9 ghz? That only a modest improvement over piledriver at the same clock.

Richland at 5Ghz gets 395. That would be quite a big jump to 547.

http://hwbot.org/submission/2435718_xcero_cinebench_r15_a10_6800k_395_cb

 

[noob2222]

CB 15 increased AMD cpus by quite a bit over 11.5.

Clock for clock its 28% faster than richland. It will be interesting to see how much improvement is ipc and how much is multicore scaling. Its still nowhere near i5 2500k.

if I had to guess, 5-10% is single core.

Tourist also mentioned trinity being 5% faster just with the new motherboard. If cb is affected by the board, that could put the cpu itself close to the 20% mark.

 

[FALC0N]

I would agree if it was a the score at stock. But at 4.9 ghz? That only a modest improvement over piledriver at the same clock.

Richland at 5Ghz gets 395. That would be quite a big jump to 547.

http://hwbot.org/submission/2435718_xcero_cinebench_r15_a10_6800k_395_cb

You and Jdwii might have a point here. I had used this number for my earlier calculation, but I mistakenly thought it was a stock score, which gave a 15-20% IPC increase. Quite reasonable, especially in light of the lower clocks. Recalculating with the correct figures, I get closer to 40% IPC increase. That one is a little shakier in my mind. I wouldn't rule it out, as IPC was a big target of the decoder changes and the clock speeds are lower, but it does raise an eyebrow.

So 40% IPC improvement, minus 10% clock speed would leave the top end Kavari part about 30% faster than Richland. And that might be a best case scenario among benchmarks as well. So it is possible that its still legit, but im more skeptical than before.

 

[de5_Roy]

^^ since amd is limited by process node and foundry, hardware ipc improvements will be most welcome for kaveri. while the cpu clockrates are seemingly low for a 95w 28nm apu, kaveri will also house a 512 shader, 720 mhz gcn igpu, sharing the power budget. that's a lot of transistors. is comparing the R7 igpu in a10 7850k to radeon hd 7750 fair? 7750 had 1.5B transistors, built on tsmc's 28nm bulk process.
along with ipc gain, another helpful part may be more granular, effective turbo controller like the one seen in richland (but very likely (read: hopefully) more tuned and improved). this is for stock settings and laptops though, since that's what almost everyone will use - apart from casual and enthusiast overclockers.
power and temp monitoring logics will also need a lot of fine tuning, same with motherboards and their bioses.

 

[bosaile]

intel haswell-e processors will dominate, 16 threads 8 cores 20mbs l3, yeah okay

 

[szatkus]

R15 was released in september. If Cinebench uses modern instruction sets (like AVX) it's quite possible. New instruction sets are usually highly optimized during first 2-3 generations after introduction.

 

[gamerk316]

Discussion about the Cinebench benchmark of Kaveri continues

http://wccftech.com/amd-kaveri-a107850k-benchmark-surfaced/

Is it legit or fake?

About where I would expect Kaveri to fall, quite honestly. Not saying if its real or not, but the numbers certainly seem to be in the ballpark.

 

[gamerk316]

I would agree if it was a the score at stock. But at 4.9 ghz? That only a modest improvement over piledriver at the same clock.

Richland at 5Ghz gets 395. That would be quite a big jump to 547.

http://hwbot.org/submission/2435718_xcero_cinebench_r15_a10_6800k_395_cb

CB 15 increased AMD cpus by quite a bit over 11.5.

Clock for clock its 28% faster than richland. It will be interesting to see how much improvement is ipc and how much is multicore scaling. Its still nowhere near i5 2500k.

if I had to guess, 5-10% is single core.

Tourist also mentioned trinity being 5% faster just with the new motherboard. If cb is affected by the board, that could put the cpu itself close to the 20% mark.

You hit the main point here, the entire CHIP is ~28% faster then Richland [for this benchmark at least], which means, on a per-core basis, IPC only went up about ~7%.

Likewise, since the benchmark loads all cores, you could take the results as the "best case" improvement of ~30% or so. Looking at the other side of the coin, since most tasks do NOT scale to four cores, the "typical" improvement would be half that, or, *gasp* 15%.

Which means AMD still has the same old problem: Unless programs scale, it gets crushed by Intel.

 

[szatkus]

I would agree if it was a the score at stock. But at 4.9 ghz? That only a modest improvement over piledriver at the same clock.

Richland at 5Ghz gets 395. That would be quite a big jump to 547.

http://hwbot.org/submission/2435718_xcero_cinebench_r15_a10_6800k_395_cb

CB 15 increased AMD cpus by quite a bit over 11.5.

Clock for clock its 28% faster than richland. It will be interesting to see how much improvement is ipc and how much is multicore scaling. Its still nowhere near i5 2500k.

if I had to guess, 5-10% is single core.

Tourist also mentioned trinity being 5% faster just with the new motherboard. If cb is affected by the board, that could put the cpu itself close to the 20% mark.

You hit the main point here, the entire CHIP is ~28% faster then Richland [for this benchmark at least], which means, on a per-core basis, IPC only went up about ~7%.

Look at this: http://browser.primatelabs.com/geekbench3/compare/32475?baseline=223722

Multicore improvement for floating point tests is only 12% better than single core.

 

[gamerk316]

Look at this: http://browser.primatelabs.com/geekbench3/compare/32475...

Multicore improvement for floating point tests is only 12% better than single core. Not ~20%.

I have a nice excel spreadsheet where I calculated everything, but found out that I can't drop in a HTML table...*glares at site admins*

So you'll simply have to make due with the aggregate results instead:

Average %increase in Integer Performance: 16.02874497%
Average %increase in Single Core Integer Performance: 13.05047673%
Average %increase in Multi-Core Integer Performance: 19.0070132%

Average %increase in Floating Point Performance: 17.44868573%
Average %increase in Single Core Floating Point Performance: 15.7498446%
Average %increase in Multi-Core Floating Point Performance: 19.14752686%

Average %increase in Memory Performance: 0.2508381%
Average %increase in Single Core Memory Performance: -0.006994888%
Average %increase in Multi-Core Memory Performance: 0.508671088%

So 16-17% on average. In the individual results though, the results were a bit odd and unexpected. For Integer, ~10-15% was the typical increase in performance, with a few tests doing significantly better and raising the average. For FP, the median occurrence was a LOT higher, typically above 20% improvement (and 4-5 cases of 30% increases to performance), with a few cases dragging down the average. FP clearly increased more across the board then Integer based performance this go around for AMD, based on the aggregate numbers, with a handful of tests dragging down the average. Memory results are obviously just noise.

I'd post the entire thing if I could find a way to tablize it on the forums. *shrug*

 

[gamerk316]

I also note, if the numbers are real (which I stress we can't confirm or deny yet), then my 15% number is holding up.

 

[szatkus]

Look at this: http://browser.primatelabs.com/geekbench3/compare/32475...

Multicore improvement for floating point tests is only 12% better than single core. Not ~20%.

I have a nice excel spreadsheet where I calculated everything, but found out that I can't drop in a HTML table...*glares at site admins*

So you'll simply have to make due with the aggregate results instead:

Average %increase in Integer Performance: 16.02874497%
Average %increase in Single Core Integer Performance: 13.05047673%
Average %increase in Multi-Core Integer Performance: 19.0070132%

Average %increase in Floating Point Performance: 17.44868573%
Average %increase in Single Core Floating Point Performance: 15.7498446%
Average %increase in Multi-Core Floating Point Performance: 19.14752686%

Average %increase in Memory Performance: 0.2508381%
Average %increase in Single Core Memory Performance: -0.006994888%
Average %increase in Multi-Core Memory Performance: 0.508671088%

So 16-17% on average. In the individual results though, the results were a bit odd and unexpected. For Integer, ~10-15% was the typical increase in performance, with a few tests doing significantly better and raising the average. For FP, the median occurrence was a LOT higher, typically above 20% improvement (and 4-5 cases of 30% increases to performance), with a few cases dragging down the average. FP clearly increased more across the board then Integer based performance this go around for AMD, based on the aggregate numbers, with a handful of tests dragging down the average. Memory results are obviously just noise.

I'd post the entire thing if I could find a way to tablize it on the forums. *shrug*

You can use Google Docs and put link here

I deliberately chose CPU with similar memory score to minimize memory brandwidth influence. Obviously many of these benchmarks highly depend on memory so it could be not too accurate.

 

[-Fran-]

That Cinebench screenshot came out of 4Chan, so there you go. It's fake.

From all the places we can trust "leaks", 4Chan is not one of them, haha.

Cheers!

 

[de5_Roy]

That Cinebench screenshot came out of 4Chan, so there you go. It's fake.

From all the places we can trust "leaks", 4Chan is not one of them, haha.

Cheers!

i hear you can mine bitcoin with apple macs
http://www.tomsguide.com/us/dont-believe-bitcoin-mining-apple,news-17987.html
everything on the internet is true, so i'll try it.

 

[-Fran-]

Well, then instead of "leaks", make that "overall all types of information", haha.

Except trolls, of course.

Cheers!