AMD CPU speculation... and expert conjecture

[szatkus]

Oh god some of you...

AVX is useless in Windows. Hardly and programs use it and most target sse2 or lower.

Secondly, AVX doesnt increase instructions per clock. It decreases the amount of instructions to do a certain task. Which is why CISC has big advantages over RISC.

It's like comparing two ways of doing a math problem that is long. One takes a lot of shortcuts and simplifies everything. the other does every menial operation. And then you finish and look at the person who took all the shortcuts and go "wow, youre very fast at doing menial math!"

IE you dont sound like you have a clue what youre talking about.

But if you want to compare CISC like x86 to RISC like ARM, the math example is a good one. CISC is about taking shortcuts to get things done faster and RISC is about doing a lot of simple operations quickly.

Right, I analyzed some time ago my collection of games for Windows. Some of them even used AVX, but only few instructions. Anyway, adopting new instruction takes years (now even more because PC ecosystem is more mature than in 90. or 00.), performance with legacy code is the most important for PC CPU.

 

[juanrga]

Juan did you even watch the AMD Core Innovation Summit? Lisa Su and Jim Keller said that x86 is just as important as ARM or did you chose to tune that part out? Keller also said that the build process will be fast, he and his team are taking things from other builds/ arch that AMD perfected in the past and implement them in the new Arch. Your assumptions that the new x86 Arch won't outperform the K12 are asinine and unfounded.
One last thing I am sure that Keller has plenty of funds to build his new x86 toy, all your nonsense just sounds like sour grapes cuz it's not all about ARM!

I think that I didn't miss anything important from the Core Innovation Conference. You can check in my tweeter account what I said then: "AMD now means AMbiDextrous". I also gave in this thread the ARM+x86 TAM numbers that AMD expect for the next four years and I mentioned more than once that the new x86 core will be a "sister core" to K12.

Once said that, I want emphasize the very important fact that AMD used to denote their flagship product with a "K": K7, K8, K10, K11... whereas the small cores (Bobcat, Jaguar, Puma+) are not. The difference is that now K12 is an ARM core not a x86 core; in fact the x86 core is still unnamed by AMD.

During the talk, Keller also mentioned how the armv8 ISA allows to extract more performance than x86, and said explicitly that the ARM core has a "bigger engine", which I interpreted as wider-issue or similar feature. If you look at other forums, other people is making similar claims:

My gut feeling tells me that the ARM core will be released first, followed by a less performing x86 variant shortly after.

I also gave the next quote from AMD head:

At AMD, we intended to be the leader in ARM CPUs

No similar comment for x86 was done. I bet you to find a similar quote for x86.

Finally, AMD finances are well-known and during the talk Papermaster mentioned something about not giving Keller enough resources.

 

[GOM3RPLY3R]

http://www.bit-tech.net/hardware/2012/11/06/amd-fx-8350-review/6

This just one example... even ancient 2600k at stock blows the dust out of the 8350.

I plan on supporting AMD in the future, right now Piledriver and Bulldozer just does not cut the performance i need for Gaming.

AMD should had dumped Bulldozer/Piledriver just after the FX-8350, but i can understand their position... they have such a limited budget and if you don`t belive it just watch the video where Keller, Rory and Lisa are all together talking about AMD, Keller jokingly asks if he can have more budget and they reply to him, Sorry no.

It`s closing to 3 years after the bulldozer Arch and AMD still pushing with Piledriver... and as far as i am aware Piledriver will extend to 2015, i have no idea what are their plans for new Desktop CPUs but i am starting to doubt we will see something new as "piledrivers" all through 2015.

2016 is the big day, but looking at past CPUs from AMD the new arch will probably hit some where between Q4, saddly/gladly (depends on your view) the FX9590 will be top dog from AMD until the new stuff.

Oh and don`t get your hopes too high in 2016, if AMD manages to match the IPC of the CPUs available NOW, it will be a huge success! but Intel will surely improve their CPUs with Skylake... the best i expect AMD to do is to get close to performance of Skylake, that`s best case scenario... let`s be realistic, but i`m in if AMD manages this.

Totally with you on that. I remember when the 9590 came out and it was the craziest thing (especically that it sold for over $700 for decent amount of time). Then to the fact that a 3.4Ghz 3570k can still blow that thing out of that water. Building on what you said, with their limited budget there's really no where to run far. Also considering the cheaper price of their CPU's some of the manufacturer errors, and other factors, it's hard to determine where they'll end up other than either failing and going bankrupt (at least for their CPUs) or just making it by.

I was browsing and I don't know if this is a error or not:

See next to the name of the 8350 on the results? What does it mean when it says, "4C / 8T"?

4 Cores / 8 Threads? That doesn't make sense though, isn't it a 8-core CPU?

Just wondering >.>

One other thing that I think most people know but I like to point it out:

In this instance this Vishera FX-8350 is running at 4.85 Ghz with a voltage of 1.476 V (on a 32nm die).

To compare, my Ivy-Bridge i5-3570k is running at 4.69 Ghz (4.7) with a voltage of 1.240 V (on a 22nm die).

Now if AMD weren't having money and budget issues, wouldn't they just try to copy or make similar products as Intel? That is what is interesting. The technology is already out there, and they really shouldn't have a problem with making a better processor, but it is their budget that is really holding them back (all in my opinion).

 

[juanrga]

Now expect 200+ reposts by Juan that "hire" actually means "fire".

This counts as another of your failed predictions.

Now expect 200+ reposts by Juan that "hire" actually means "fire".

Be nice he is from Spain they have a 25% unemployment rate its probably hard to tell the difference between hired and fired

LOL

My favorite part of this so far. Juan likes to think of replacements smarter people at Amd like to think of newer markets that co-exist with the older stuff

It is hard to say what is your new misunderstanding of my words, because you are being too cryptic and avoiding details or exact quotes.

 

[juanrga]

@colinp nevermind was thinking 4970k not 4790k, meh, what I get for working and posting. (They have launched the 8 core, right? Or was that next cycle?)

The 8-core Haswell has not been launched still. Moreover, it is named 5960X.

 

[esrever]

High end desktop is dead. No money in it for AMD. Them trying to bring to market a new CPU would only lose them money at this point.

 

[juanrga]

One thing about x86 vs ARM that I don't see mentioned much is what happens when you get added instructions in an environment you can use them. Haswell brought AVX2 and look what it did for Dolphin Emulator. It's ridiculously faster. ARM might be catching up to x86 when the x86 is running nothing higher than SSE2 code, but if you were to compare ARM IPC to Haswell IPC in something like Dolphin Emulator when Haswell is using AVX2, it'd be an absolute bloodbath.

Even Haswell vs IB is an absolute bloodbath. And ARM being RISC while x86 being CISC will always have that sort of situation. IPC is meaningless comparison if both CPUs can do 20 instructions in a given unit of time, but the x86 is doing FMA and all these operations in single instructions while the RISC is doing something like a simple FMA in a few instructions.

Piledriver/Steamroller only gives 4 FLOP/core when using SSE2 code.

An standard A57 core can bring 8 FLOP/core when using ARM NEON extensions.

The server-like Piledriver/Steamroller arch has to run FMA4 code to match the 8 FLOP/core of the phone-like A57.

I have seen some diagram of a future ARM core (server-like) that is now under testing and that will be fabbed the next year and each ARM core includes 256-bit Neon units plus SMT4. This ARM core will bring 16 FLOP/core plus SMT4. Engineers claim their core will provide 90% of the single thread performance of a Xeon Haswell core.

Your discussion about IPC ignores two points: first because the modern x86 CPUs are internally RISC machines. X86 CISC code is translated to RISC-like uops which are then run by the execution units. E.g. a given x86 instruction can be translated to 200 uops, and the execution units run those 200 uops.

When I say that a given core can do up to 16 FLOP/core, this is independent of how many uops the engine needs internally to provide that throughput. It doesn't matter if the core internally runs one uop or if runs 700 uop.

Second, when I discuss IPC I am not referring to the native ISA instructions but to the source code instructions. E.g. how many instructions of this piece of 1000 lines of C can the CPU run per second? This is why we can compare the performance of, say, C-ray on different CPUs with different ISAs. If chip A @ 3GHz spendz 30 seconds on C-Ray bench and chip B @ 3GHz spendz 15 seconds, one can say that the Chip B has 2x higher IPC.

 

[szatkus]

Second, when I discuss IPC I am not referring to the native ISA instructions but to the source code instructions.

It's usually compiled.

 

[etayorius]

Second, when I discuss IPC I am not referring to the native ISA instructions but to the source code instructions.

It's usually compiled.

I don`t think i qouted that... maybe you were trying to quote Juan?

 

[8350rocks]

Jim Keller has had tons of time in the industry working on projects since K7...if he says it will be better, I have no reason to doubt. Though, like those I know at AMD, I consider it a win if they are within 5%.

To be accurate, we are discussing what you are claiming that someone said you about what Keller said him... Also your claim is morphing at a fast rate: first, you said us that AMD "will be better by skylake". Next you added that "if they fall short of predictions they will be 'on par'", now you are already opening the door to 5% behind Skylake.

To match Skylake core (+- 10%), AMD would multiply the per core performance by up to a factor of eight (8)

Steamroller: 8 FLOP/core
Skylake: 64 FLOP/core,

and would beat a process disadvantage as well: 14/16nm (Glofo/TSMC) vs. Intel 10nm.

Moreover, AMD is bringing MANTLE to reduce CPU overhead. If AMD releases a CPU that is faster than anything from Intel and then brings MANTLE in final form to the arena, AMD will be competing against themselves because MANTLE would help Intel to remain competitive with an inferior CPU.

In short, this idea of AMD's 2016 x86 core outperforming Skylake on performance doesn't make sense: neither technologically, nor economically, nor strategically.

I agree, there is no chance AMD cant even match Skylake, even less outperform it with their next Arch... BUT, AMD does not have to beat Intel in the IPC game, it only needs to be competitive and offer better prices and AMD would be back at the game.

Skylake will probably be a good 15-25% faster than what ever AMD Throws with Keller, but hey! 25% behind is much better than being 60% behind, just as they are now.

So much hyperbole...

Show me where AMD is 60% behind in FPS in any game 8350 vs. 4670k or 4770k

Please, I am waiting for examples.

Couldn't ask for anything easier

(it's not 4670, so you can add few FPS)
http://techreport.com/r.x/amd-fx-8350/skyrim-fps.gif
http://techreport.com/r.x/amd-fx-8350/civv-lgv.gif

There are a lot of charts like this at Google Images.

If you want to cherry pick...there is a benchmark for the FX 8350 in this thread somewhere...(pages and pages ago) where the 8350 was 800% faster than the 3930k.

Show me a new game that is not run on basically 1-2 cores that does that...you might as well have put up iTunes or some other retarded one sided benchmark.

 

[8350rocks]

One thing about x86 vs ARM that I don't see mentioned much is what happens when you get added instructions in an environment you can use them. Haswell brought AVX2 and look what it did for Dolphin Emulator. It's ridiculously faster. ARM might be catching up to x86 when the x86 is running nothing higher than SSE2 code, but if you were to compare ARM IPC to Haswell IPC in something like Dolphin Emulator when Haswell is using AVX2, it'd be an absolute bloodbath.

Even Haswell vs IB is an absolute bloodbath. And ARM being RISC while x86 being CISC will always have that sort of situation. IPC is meaningless comparison if both CPUs can do 20 instructions in a given unit of time, but the x86 is doing FMA and all these operations in single instructions while the RISC is doing something like a simple FMA in a few instructions.

Piledriver/Steamroller only gives 4 FLOP/core when using SSE2 code.

An standard A57 core can bring 8 FLOP/core when using ARM NEON extensions.

The server-like Piledriver/Steamroller arch has to run FMA4 code to match the 8 FLOP/core of the phone-like A57.

I have seen some diagram of a future ARM core (server-like) that is now under testing and that will be fabbed the next year and each ARM core includes 256-bit Neon units plus SMT4. This ARM core will bring 16 FLOP/core plus SMT4. Engineers claim their core will provide 90% of the single thread performance of a Xeon Haswell core.

Your discussion about IPC ignores two points: first because the modern x86 CPUs are internally RISC machines. X86 CISC code is translated to RISC-like uops which are then run by the execution units. E.g. a given x86 instruction can be translated to 200 uops, and the execution units run those 200 uops.

When I say that a given core can do up to 16 FLOP/core, this is independent of how many uops the engine needs internally to provide that throughput. It doesn't matter if the core internally runs one uop or if runs 700 uop.

Second, when I discuss IPC I am not referring to the native ISA instructions but to the source code instructions. E.g. how many instructions of this piece of 1000 lines of C can the CPU run per second? This is why we can compare the performance of, say, C-ray on different CPUs with different ISAs. If chip A @ 3GHz spendz 30 seconds on C-Ray bench and chip B @ 3GHz spendz 15 seconds, one can say that the Chip B has 2x higher IPC.

No, this is precisely the misinterpretation he is referring to...if you can compile something to run AVX2 versus another CPU running it at SSE2, then the AVX2 CPU looks much faster by default, even if the IPC is similar...because the capability to do different operations can be shortcut into bigger chunks, and thus can be done more quickly...

 

[szatkus]

Jim Keller has had tons of time in the industry working on projects since K7...if he says it will be better, I have no reason to doubt. Though, like those I know at AMD, I consider it a win if they are within 5%.

To be accurate, we are discussing what you are claiming that someone said you about what Keller said him... Also your claim is morphing at a fast rate: first, you said us that AMD "will be better by skylake". Next you added that "if they fall short of predictions they will be 'on par'", now you are already opening the door to 5% behind Skylake.

To match Skylake core (+- 10%), AMD would multiply the per core performance by up to a factor of eight (8)

Steamroller: 8 FLOP/core
Skylake: 64 FLOP/core,

and would beat a process disadvantage as well: 14/16nm (Glofo/TSMC) vs. Intel 10nm.

Moreover, AMD is bringing MANTLE to reduce CPU overhead. If AMD releases a CPU that is faster than anything from Intel and then brings MANTLE in final form to the arena, AMD will be competing against themselves because MANTLE would help Intel to remain competitive with an inferior CPU.

In short, this idea of AMD's 2016 x86 core outperforming Skylake on performance doesn't make sense: neither technologically, nor economically, nor strategically.

I agree, there is no chance AMD cant even match Skylake, even less outperform it with their next Arch... BUT, AMD does not have to beat Intel in the IPC game, it only needs to be competitive and offer better prices and AMD would be back at the game.

Skylake will probably be a good 15-25% faster than what ever AMD Throws with Keller, but hey! 25% behind is much better than being 60% behind, just as they are now.

So much hyperbole...

Show me where AMD is 60% behind in FPS in any game 8350 vs. 4670k or 4770k

Please, I am waiting for examples.

Couldn't ask for anything easier

(it's not 4670, so you can add few FPS)
http://techreport.com/r.x/amd-fx-8350/skyrim-fps.gif
http://techreport.com/r.x/amd-fx-8350/civv-lgv.gif

There are a lot of charts like this at Google Images.

If you want to cherry pick...there is a benchmark for the FX 8350 in this thread somewhere...(pages and pages ago) where the 8350 was 800% faster than the 3930k.

Show me a new game that is not run on basically 1-2 cores that does that...you might as well have put up iTunes or some other retarded one sided benchmark.

I didn't cherry pick, it was third sensible (other was about GPU) chart in Google Images.
I remember that benchmark, also I remember some benchmark where A10 beats i5, but these were really specific cases.

Overall performance in actual software (which usually sucks in multi-threading) is the most important thing, when you compare CPUs.

 

[szatkus]

Second, when I discuss IPC I am not referring to the native ISA instructions but to the source code instructions.

It's usually compiled.

I don`t think i qouted that... maybe you were trying to quote Juan?

Sorry, it was Juan.

 

[juanrga]

jdwii you might as well buy a console if your recommending a pentium and 250X. You proclaim to build many of these pentium gaming boxes but yet you have not tested one? Or is it you only test when you unload services and background programs to fool the buyer into thinking how great it is. Nice way to do business.

+1

 

[juanrga]

Whatever. They have them, they know how they works, when they spot something better than their own parts they could use that knowledge to improve own blocks. Ideas aren't patented.

There is no special ideas around the A57 core. It is a standard implementation of ARMv8 ISA. Most companies purchased an ISA license and are developing/implementing their own custom cores. The only reason why AMD is using A57 cores now is because they are late at the party and need something to compete before their K12 core is ready.

 

[juanrga]

I said: Jim Keller expects them to be ahead by skylake

Your original words were:

JK says they will be better by skylake.

http://www.tomshardware.co.uk/forum/352312-28-steamroller-speculation-expert-conjecture/page-283#13631771

Jim Keller has had tons of time in the industry working on projects since K7...if he says it will be better, I have no reason to doubt.

http://www.tomshardware.co.uk/forum/352312-28-steamroller-speculation-expert-conjecture/page-283#13624062

You have changed the words from certainty to hope.

In any case when I was preparing the BSN article about Kaveri the AMD Engineering Dept. claimed that Steamroller would bring 45% IPC gains. I reduced that by 15%, then reduced it by another 5% and used the remaining figure for my article. Final measurements have shown that my analysis of the Steamroller architecture was accurate up to the single digit percent.

I note that you have ignored any of my technical, economic, and strategic arguments about the post-Excavator arch.

 

[juanrga]

No, this is precisely the misinterpretation he is referring to...if you can compile something to run AVX2 versus another CPU running it at SSE2, then the AVX2 CPU looks much faster by default, even if the IPC is similar...because the capability to do different operations can be shortcut into bigger chunks, and thus can be done more quickly...

No. That is a pretty trivial comment. It is so trivial as mentioning that the octo-core FX-8350 gives a maximum throughput of 256 GFLOP/s only if the code run on the 8 cores. One would don't confound the IPC allowed by the architecture with software that is or isn't optimized for that architecture.

I would like to point one thing out...as at the time we did not know...

When Jim Keller said, "AMD are on track to catch Intel in high performance cores"

The exact quote was:

AMD are on track to catch up on high performance cores

And the funny part is that I know to whom wrote that article.

 

[8350rocks]

@juan

I have never changed what Jim Keller said. He says it will be better. I do not doubt.
My source says if they shoot for better, and fall a bit short, they are still on par. Which I think is reasonable.

Either way, I am done talking with you...we will see who is right. I will hang your head on a pike when you are wrong beyond doubt for all to see as well. You will not be able to twist your words about ARM we have 100+ pages of your rubbish.

 

[Fidgetmaster]

AMD being late to the party and needing something to Compete....That sums them up very well currently haha

 

[vmN]

@8350rocks

You most understand why people aren't jumping right onboard on the "AMDs next x86 arch will be better than skylake". We have little to no information on the architecture itself.

I have stated, that I doubt AMD's next x86 arch will reach skylakes performance (first, we don't even know skylakes performance yet, we are only estimating based on their earlier architectures, and some of the things it features (which we know for certain, like AVX512)).

Don't hype the hypetrain to much without anything backing it up.

I'm not saying I'm not excited about their upcoming x86 arch, I'm just waiting for some more information, to properly place and estimation on it's performance.

 

[jdwii]

jdwii you might as well buy a console if your recommending a pentium and 250X. You proclaim to build many of these pentium gaming boxes but yet you have not tested one? Or is it you only test when you unload services and background programs to fool the buyer into thinking how great it is. Nice way to do business.

I see now you just wanted to argue with someone yesterday.

Although the discussion over banks(although i stand by what i say) will be stopped i will stand by what i claim a Pentium and a 250X is Superior to gaming then a A10 i ran several benchmarks many times. I stand by what i sell and the people who buy the builds i make(i don't own a business its just free lance)never complain about anything. I do not turn off any background services and i always install AVG free. Under general use i will admit a A10 is better it even feel's faster but for gaming no, just no unless its also paired with a video card.

 

[jdwii]

Spoiler

Now expect 200+ reposts by Juan that "hire" actually means "fire".

This counts as another of your failed predictions.

Now expect 200+ reposts by Juan that "hire" actually means "fire".

Be nice he is from Spain they have a 25% unemployment rate its probably hard to tell the difference between hired and fired

LOL

My favorite part of this so far. Juan likes to think of replacements smarter people at Amd like to think of newer markets that co-exist with the older stuff

It is hard to say what is your new misunderstanding of my words, because you are being too cryptic and avoiding details or exact quotes.

You always claim how Arm will replace this and that. Arm will co-exist with X86 for the foreseeable future.

 

[jdwii]

@8350rocks

You most understand why people aren't jumping right onboard on the "AMDs next x86 arch will be better than skylake". We have little to no information on the architecture itself.

I have stated, that I doubt AMD's next x86 arch will reach skylakes performance (first, we don't even know skylakes performance yet, we are only estimating based on their earlier architectures, and some of the things it features (which we know for certain, like AVX512)).

Don't hype the hypetrain to much without anything backing it up.

I'm not saying I'm not excited about their upcoming x86 arch, I'm just waiting for some more information, to properly place and estimation on it's performance.

Amd seems to have management fixed now vs back in the day where it had problems read this its old
http://kotaku.com/visit-south-park-thanks-to-the-magic-of-virtual-realit-1600134381?utm_campaign=Socialflow_Kotaku_Facebook&utm_source=Kotaku_Facebook&utm_medium=Socialflow

Bulldozer was once a dropped project as you probably already knew. Something or someone brought it back. Intel has a lead in their foundries compared to everyone else. I do not believe its possible for Amd to ever be faster overall compared to Intel but it did happen back in the P4 day's. Juan makes some good points on this however, and i kinda agree with him on them.

What Amd really needs is their prediction unit to be more accurate back in the early BD day's the team claimed that the 3rd ALU was barley touched on phenom but then we see 4 per core on Intel. They have a lot of work to do to improve on this.

 

[szatkus]

Whatever. They have them, they know how they works, when they spot something better than their own parts they could use that knowledge to improve own blocks. Ideas aren't patented.

There is no special ideas around the A57 core. It is a standard implementation of ARMv8 ISA. Most companies purchased an ISA license and are developing/implementing their own custom cores. The only reason why AMD is using A57 cores now is because they are late at the party and need something to compete before their K12 core is ready.

A57 includes thousands of blocks. Chances are that some of them are better than ones in AMD's library.

Every engineer will understand it, but again you're smarter than everyone else.

 

[szatkus]

jdwii you might as well buy a console if your recommending a pentium and 250X. You proclaim to build many of these pentium gaming boxes but yet you have not tested one? Or is it you only test when you unload services and background programs to fool the buyer into thinking how great it is. Nice way to do business.

I see now you just wanted to argue with someone yesterday.

Although the discussion over banks(although i stand by what i say) will be stopped i will stand by what i claim a Pentium and a 250X is Superior to gaming then a A10 i ran several benchmarks many times. I stand by what i sell and the people who buy the builds i make(i don't own a business its just free lance)never complain about anything. I do not turn off any background services and i always install AVG free. Under general use i will admit a A10 is better it even feel's faster but for gaming no, just no unless its also paired with a video card.

This looks like next neverending Phenom II X4 vs i3 discussion, where no one is right, because every game is different.