AMD CPU speculation... and expert conjecture

[JAYDEEJOHN]

Similar to old K8 @ 2.6 vs C2D @ 1.83

 

[$hawn]

I repeat again, all this is just my fancy guess work,

i see that you missed the part where we need to multiply by ipc factor (i.e, multiply fx core by 0.66 i.e, 800x0.66) and where we neet to consider clock diffence too

I've already considered IPC into the equation, as my calculations are based on the assumption that a 5GHz PD core is equivalent only to a 3.3GHz SB core. Just read thru it one more time, IMHO my calculations stand correct.

 

[8350rocks]

I repeat again, all this is just my fancy guess work,

i see that you missed the part where we need to multiply by ipc factor (i.e, multiply fx core by 0.66 i.e, 800x0.66) and where we neet to consider clock diffence too

I've already considered IPC into the equation, as my calculations are based on the assumption that a 5GHz PD core is equivalent only to a 3.3GHz SB core. Just read thru it one more time, IMHO my calculations stand correct.

But, the difference in IPC currently is only 32 IPC (8 core AMD) vs. 36 IPC on i7 SB...

The difference is only the efficiency of the instructions handling...or were you aware of the actual architecture IPC values?

IPC is not the gap...the front end architecture is...that's why Steamroller will be so amazing...30% increase in effectiveness and IPC increase from 32 to 44...(which will be more IPC than Ivy, and likely Haswell with the exception of the 6 core intel's they will be 48 IPC in Ivy-E)

 

[sarinaide]

From what I am told Richland is impressive, higher clocks at lower power and a sizeable improvement in CPU+GPU performance with only a 44mhz bump on frequency. This is where AMD is doing right.

It is looking like GT3 will only be viable on notebooks/laptops but from what I heard top end GT3 notebooks are looking up on $1500 which will make it limited to certain people, a hard core user wanting a laptop with significant grunt will still go for a much cheaper option with AMD or Nvidia graphics, so for that GT3 actually failed in that it was intended to remove the need for discrete graphics. The other aspect is Richland notebooks at half the cost also showing massive iGPU gains could depending on the persons needs be a strong price point competitor and apeal more to the market that cannot afford expensive systems.

Its very interesting to see the differing approaches though as above AMD on very little clock bumps is achieving on proportion much better gains relative to intels beast which is impressively clocked at 1.3ghz, architecturally its impressive and testament to Intel's FAB, but if you are doing a relative inspection of intel vs AMD HSA, for intels next iGPU to match Kaveri it will basically need 4-5x the resources and probably 50% clock speed bump which is not sustainable, again it comes down to what a GPU really is and AMD and Nvidia are exceptional in that regard, their GPU's are getting smaller and faster and more efficient. It begs the question will Intel buy or join with Nvidia, personally I doubt it, Nvidia is doing well on their own and the last time they partnered intel they got burned so there is no incentive for team green.

 

[noob2222]

Just curious, links to real world apps?

All these numbers I've put up are a result of reading numerous comparison articles, some common sense and a little guess work. Do NOT treat this as scientific data!!

If you want to know how i arrived at the 50% faster value, have a look at this AAC encoding chart (audio encoding is my favourite single core performance measuring yardstick).
http://media.bestofmicro.com/Y/4/357628/original/itunes.png

SB takes 62 seconds @3.8GHz turbo.
As PD takes 84 secs @4.2GHz turbo, so it should logically take 93 seconds @3.8GHz

IPC ratio of PD/SB in AAC encoding = 62/93= 0.66. In other words, SB is 1/0.66 => 1.51x PD IPC, ie, its 51% faster.

I repeat again, all this is just my fancy guess work, and for all you know I could be very wrong. Do correct me if I am. My only intention was to put into perspective what a 5GHz centurion chip would look like against a SB-E chip. Cheers

relying on one lopsided benchmark will give you your lopsided point of view.

how about using something thats more to the norm as amd/intel goes? 50% rofl. on average its actually closer to 20% as seen here. After all, if its the 50% that you claim, this should never happen:

so whats the problem with Itunes? why is it so far off from all other results? Does it actually reflect RL results? only if you use it, othewise NO.

kinda funny that Itunes and Cinebench run close to the same single threaded result, and Cinebench is known to be compiled on ICC.

 

[noob2222]

The other missing part of your assumption is they both are running at their max turbo core speed, same for taking into account intel's base speed = 4 core speeds.

ran average of 3.9 ghz and not the maximum of 4.2

averaged to 3.5 ghz and not its maximum of 3.6, but look at this one:

your 3.3ghz SB never reached 3.3Ghz, it stayed at 3.4 on a full 4 core cpu encoding run.

Truth of the matter is, there are soo many variales that IPC is a defunct term. you have turbo speeds changing the results, software itself (30% speedup for amd by using povray over itunes), wich functions does the cpu support (sse, avx, fma) as well as countless other factors.

 

[noob2222]

On another note, DDR 4 memory looks like its going to start around CL 13-14 for 1866. If you were hoping DDR4 was going to lower latency, sorry, the answer looks to be NO.

http://www.micron.com/products/dram/ddr4-sdram#fullPart&236=0

 

[sarinaide]

Nice work noob whomever says Intel is 50+% faster has certainly bitten the sensationalist apple, basically Intel and Nvidia have been able to spin up stories that are completely falsified and has the effect of taking marketshare, the most recent was that FCAT nonsense which was trying to tell people that Crossfire microstutters all the time yet most whom have crossfire will say while not perfect is anything but microstuttering 24-7, most whom also SLI will almost say that SLI vs CF is almost the same just that Nvidia have better SLI support on their drivers. What happened they spun up a story that every tom (no pun) dick and harry latched onto like pitbull.


Anyways at no point was Intel 50% faster in x86 operations, that number is inflated and I would say best case scenario is 15-20% top limit.

As for DDR4, its bandwidth not latency driven and its not like AMD or Intel does much with tight latencies anyway, bandwidth has the sizeably bigger improvement.

so whats the problem with Itunes? why is it so far off from all other results? Does it actually reflect RL results? only if you use it, othewise NO.

kinda funny that Itunes and Cinebench run close to the same single threaded result, and Cinebench is known to be compiled on ICC.

The first bit is MEME ready. "In real life, I Itunes"

Basically thats what it ultimately is, who ever said that iTunes was ever a meaningful synthetic, first its apple and well that about runes it for life and honestly have you seen how rubbish it is to navigate around, I hate to say this but Windows Media Player is infinitely better than iTOONS.

Again its the way Intel is marketed and not in part without the complience from online review sites, they can mask around the truth to create a sensation that Intel is 2x faster than a AMD system when as most of us whom own both will tell you thats hardly true and its only in instances, then when you scale to real world apps that never manifests into major differences.

 

[-Fran-]

This is why turbo schemes obscure the big picture, ugh.

Still, I think my assertion regarding the OC between FX'es and the Core i series stand. For every Mhz increase Intel gets, AMD has to double it (3:2 ratio was my guesstimation). I know for stocks it doesn't matter, but we are all from that special 5% (or less) that actually should care more for OC figures than stock. Anyone with energy to make something on that premise? truegenius? haha.

Now, to be fair. I just assembled a FX4300 for a friend and OC'ed to 4.3Ghz (225x19); it's faster, side by side, than a i3 on everything. Even games. Now, it was a gap of around 1Ghz, so it didn't come to any surprise. Still, for the price, it's a better buy when OC'ed.

Cheers!

 

[JAYDEEJOHN]

http://semiaccurate.com/2013/04/15/putting-intels-new-hd-4000-drivers-to-the-test/
Thought I would drop this here for entertainment.
Not looking good heheh

 

[sarinaide]

http://semiaccurate.com/2013/04/15/putting-intels-new-hd-4000-drivers-to-the-test/
Thought I would drop this here for entertainment.
Not looking good heheh

APU wins flawless victory.......FATALITY!

HD4000 is horrible even in single player, well the good news is HD4600 may pass 30 FPS in BF3. If you want graphics on chip AMD is just so head and shoulders above the competition its pointless to consider the alternative.

[EDIT] I was told these new drivers where going to be extraordinary, on the median of improvement it was about a few percent, thats not going to make me sell my Trinity, it is entertaining though to see all the sad Intelmaniacs who last year touted Haswell the "trinity slayer" the "skyrim on ultra at 1080" and lately the "GT650M slayer" so far its looking well, well, well short like a leprechaun in butt kicking contest. Even if by some miracle GT3 was equal to a GTX650M, it will cost more than a system with a GTX650M so not really achieving much and nobody knows power yet but what GT3 is packing it will be thirsty.

 

[bobbybamf12]

FM2 WHAT!!!!??????
WHAT!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
WHYYYYYYYYYYYYY WOULDDDDDDDDDDDDD YOUUUUUUUUU DOOOOOOOOO THAT!?!?!?!?!?!?!?!?!!??

(HEAD EXPLODED)

 

[sarinaide]

Because I can

 

[esrever]

http://www.techpowerup.com/182791/AMD-Works-on-Next-TWKR-Chip-Based-on-Vishera.html

So the 5GHz chip will be for overclockers to break records with and not for consumers. This makes more sense. If you want to have the most mhz, AMD is where you go and not having to buy dozens of 8350s to bin them is a plus.

 

[Cazalan]

If GT3 ends up costing an arm and a leg it will only be useful in extreme low profile UltraBooks. Everyone else will just bundle a discrete. It's starting out as an experiment for Intel. How high can they charge before people lose interest. Maybe Apple buyers will gladly pay the premium but not many others.

If the CPU can also use the cache as L4 then some other operations may be sped up as well.

 

[Cazalan]

http://www.techpowerup.com/182791/AMD-Works-on-Next-TWKR-Chip-Based-on-Vishera.html

So the 5GHz chip will be for overclockers to break records with and not for consumers. This makes more sense. If you want to have the most mhz, AMD is where you go and not having to buy dozens of 8350s to bin them is a plus.

That's just their spin on the same single leak from Hexus. The entire rumor is off of one source, that's been replicated 40+ times.

 

[griptwister]

APU's are beast lol. And "Because I can" is always a great answer lol.

I was thinking about this...

Even if AMD has a slightly lower IPC on their steamroller chips, it seems as if SR will have higher clockspeeds than intel and have OC headroom. IMO, AMD might finally take the crown back.

I'm a bit curious to see how this 64mb L4 cache works out for intel, wouldn't things have to be optimized to use the L4?

 

[de5_Roy]

if hexus' rumor is true, 'centurion''s pricing is good. goes to show that amd will also charge super high price when they get a chance. technically, centurions will be 'massively' undercutting core i7 3970x (not the 'cheaper' core i7 3930k that no one cares about), so amd is still presenting a lot of value.

 

[gamerk316]

Just curious, links to real world apps?

All these numbers I've put up are a result of reading numerous comparison articles, some common sense and a little guess work. Do NOT treat this as scientific data!!

If you want to know how i arrived at the 50% faster value, have a look at this AAC encoding chart (audio encoding is my favourite single core performance measuring yardstick).
http://media.bestofmicro.com/Y/4/357628/original/itunes.png

SB takes 62 seconds @3.8GHz turbo.
As PD takes 84 secs @4.2GHz turbo, so it should logically take 93 seconds @3.8GHz

IPC ratio of PD/SB in AAC encoding = 62/93= 0.66. In other words, SB is 1/0.66 => 1.51x PD IPC, ie, its 51% faster.

I repeat again, all this is just my fancy guess work, and for all you know I could be very wrong. Do correct me if I am. My only intention was to put into perspective what a 5GHz centurion chip would look like against a SB-E chip. Cheers

In theory, yes, but the heavily shared arch of BD/PD makes the computation a bit tougher, as increasing the clockspeed also effects the shared backend. I'd expect BD/PD scales slightly faster then linearly with clockspeed.

From what I've seen, a BD/PD MODULE is about 75-80% the performance of a Intel core, at the same clock. AMD clocks more agressively though...

 

[gamerk316]

But, the difference in IPC currently is only 32 IPC (8 core AMD) vs. 36 IPC on i7 SB...

The difference is only the efficiency of the instructions handling...or were you aware of the actual architecture IPC values?

IPC is not the gap...the front end architecture is...that's why Steamroller will be so amazing...30% increase in effectiveness and IPC increase from 32 to 44...(which will be more IPC than Ivy, and likely Haswell with the exception of the 6 core intel's they will be 48 IPC in Ivy-E)

IPC is NOT a flat value, and varies by workload. Throw in the shared backend, and you can easily get pipeline stalls (which is blamed as one of the reasons for BD's poor single threaded performance). This is especially notable in FP workloads, as the FP scheduler isn't shared like the integer one is.

So anyone claiming to have absolute numbers on IPC is kidding themselves. You can only solve IPC PER APPLICATION.

I'll use this as an example of the math:

As its single threaded, I can safely disregard any pipeline stalls. I'm going to do the math assuming no turbo for simplicity sake though, so these numbers won't be perfect...

I'll compare the 4300 and 2500k.

AMD FX-4300: 3.8GHz
Intel i5 2500k: 3.3GHz

Solving for IPC:

Time = NumberCores * Clockspeed * IPC

IPC = Time / NumberCores * Clockspeed

For AMD (remember: Single Threaded):

IPC = 236.3 / 1 * 3.8
IPC = 236.3 / 3.8
IPC = 62

For Intel (2500k):

IPC = 274.9 / 1 * 3.3
IPC = 274.9 / 3.3
IPC = 83

And just for kicks, the 3570k:

IPC = 302.2 / 1 * 3.4
IPC = 302.2 * 3.4
IPC = 88

Now lets solve multithreaded:

FX-4300:

IPC = IPC = 983.6 / 6 * 3.8
IPC = 983.6 / 22.8
IPC = 43

Note how IPC decreased? Thats likely either a failure to fully utilize all the cores, or the shared backend robbing performance.

2500k:

IPC = 1012.8 / 4 * 3.3
IPC = 1012.8 / 13.2
IPC = 77

Hence why the i5 at a lower clock still beat the 4300: Superior IPC. But note how IPC only dropped by 6, compared to the drop of 19 for AMD: This indicates that the i5 actually scales better then BD/PD (again: could be pipeline stall, lack of core loading, and other factors).

3570k:

IPC = 1108.9 / 4 * 3.4
IPC = 1108.9 * 13.6
IPC = 81.5

Same story here: IPC drops slighty (7.5), but less then half as much as AMD. Farther indication theres a problem somewhere as workload starts to scale (not a good sign for an arch designed to scale).

As far as this PARTICULAR application goes: Intel has far superior IPC, and scales better in the multithreaded bench, even if it looses in pure performance.

Farther, you can break down IPC per core, which nets this:
AMD FX-4300: 7
Intel i5 2500k: 19.25
Intel i5 3570k: 20.375

Hence why AMD only wins when clocked higher and when all the cores are used: Its per core performance is less then half as much. (Do remember that shared backend though).

So yeah, your IPC numbers are kinda worthless in hindsight. You can only solve PER APPLICATION, as loading factors will vary with different apps.

 

[sarinaide]

APU's are beast lol. And "Because I can" is always a great answer lol.

I was thinking about this...

Even if AMD has a slightly lower IPC on their steamroller chips, it seems as if SR will have higher clockspeeds than intel and have OC headroom. IMO, AMD might finally take the crown back.

I'm a bit curious to see how this 64mb L4 cache works out for intel, wouldn't things have to be optimized to use the L4?

1) Yes the APU is good, and its only the beginning, nobody has anything that can match it within realms of reason so it is the most technologically advanced CPU on the market. The upsides are immense.

2) Lets hope that AMD close up the IPC stakes enough to make team blue fret their knickers a bit

3) GT3 comes with a premium price tag and as before 1.3ghz is really high clocks, I can't see that kind of proliferation sustainable when Intel are a) power conscious and b) AMD's graphics SIMD will get smaller, faster and more efficient then scale well. Dedicated GDDR5 will unleash the Radeon engine and these horses will gallop.

 

[gamerk316]

3) GT3 comes with a premium price tag and as before 1.3ghz is really high clocks, I can't see that kind of proliferation sustainable when Intel are a) power conscious and b) AMD's graphics SIMD will get smaller, faster and more efficient then scale well. Dedicated GDDR5 will unleash the Radeon engine and these horses will gallop.

GPU wise? Yeah.

CPU wise? I very much doubt it. I really can't stress how much I dislike GDDR5 as system memory. Nevermind the cost...

 

[sarinaide]

3) GT3 comes with a premium price tag and as before 1.3ghz is really high clocks, I can't see that kind of proliferation sustainable when Intel are a) power conscious and b) AMD's graphics SIMD will get smaller, faster and more efficient then scale well. Dedicated GDDR5 will unleash the Radeon engine and these horses will gallop.

GPU wise? Yeah.

CPU wise? I very much doubt it. I really can't stress how much I dislike GDDR5 as system memory. Nevermind the cost...

Well from what has been said the iGPU will exclusively use its GDDR5, there is not a lot but there need not be. The CPU will use system memory DDR3/4. They mentioned 2 controllers and a ARM chip to manage the pageable space in system memory for HSA operations. Some speculation is HSA based applications will run up to 50% faster, the iGPU on GDDR5 will improve by a sizeable margin.

DDR4 latencies are not looking to impressive, but then again as you SPD goes up the latencies fall with that rise in bandwidth, that's the trade off, if you want tight latencies get DDR2 in there. Since bandwidth has the bigger influence, I will take that over latency.

 

[griptwister]

@sarinaide: Heck yeah man! These little APUs are incredible! I've got one of the older model APU desk tops. It's perfect for my dad's photo shop usage!

@gamerk316: Once the steamroller APUs come around, I'm convinced the CPU side of things will change. I'm willing to bet that the SR APUs will be competitive with the i5s in some areas.

If the Steamroller APUs come out before the SR performance line drops, I might just go that route! Especially if I can get 6 decent cores at a great price!!!

 

[bobbybamf12]

APU's are beast lol. And "Because I can" is always a great answer lol.

I was thinking about this...

Even if AMD has a slightly lower IPC on their steamroller chips, it seems as if SR will have higher clockspeeds than intel and have OC headroom. IMO, AMD might finally take the crown back.

I'm a bit curious to see how this 64mb L4 cache works out for intel, wouldn't things have to be optimized to use the L4?

I don't know about that Intel next cpu will have an even lower TDP then they do now, So they will probably have a higher overlocking headroom compare to AMD. If AMD sticks with 125w then the overclocking will probably be on par as it is now but I'm sure they will lower their TDP also.

Of course if you are going to overclock an Intel cpu you wll need to buy a K series. That's where AMD has it's advantage. I don't believe steamroller will be the cpu that surpasses Intel but i can see excavator doing so.