AMD CPU speculation... and expert conjecture

[sarinaide]

Adobe CS7 ?

Dont know if there will be HSA logos, i think it will, but if it is more than those other Adobe apps (which i think are not HSA cert) then the discussion of CPU core performance almost be over

http://community.amd.com/community/amd-blogs/amd-unprocessed/blog/2013/06/17/amd-performance-screams-with-just-released-adobe-photoshop-cc-and-adobe-premiere-pro-cc

" Poor lame AMD can't compete because the CPU is behind 20% ( on dubious benchs) ... its a beating! " frankly would seem quite ridiculous then

and there are a WHOLE LOT more(edit) ... including for Android (ARM) lol
http://www.amd.com/us/vision/shop/cool-apps/Pages/cool-apps.aspx

Thats the new Adobe Premier Pro, this is the first genuine HSA application which the HSA foundation has said many will follow but already Richland is developing a substantial lead, Kaveri will be nothing short of mindblowingly fast with hUMA adaptations.

 

[juanrga]

Supposed benchmarks of Steamroller FX chips

http://northwood.blog60.fc2.com/blog-entry-6834.html

AMD Steamroller FX 9650 - 4.5 Ghz / 4.8 Ghz Turbo Benchmarks! lol

If that is truth, than its not 32% better than the Fx8350, by the test results...

Lets cont only POV-RAY, the Cinebench doesn't count, after so many past posts i think nobody is going to be fouled.

POV-RAY
FX-8350: 1505
FX-9650: 1985

That is a 32% gain.

The Fx8350, in high performance, that is, in multithreading which is the only truly high performance metric( there isn't any other, ILP is dead) is already ~10% above the i7 3770 by that test... so it could be similar performance to the 4770.

Last Openbenchmarks show that a FX-8350 at stock is so faster or a bit more than the 4770k in multhreading. Moreover the FX OC better than Haswell.

 

[gamerk316]

I think games that run on PCs were simply ports.

Sometimes, yes. Red Faction: Guerrilla was a good example of a game ported to the PC after the fact. More often though, the PC port is handled in parallel, or is the lead platform. The "PC Port" argument was created to try and explain why we aren't seeing significant jump in graphics anymore, and its easier to blame consoles then the fact that Rasterization is tapped out [hence the focus on new AA and other post-processing effects].

 

[gamerk316]

If the FX9450 is similar clocks to the 8350, but then its only 10% above Piledriver by that test on POV-ray results !... so from where comes that 32% ?

It looks like the 9650 is 4.5GHz base clock, but can't find info on the 9450's clock. For reference: The 8350's base clock is 4GHz. So at least some of the performance is due to clock, not IPC. Without know the 9450's clock, really hard to make a comparision.

As far at the 8350/9650 comparison, we can solve the IPC difference based on the info we have:

For POV Ray:

Performance = IPC * Clockspeed * Number of Cores [Number of Cores is the same for both CPU's, so this cancels out]

8350: 1505 = IPC * 4
8350: IPC = 376.25

9650: 1985 = IPC * 4.5
9650: IPC = 441.1

Now to solve the performance difference:

IPC Difference = 100 - 376.25 / 441.1 * 100
IPC Difference = ~15%

For CineBench (single):

8350:

1.11 = IPC * 4
IPC = .2775

9650:

1.47 = IPC * 4.5
IPC = .3267

IPC Difference = 100 - .2775 / .3267 * 100
IPC Difference ~15%

So far, looks like 15% IPC gains. One more to go:

Cinebench (Multi)

8350:

6.93 = IPC * 4
IPC = 1.7325

9650:

9.15 = IPC * 4.5
IPC = 2.03

IPC Difference = 100 - 1.7325 / 2.03 * 100
IPC Difference ~ 15%

So it looks like Steamroller gains 15% in IPC over Piledriver, based on three benchmarks. Not bad for a generational leap.

And remember I'm solving for the DIFFERENCE in IPC, not the actual IPC. I've also simplified the equation and made a few assumptions (100% core loading). Quite possible the IPC number is slightly understated if SR isn't being pushed to 100%, or overstated if Turbo is being used excessively. But I'd say 15% is probably fairly accurate for CPU based benchmarks.

 

[gamerk316]

I will say, based on my analysis, SR is about where AMD needs to be at the top end to really compete with Intel for enthusiasts. Its what BD *should* have been when it was released. Very attractive buy if the price is right.

Really wish I know the 9450's base clock; would love to see if the 15% IPC increase holds up in a lower-clocked processor, or if other bottlenecks could come into play at a lower clock.

 

[Ranth]

I'm confused. Have AMD given steamroller fx's numbers yet? Or are the 9450/9650 numbers just rumors? I hope they are otherwise AMD is making it much more confusing with the numbering.

 

[The Q6660 Inside]

Not to change the topic too dramatically, but did anyone else catch this:

http://www.tomshardware.com/news/AMD-A10-6800K-Overclock-APU,23108.html

A10-6800K overclocked to over 8GHz on all cores. Impressive as I wouldn't have believed an APU able to overclock like that. Also 1.992 vcore - holy cow!! Even I wouldn't have believed an AMD APU or CPU could handle that kind of vcore without instant death.

The Richland APU is more capable than I gave it credit for, makes one wonder what Kaveri will be capable of. Still waiting for any information on Steamroller FX though!!

The fine print on that is throttling.

Cheers!

At -185*C there will be no thermal throttling, and I believe he validated it on HWBot so it was stable enough to be valid.

Sarinade, this is your chance to take a 6800K under LN2 and beat that record :3

Anyways, I am amazed at how the A10-6800K managed to break 8GHz. With this amazing feat, I am sure the Kaveri A10 XXXXK will break the FX8150's World Record.

@Juanrga, I agree that GT3e/Iris was a dare by Intel that did not deliver for the high-performance market (650M+ sector) that it was hyped to be and is quite power hungry, like HasBurst™ in general. You might as well get a 7970M+i7 laptop for the same price and have faaaar better frames.

Daily bonus video thingy time- http://www.youtube.com/watch?v=zUc6znC848o

Nah LN2 is to much work these days, I am into lower power APU's and killing people on BF servers

Iris was a case of to much improvement, and the rigidity of its pricing which is a lot in part to Intel flogging billions on Iris means its basically between a rock and hard place, as before Iris is only fast at lower resolutions sub 16:9 resolutions, crank it up to 16x9 and 19x10 resolutions and Iris loses around 65% of its performance to the point its only as fast as a HD7660D despite twice the bandwidth (tested with 2133 bandwidth ie:34GB/s). Since a 650M can play med settings at 1080, Iris is nowhere close to a 650M, its like 40% slower at that res and settings.

Sarinade, this is your chance to take a 6800K under LN2 and beat that record :3

Anyways, I am amazed at how the A10-6800K managed to break 8GHz. With this amazing feat, I am sure the Kaveri A10 XXXXK will break the FX8150's World Record.

If that kaveri comes only in "bulk" process, i bet you €1000 that it cannot... not even that Richland, not even close lol (TH could mediate the deal lol)

If it is 28nm PD-SOI then the bet is off ok!?

8150 passed 9ghz on a single core, the 8350 can do around 7.0ghz on 8 cores.

Refuted: BF4, other Frostbite 3 games to be 'optimized exclusively for AMD'
http://techreport.com/news/24976/refuted-bf4-other-frostbite-3-games-to-be-optimized-exclusively-for-amd
Rumor: AMD Gets Exclusive Optimization for all Frostbite 3 Games
http://www.pcper.com/news/Graphics-Cards/Rumor-AMD-Gets-Exclusive-Optimization-all-Frostbite-3-Games

I got a BF4 alpha trial pass, but as they say in the invite a lot of content is cut down to minimal including graphics engines, the sole purpose is to play and offer opinion on how they can improve the game by Beta.
#LONGPOST

Perhaps AMD will change their mind at the last second and get a 28nm or PD-SOI FD-SOI process . Realistically, it looks like AMD has just gotta adapt. However, I am still confident Kaveri A10-XXXXK will beat the 8150's broken record. I need those $1000 so I can build a Steamroller FX 8XXX/9XXX rig to replace my beloved i7-920 C0@3.8 main rig and hand-me-down but not really mine QX9770@3.7 gaming rig. Talk about power consumption . The FX9650 name seems odd, considering it was taken by the Phenom I (AKA: the 8150 of 2007).

 

[viridiancrystal]

Yes, XB1's presale numbers are abysmal and to add an interesting note to the AMD v. Intel debate...it looks like Intel is "slipping". Instead of Broadwell next year, they are doing a haswell "refresh".

http://www.fudzilla.com/home/item/31701-intel-plans-haswell-refresh-in-q2-2014

Imagine Steamroller launching and having to compete with hasfail for a full 18 months, by which time, if broadwell does finally poke it's head out of obscurity, the Excavator will likely be hitting the shelves.

Interesting...eh?

I like how you always say hasfail, as you know its amazing. Haswell is revolutionary, my brother was right: biggest improvement since 2003 when pentium m cpus were released on the mobile market. My brother said my ivy bridge laptop would look 10 years old when haswell would come out and he was right[strike] lol[/strike]. I wish I had waited for haswell now [strike]and[/strike]. The igpu has improved significantly, so much so amd's apus get crushed compared to intel's higher end hd 5000 series igpu. I remember reading 32nm pch, compared to the 65nm pch motherboard that was out for years [strike]and[/strike]. Not only that haswell takes [strike]a lot[/strike] less power on idle as well, as the igpu is 2x better performance per watt then ivy bridge, as the intel demonstration showed.

Haswell is the biggest improvement seen in years in mobile market, you could say haswell is a game changer, and amd must be scared now that their apus are pointless. Intel is in a position where in 2016, if haswell's succesor came out intel would still be ahead of amd, as they are ahead on performance per watt, and they can release [strike]a lot[/strike] more cpu cores.

Now let me see if I can take this seriously. Nope, did not help.

Spoiler

Supposed benchmarks of Steamroller FX chips

http://northwood.blog60.fc2.com/blog-entry-6834.html

My attempt at de-enrishifying your link. I fear that some will be lost in translation but perhaps somebody else can do better.

4.5 Ghz/4.8 Ghz Turbo Benchmarks!
9650 AMD Steamroller FX (AMDFX) achieves performance 32% greater rather than 45% greater. And the second generation "Piledriver" as compared to the first "Bulldozer" to "Steamroller". Achievements of 30% Ops/cycle increase in the ratio of "Piledriver" to "Steamroller", that is an IPC improvement of more than 20% has been achieved. The result, an extent to which is less than 25-30% in comparison, "Haswell" improvement is half way between "SandyBridge" and "Nehalem" in IPC. The scores of the benchmarks are listed as follows.

Core i7
3770K Core i7
3970X FX-8350 Steamroller
FX-9450 FX-9650
Number of Cores 4-Core
8-thread 6-Core
12-thread 8-Core
4-module 8-Core
4-module 8-Core
4-module
POV Ray 3.7 Multithread 1363 1888 1505 1636 1985
CineBench
11.5 Sinlge 1.66 1.62 1.11 1.21 1.47
Multi 7.52 10.84 6.93 7.55 9.15

However, note this small detail in that "32% increase in performance over the FX 8350 determined with" is attached. The possibility of stuff that made ​​up by 1.32 times the score of the FX-8350 seems to be high. There is a difference of about 20% performance in a single core of "Steamroller" and "Haswell". In a system with two high end graphics cards, the FX-9650 shows gaming performance equivalent to the i7 2600K, but is still inferior by one step to the i7 4770K. However, in titles that benefit from a core count of more than 4 cores, such as Crysis 3, the FX-9650 will achieve the same performance as rival product i7 4770K. And an improved IPC of 20% and a frequency enhancement of 10% will achieve performance improvements of 32% over "Piledriver". On the other hand, "Steamroller" with respect to overclocking is worse than ever. AMD poured a lot of effort to lower the power consumption of "Steamroller". But this effort was a way of achieving a higher frequency per TDP. The overall power consumption draws around 5-8% less than the FX-8350 as a result of that effort.

OC 43% can be from 3.50GHz to 5.00Ghz for the i7 4770K but, FX-9650 will remain at about 20% and 5.40GHz from 4.50GHz. It is assumed that the performance of the FX-9650 falls between the i7 4960X and i7 4770K. Prices also reflect that. Chips of 3.80GHz / TC 4.20GHz are to be lower model as the product you want to compete with i7 4770K separately seems to exist for this. Expected prices will be $ 449 for chip of 4.50GHz / TC 4.80GHz as a flagship. Chips of 3.80GHz / TC 4.20GHz and below would be around $ 249 than this. In other words, between the FX-9650 and the i7 4930K, the FX9450 is not in a position to counter the i7 4770K. I am referring to the possibility of FX CPU "Centuian" with 12-core / 6-module referred to the topic of "Kaveri" APU of 6-core/3-module in the last paragraph down. "Centurian", by the way, (code-named) is featured as the name of FX CPU of 5GHz previously. ... And even if you recall in a very suspicious, familiar story, some leading to the frenzy of the previous appearance of "Bulldozer". I welcome with open arms if this is true, but it remains too dubious of rumors about the "it 's Na glad you were very real" at the moment. AMD Steamroller Die shots leaked in photos of the die are "Steamroller" have been published. I suspect that this die photo is different or legitimate, and there is a possibility in various ways. I say what changes is applied significantly "Steamroller" thing you can see from this photo anyway. It is a remarkable branch prediction and front-end, the instruction cache data, especially in the FP unit. Looks significantly larger, as it accounts for a large area of two times compared to "Bulldozer / Piledriver" FP unit. Speculation seems to be flying around a lot. A mark shall be added that the 256-bit FMA × 2 128-bit FMA × 2 up to now, integer arithmetic unit of 2 + AGU × 2 ALU × the "floating-point unit There is also quite intense speculation that the data cache also changed as such "is increased to 32kB from 16kB and change, the ALU × 4 + AGU × 4 from the configuration.

Also worth noting: http://technewspedia.com/kaveri-amd-apus-have-integrated-pcie-3-0-controller/

I remember reading, I think somewhere way back on the bulldozer rumors thread, that much of the reason Intel did so much better in games was due to the on-die PCIe controller. I doubt it really does much, just food for thought.

 

[lilcinw]

/snip

http://www.freepatentsonline.com/y2012/0258611.html
(first hand reveal lol )

/snip

This is the most interesting link I have seen in this thread for some time. Thanks!

I assume this is AMD's response to Xeon Phi (where is Phi btw?) and will be relegated to servers for HPC work. If I am not mistaken the patent was filed prior to the SeaMicro acquisition so I wonder if this is still a planned product or if Freedom Fabric has usurped it.

No its an extension to current x16 PCIe slots... like (PCIe)_____________ + _____ (sideband), on the same row, so the PCIE slot gets an "extension". The remarkable is that you can use in that slot either a PCIe adapter(no sideband needed) or a HTX (hypertransport) adapter, being the great advantage of this last one, if not bandwidth(could be) at least latency and Cache Coherency. Yes it could be a response to Xeon Phy, FireGL can be HTX cache coherent and be a truly "co-processor"...

But also the patent foresees the same in "integrated" designs, that is, ALL "processores" (CPU & APU) from AMD are already an Hypertransport switch in the XBAR (integrated northbridge)... they would pass also being a PCI Xbar of sorts... meaning truly integrated.

One possible great advancement is all Radeon adapters coming with that "SLOT" in mind, on intel and legacy they could function PCIe, on advanced AMD platforms they would function HTX. Also then the X2 adapters will be gone, all could be truly MCM now, but perhaps this MCMs then one only HTX (no PCIe, makes sense).

I understand that. My point was that with the HT link added to PCIe bandwidth it would allow for a co-processor that performs more like a multi-socket system than a GPU\Phi assisted one. I assume that the HT link allows the co-processor direct access to main memory and possibly direct access from the host to the GPU\Phi memory.

The real question is whether this will be relegated to the Server\Workstation markets or if it will show up in enthusiast platforms.

 

[hcl123]

@ gamerk316 ; juanrga

I made some e "hasty" calculations alright, but why bother and insist, that "SR" tests seems has fake as they come... and even IF it isn't, must be a preliminar ES, the setup isn't indicated, what DRAM(edit) what platform... so in any case is not representative at all

I think the AMDFX blog (from where that chinese site picked the "speculation") did a "speculation" preview, that is why the top FX 9650 ends on top of 32% like a clock LOL... only they forgot to mention that they counted with clock improvement also... quite not to take seriously.

 

[hcl123]

I understand that. My point was that with the HT link added to PCIe bandwidth it would allow for a co-processor that performs more like a multi-socket system than a GPU\Phi assisted one. I assume that the HT link allows the co-processor direct access to main memory and possibly direct access from the host to the GPU\Phi memory.

The real question is whether this will be relegated to the Server\Workstation markets or if it will show up in enthusiast platforms.

the patent is not given yet, so its not for soon.

Yet when will be, according to past AMD positioning, it will be only a "professional" feature, perhaps the new RD1000 chipset will have it on C32 and G34 sockets platforms, doubt AM3+ will have ( its EOL, but if it will i salute AMD lol).

So perhaps the "enthusiasts" can transition to C32 or G34... don't know, but it will be a good bet, both are LGA.

 

[8350rocks]

Just to set the record straight, FX 8350 overclocked to 8.176 GHz on all 8 cores...a new world record (For all cores active):

http://hwbot.org/newsflash/1840_amd_fx_8350_overclocked_to_8.176_ghz_with_8_cores_enabled

 

[gamerk316]

@ gamerk316 ; juanrga

I made some e "hasty" calculations alright, but why bother and insist, that "SR" tests seems has fake as they come... and even IF it isn't, must be a preliminar ES, the setup isn't indicated, what DRAM(edit) what platform... so in any case is not representative at all

I think the AMDFX blog (from where that chinese site picked the "speculation") did a "speculation" preview, that is why the top FX 9650 ends on top of 32% like a clock LOL... only they forgot to mention that they counted with clock improvement also... quite not to take seriously.

Working on the assumption the DRAM/Platform is the same for both platforms, the results are still valid. So the question becomes "are the numbers real"? If so, then it looks like 15% improvements strictly on IPC, plus performance due to higher clocks.

 

[hcl123]

Perhaps AMD will change their mind at the last second and get a 28nm or PD-SOI FD-SOI process . Realistically, it looks like AMD has just gotta adapt. However, I am still confident Kaveri A10-XXXXK will beat the 8150's broken record. I need those $1000 so I can build a Steamroller FX 8XXX/9XXX rig to replace my beloved i7-920 C0@3.8 main rig and hand-me-down but not really mine QX9770@3.7 gaming rig. Talk about power consumption . The FX9650 name seems odd, considering it was taken by the Phenom I (AKA: the 8150 of 2007).

You wish lol

28nm PD-SOI was never announced... only a 28 SHP ( 32 SHP is PD-SOI) appeared in some GloFo charts, but GloFo never talked what that is.

What is revealed is that the first versions of kaveri will be bulk, most probably low power low clock "mobile" parts(edit), so bulk is not a bad bet, and also will not have GDDR5 on socket.

Since Richland is already > 20% average on graphics/games above Trinity
http://translate.googleusercontent.com/translate_c?act=url&depth=3&hl=en&ie=UTF8&prev=_t&rurl=translate.google.com&sl=auto&tl=en&u=http://pclab.pl/art53197.html&usg=ALkJrhiuaznyeIjQEFEYT7b1jGPX9mYuiQ

Kaveri being GCN can be even a little more than that to Richland (>20%)... yet the same 384sp, for low power DDR3 2133. More than enough to beat the Iris and no need for 512sp ( iris is 160) or GDDR5.(edit)

After that is anybody guess, if Kaveri Desktop and FX will not be PD-SOI, then AMD is wasting a good competitive advantage, but this part is not on bet lol...

EDIT

Working on the assumption the DRAM/Platform is the same for both platforms, the results are still valid. So the question becomes "are the numbers real"? If so, then it looks like 15% improvements strictly on IPC, plus performance due to higher clocks.

Exactly has above... that is 2 questions 2 assumptions, not 1... if Steamroller will only be in "bulk" variants a very good assumption is that it will never have higher clocks than the actual Vishera/Trinity.

So that "speculation" along with revealing only 15% IPC gain, not 32%, is also adventing PD-SOI...

(i'm hoping 28nm PD-SOI, but i will not bet on it lol .. better said, if PD-SOI (or FD-SOI for that matter, which Glofo already announced ) then all bets are off...)

 

[The Q6660 Inside]

Perhaps AMD will change their mind at the last second and get a 28nm or PD-SOI FD-SOI process . Realistically, it looks like AMD has just gotta adapt. However, I am still confident Kaveri A10-XXXXK will beat the 8150's broken record. I need those $1000 so I can build a Steamroller FX 8XXX/9XXX rig to replace my beloved i7-920 C0@3.8 main rig and hand-me-down but not really mine QX9770@3.7 gaming rig. Talk about power consumption . The FX9650 name seems odd, considering it was taken by the Phenom I (AKA: the 8150 of 2007).

You wish lol

28nm PD-SOI was never announced... only a 28 SHP ( 32 SHP is PD-SOI) appeared in some GloFo charts, but GloFo never talked what that is.

What is revealed is that the first versions of kaveri will be bulk, most probably low power low clock "mobile" parts(edit), so bulk is not a bad bet, and also will not have GDDR5 on socket.

Since Richland is already > 20% average on graphics/games above Trinity
http://translate.googleusercontent.com/translate_c?act=url&depth=3&hl=en&ie=UTF8&prev=_t&rurl=translate.google.com&sl=auto&tl=en&u=http://pclab.pl/art53197.html&usg=ALkJrhiuaznyeIjQEFEYT7b1jGPX9mYuiQ

Kaveri being GCN can be even a little more than that to Richland (>20%)... yet the same 384sp, for low power DDR3 2133. More than enough to beat the Iris and no need for 512sp ( iris is 160) or GDDR5.(edit)

After that is anybody guess, if Kaveri Desktop and FX will not be PD-SOI, then AMD is wasting a good competitive advantage, but this part is not on bet lol...

EDIT

Working on the assumption the DRAM/Platform is the same for both platforms, the results are still valid. So the question becomes "are the numbers real"? If so, then it looks like 15% improvements strictly on IPC, plus performance due to higher clocks.

Exactly has above... that is 2 questions 2 assumptions, not 1... if Steamroller will only be in "bulk" variants a very good assumption is that it will never have higher clocks than the actual Vishera/Trinity.

So that "speculation" along with revealing only 15% IPC gain, not 32%, is also adventing PD-SOI...

(i'm hoping 28nm PD-SOI, but i will not bet on it lol .. better said, if PD-SOI then all bets are off...)

Aww.......my i7-920+4870 rig is crying to become a router! Steamroller may cap out @7.5-8GHz on extreme cooling if the bulk process is to be adopted, but I am sure stock clocks will be identical. They always have their Asetek CLC to fall back on the 8-Core SKUs. I wish they could go 28nm SOI, but sacrifices have to be made for the company's own good.

 

[hcl123]

You wanna bet on that ? ... i can use $1000 lol

 

[The Q6660 Inside]

You wanna bet on that ? ... i can use $1000 lol

Alright, the glorious PayPal master race begins.

 

[8350rocks]

You guys do know that GloFo announced HKMG scalability down to 28nm right? That's the process for the 28nm SHP. It's FD-SOI as well...which means that would be a good performance bump over bulk or even PD-SOI.

http://www.eetimes.com/design/eda-design/4406513/GloFo-commits-to-FDSOI

That's interesting that initial production of SoC manufacturing will begin in Q4 2013 and full scale ramps up Q1 2014...what other timing does that sound like to you?

 

[The Q6660 Inside]

You guys do know that GloFo announced HKMG scalability down to 28nm right? That's the process for the 28nm SHP. It's FD-SOI as well...which means that would be a good performance bump over bulk or even PD-SOI.

http://www.eetimes.com/design/eda-design/4406513/GloFo-commits-to-FDSOI

That's interesting that initial production of SoC manufacturing will begin in Q4 2013 and full scale ramps up Q1 2014...what other timing does that sound like to you?

Steamroller! AMD might be able use 28nm FD-SOI instead of "bulk".

 

[hcl123]

Alright, the glorious PayPal master race begins.

No LN2 or similar is valid... >7.5ghz on water, "stable" ? .. else define extreme.. but enough to withstand a full set of bencjmarks

(c'mon SNB already caped very close to that on 32nm, even is was only good to take a couple of shots)

 

[8350rocks]

You guys do know that GloFo announced HKMG scalability down to 28nm right? That's the process for the 28nm SHP. It's FD-SOI as well...which means that would be a good performance bump over bulk or even PD-SOI.

http://www.eetimes.com/design/eda-design/4406513/GloFo-commits-to-FDSOI

That's interesting that initial production of SoC manufacturing will begin in Q4 2013 and full scale ramps up Q1 2014...what other timing does that sound like to you?

Steamroller! AMD might be able use 28nm FD-SOI instead of "bulk".

 

[The Q6660 Inside]

Alright, the glorious PayPal master race begins.

No LN2 or similar is valid... >7.5ghz on water, "stable" ? .. else define extreme.. but enough to withstand a full set of benchmarks

(c'mon SNB already caped very close to that on 32nm, even is was only good to take a couple of shots)

[strike]FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUU[/strike] Extreme cooling (IMO) means anything from poor man's LN2 (Alcohol+Dry Ice) up to Liquid Helium :3.

 

[hcl123]

Steamroller! AMD might be able use 28nm FD-SOI instead of "bulk".

The first iterations was revealed by an AMD "official" in an interview to be "bulk", which i assume will be "mobile" parts... fits nicely.

The rest or was not revealed or is bulk also. Case SR has a longer pipeline and much more speed daemon, and the IPC improvement can yet be a good bit superior to 30%... which is kind of counter-intuitive, greater IPC (fatter cores) and a good speed improvement...then bulk might not be that bad if clocks are very close to Richland.

That EEtimes article is old news, FD-SOI at GF see here

http://www.advancedsubstratenews.com/2013/04/gfs-two-flavors-of-fd-soi-kengeri-explains-exclusive-asn-qa/

http://www.advancedsubstratenews.com/2013/05/more-good-fd-soi-news-from-date-conference-st-leti-mentor-cmp/

The real interesting part for high performance is the "maximum" process variant at GF, but really tuned for it only at the 14nm. (2016/17) ... OTOH 20nm SHP might had transitioned for FD-SOI "maximum" first iteration (2014/15), 28nm should be all low power(bulk and FD-SOI).

I would bet more on 28nm PD-SOI than FD-SOI.
http://semiaccurate.com/forums/showpost.php?p=133926&postcount=9

http://translate.google.com/translate?langpair=auto|en&u=http%3A%2F%2Fwww.planet3dnow.de%2Fcgi-bin%2Fnewspub%2Fviewnews.cgi%3Fid%3D1360178291
Original link http://www.planet3dnow.de/cgi-bin/newspub/viewnews.cgi?id=1360178291
(good article with the most recent graph, comparing old graph and FD-SOI news)

IMAGE
http://www.planet3dnow.de/photoplog/file.php?n=22812&w=o

 

[hcl123]

(uuuutooo lol) Extreme cooling (IMO) means anything from poor man's LN2 (Alcohol+Dry Ice) up to Liquid Helium :3.

OK, but since you love so much that crap bulk, which is SOOOO SUPERIOR... than lame AMD is over 8Ghz if with bulk than it should be above (>) 9Ghz (otherwise how could be improvement ?.. that was always the point)... and at least stable enough to run a couple of becnhmarks ?

 

[noob2222]

Supposed benchmarks of Steamroller FX chips

http://northwood.blog60.fc2.com/blog-entry-6834.html

AMD Steamroller FX 9650 - 4.5 Ghz / 4.8 Ghz Turbo Benchmarks! lol

If that is truth, than its not 32% better than the Fx8350, by the test results...

Lets cont only POV-RAY, the Cinebench doesn't count, after so many past posts i think nobody is going to be fouled.

POV-RAY
FX-8350: 1505
FX-9650: 1985

That is a 32% gain.

The Fx8350, in high performance, that is, in multithreading which is the only truly high performance metric( there isn't any other, ILP is dead) is already ~10% above the i7 3770 by that test... so it could be similar performance to the 4770.

Last Openbenchmarks show that a FX-8350 at stock is so faster or a bit more than the 4770k in multhreading. Moreover the FX OC better than Haswell.

If the FX9450 is similar clocks to the 8350, but then its only 10% above Piledriver by that test on POV-ray results !... so from where comes that 32% ?

It looks like the 9650 is 4.5GHz base clock, but can't find info on the 9450's clock. For reference: The 8350's base clock is 4GHz. So at least some of the performance is due to clock, not IPC. Without know the 9450's clock, really hard to make a comparision.

As far at the 8350/9650 comparison, we can solve the IPC difference based on the info we have:

For POV Ray:

Performance = IPC * Clockspeed * Number of Cores [Number of Cores is the same for both CPU's, so this cancels out]

8350: 1505 = IPC * 4
8350: IPC = 376.25

9650: 1985 = IPC * 4.5
9650: IPC = 441.1

Now to solve the performance difference:

IPC Difference = 100 - 376.25 / 441.1 * 100
IPC Difference = ~15%

For CineBench (single):

8350:

1.11 = IPC * 4
IPC = .2775

9650:

1.47 = IPC * 4.5
IPC = .3267

IPC Difference = 100 - .2775 / .3267 * 100
IPC Difference ~15%

So far, looks like 15% IPC gains. One more to go:

Cinebench (Multi)

8350:

6.93 = IPC * 4
IPC = 1.7325

9650:

9.15 = IPC * 4.5
IPC = 2.03

IPC Difference = 100 - 1.7325 / 2.03 * 100
IPC Difference ~ 15%

So it looks like Steamroller gains 15% in IPC over Piledriver, based on three benchmarks. Not bad for a generational leap.

And remember I'm solving for the DIFFERENCE in IPC, not the actual IPC. I've also simplified the equation and made a few assumptions (100% core loading). Quite possible the IPC number is slightly understated if SR isn't being pushed to 100%, or overstated if Turbo is being used excessively. But I'd say 15% is probably fairly accurate for CPU based benchmarks.

The reason these numbers are all the same ... they are all calculated theoretical numbers. I thought they looked familiar, parroted information from may 22. http://amdfx.blogspot.com/2013/05/amd-steamroller-fx-9650-45-ghz-48-ghz.html

nothing was tested, this is what theoretical performance calculated. RL IPC gains are NEVER LINEAR.