The Official Kaveri Review Thread (A10-7850K, etc)

[Erenhardt]

Why the throttling?

 

[rainy]

Especially in discrete GPU gaming, PD really outclassed BD (~13,5% faster). Same cant be said about SR vs PD (a mere ~5.8% faster despite lots of architectural improvements).

Quite obviously you have forgotten a "small" detail: Vishera base clock is 400 MHz (circa 11 percent) higher than Bulldozer.
At 3.2 GHz FX-8350 could look rather poor vs FX-8150 (except power consumption).

Btw, Steamroller advantage over Piledriver is close to 10 percent rather than 5.

 

[Torn Mind]

I said double of % Piledriver gained against Bulldozer.
My english is so bad?

If Bulldozer takes 10 seconds to complete a task at 3.3 GHz and with one core, then Steamroller should take only 5 seconds at the same clockspeed and core count if IPC doubled. Likewise, a 2 hour encode job on a Bulldozer chip should take only one hour on a Steamroller chip with the same clockspeed and core count. That is what happens if IPC is doubled and nothing else changes.

IPC is one of the components of IPS(instructions per second). The others are clockspeed and core count. Yes there are even more variables that make smaller impacts, such as cache, but that makes things complicated.

IPS= IPCxClockspeed for one "core". Double IPC, and you double IPS if Clockspeed does not change.

Principal level of performance*(1+x) can represent increase in CPU performance.

The principal level of performance can be represented with any arbitrary number, but the number "1" leaves the least amount of confusion behind.

So, if Bulldozer's level of performance is 1, then Piledriver might be 1.15 faster.
1*(1+ .15)=1.15

Steamroller might be 1.25 times faster than Piledriver.
1.15(1+.25)=1.4375

In order to turn these into percentages, it is as simple as multiplying by 100%.

Seems like gauging CPU performance gains is similar to calculating interest payments. If I had $100, and they payed me 100% interest, I would get $200 at the end of the period.

 

[OCGuy]

http://semiaccurate.com/2012/11/05/amds-bulldozer-core-compared-with-piledriver/

http://semiaccurate.com/2014/01/23/kaveri-versus-richland-performance-per-clock-comparison/

I don't think that shows what you think it shows.

 

[Sweepr]

Quite obviously you have forgotten a "small" detail: Vishera base clock is 400 MHz (circa 11 percent) higher than Bulldozer.
At 3.2 GHz FX-8350 could look rather poor vs FX-8150 (except power consumption).

Btw, Steamroller advantage over Piledriver is close to 10 percent rather than 5.

I didnt. It was a clock per clock comparison but quite obviously you have forgotten this small detail. Even Haswell did better compared to Ivy Bridge than Steamroller vs Piledriver @ equal clocks in Hardware.fr's test suite. Amazing isnt it?

 

[Abwx]

Quite obviously you have forgotten a "small" detail: Vishera base clock is 400 MHz (circa 11 percent) higher than Bulldozer.

Not sure who forgotten a detail since 13.5% is the difference
at equal frequency for games...

http://www.hardware.fr/articles/880-6/bulldozer-vs-piledriver-4-ghz.html

 

[OCGuy]

RR just said record sales of A8 and A10 APUs in Q4.
Edit: also,

How were your AMD sales personally? You do sell exclusively AMD APUs on Greek tech boards, right?

http://translate.google.com/transla...dbox=0&usg=ALkJrhgIzPCTIn2cWhVhH8xfL8yFshENuA

 

[Abwx]

I didnt. It was a clock per clock comparison but quite obviously you have forgotten this small detail. Even Haswell did better compared to Ivy Bridge than Steamroller vs Piledriver @ equal clocks in Hardware.fr's test suite. Amazing isnt it?

They had thoses results because they got 7% lower IPC
in winzip and 7zip and that they averaged only 13 softs
comprising thoses two ones , hence their 7.5% better IPC
only but this number would be accurate if 15% of the eventual
softs would yield thoses 7% losses , wich is greatly overestimated
that s why they took care to mention that the real number is rather
10% , so not worse than IB/HW , moreover if we take account
of HW new ISAs like AVX2 that help it artificialy boost its IPC
when compaired to its predecessor.

 

[AtenRa]

How were your AMD sales personally? You do sell exclusively AMD APUs on Greek tech boards, right?

Business doing great thank you

 

[monstercameron]

http://news.softpedia.com/news/AMD-...-on-PCs-With-Dual-Rank-RAM-DIMMs-419886.shtml

dual rank dimms, does this make sense?

 

[zir_blazer]

http://news.softpedia.com/news/AMD-...-on-PCs-With-Dual-Rank-RAM-DIMMs-419886.shtml

dual rank dimms, does this make sense?

Kaveri can maximize bandwidth on such modules (even though there are some with even four ranks) and, thanks to DDR3-2400 support, can push system performance 15.2% higher than on DDR3-1866 memory.

Looks like apples to oranges comparison to me, are they talking about Frequency or ranks? Besides, all Quad Rank modules are Buffered/Registered. I doubt these work with Kaveri.

 

[Abwx]

Looks like apples to oranges comparison to me, are they talking about Frequency or ranks? Besides, all Quad Rank modules are Buffered/Registered. I doubt these work with Kaveri.

And why shouldnt they work.?.Because of your doubts.?.

Hardware.fr reviewer is posting numbers in HFR forum using 4xSR...

http://forum.hardware.fr/hfr/Hardwa...ure-steamroller-sujet_924911_129.htm#t9026699

 

[zir_blazer]

And why shouldnt they work.?.Because of your doubts.?.

How many Desktop platforms have you seen working with Registered/Buffered RAM? On LGA 1150, you simply can't. And on AMD side, I never heared about anyone trying on AM3 or FM1/2, but chances are they don't worth because as far that I know, the Motherboard traces on platforms intended to work with Buffered DIMMs are wired differently. So yes, I have doubts. Got an answer for those?

 

[el etro]

I don't think that shows what you think it shows.

6% from BD to PD and 11% from PD to SR.

 

[Abwx]

How many Desktop platforms have you seen working with Registered/Buffered RAM? On LGA 1150, you simply can't. And on AMD side, I never heared about anyone trying on AM3 or FM1/2, but chances are they don't worth because as far that I know, the Motherboard traces on platforms intended to work with Buffered DIMMs are wired differently. So yes, I have doubts. Got an answer for those?

I dont think that they are wired differently but rather than the traces
must be designed such that they work with dram circuits that are
indeed compatible voltage and protocol wise but whose dynamic
electric caracteristics are different due to buffering that will eventualy
change their high frequency impedance , hence care is mandated in their
design to have working transmission lines with no reflexions.

Edit : Memory controler must be designed accordingly
and we know that Kaveri will be used in low cost servers,
as such it can use ECC memory , i m quite sure that they
didnt neglect buffered ram given the server purposes
and ECC capability.

 

[Dresdenboy]

Kaveri competes CPU wise with Celerons and Pentiums at half the price. Those CPUs are also much more power efficient. So while the GPU part is an upgrade. Its a downgrade in about every other metric out there, including cost. And with the price increase for stagnant performance. AMD made sure OEMs will deslect the chips in even higher amount than they did with previous chips.

Is the Pentium still faster in apps which might benefit from AVX or AES since it doesn't support these extensions? No HTr, no GPU (which of course also needs power in Kaveri). So the Pentium capable to run 2 threads, no AVX, no AES, no GPU still needs a Kaveri-esque TDP? These chips might be harvested.

 

[USER8000]

How were your AMD sales personally? You do sell exclusively AMD APUs on Greek tech boards, right?

http://translate.google.com/transla...dbox=0&usg=ALkJrhgIzPCTIn2cWhVhH8xfL8yFshENuA

Why are you stalking someone on the internet?

 

[Dresdenboy]

[...] moreover if we take account
of HW new ISAs like AVX2 that help it artificialy boost its IPC
when compaired to its predecessor.

AVX2 actually reduces IPC. That's done by reducing execution overhead per single op (where multiple happen do be done per SIMD instruction) and/or with wider units.

 

[AtenRa]

AVX2 actually reduces IPC. That's done by reducing execution overhead per single op (where multiple happen do be done per SIMD instruction) and/or with wider units.

Well yea, we actually mean CPI (Cycles Per Instruction) or single thread performance when we talk about IPC here

 

[sefsefsefsef]

6% from BD to PD and 11% from PD to SR.

Just because this guy's English is a little rough, I want to clarify the point he's been making all along (but he couldn't articulate very well).

He *doesn't* claim that IPC doubled from BD to SR (although based on the way he said it, it kind of looks like this is what he's claiming). He claimed that the *improvement* from BD to PD was doubled by SR. So if PD was X% better than BD, then SR would be 2X% better than BD.

This is a perfectly reasonable claim, IMO.

 

[frozentundra123456]

Unfortunately, ipc does not tell the whole story. From BD to PD there was an IPC *and* a clockspeed increase for an overall nice final increase in cpu performance. For kaveri, the increase in ipc was pretty much negated by the lower clockspeed.

 

[AtenRa]

Unfortunately, ipc does not tell the whole story. From BD to PD there was an IPC *and* a clockspeed increase for an overall nice final increase in cpu performance. For kaveri, the increase in ipc was pretty much negated by the lower clockspeed.

That was only for the A10-7850K which is unlocked by the way. The A8-7600 has both a higher IPC and higher Clock Speed than equivalent 45W TDP Richland A10-6700T.

 

[Shivansps]

I do not belive the A8-7600 to be working in a 45W TDP window.

 

[el etro]

Just because this guy's English is a little rough, I want to clarify the point he's been making all along (but he couldn't articulate very well).

He *doesn't* claim that IPC doubled from BD to SR (although based on the way he said it, it kind of looks like this is what he's claiming). He claimed that the *improvement* from BD to PD was doubled by SR. So if PD was X% better than BD, then SR would be 2X% better than BD.

This is a perfectly reasonable claim, IMO.

My portuguese(Brazil official language) sucks equally, i would not explain it better than i explain to you in English....

I said that PD have IPC improved X% over BD, and SR have IPC improved (X*2)% over PD. That's semiaccurate and other not so known bloggers/forumers claim.


--------------------

Unfortunately, ipc does not tell the whole story. From BD to PD there was an IPC *and* a clockspeed increase for an overall nice final increase in cpu performance. For kaveri, the increase in ipc was pretty much negated by the lower clockspeed.

This. If you are talking about Kaveri on 95w TDP range.

If Kaveri will be a failure or not we will know when the reviews of first Kaveri powered notebook come to market.

 

[AtenRa]

I do not belive the A8-7600 to be working in a 45W TDP window.

Why not ??