The Kaveri Pre-Launch Thread (A10-7800 and A10-6800k @3,5 Ghz)

[Abwx]

A singlethread regression in Geekbench 3, similar to Cinebench but not as bad. This is a big disappointment. Gaming benchmarks not much faster than Richland, I didn't expect this. I thought Kaveri could bring bigger gains due to the shrink and new GPU architecture.

In Geekbench 3 it has 10% better single core score than a 4.1GHz
Richland , advantage increase to about 16% on the multi core test,
if thoses tests are genuine of course.

The APUs at the bottom are two Richland , on top of the list
are two 3.5G Kaveri s ES using a single memory channel as well as a
Richland, the rest are LGA Kabini :

http://browser.primatelabs.com/geekbench3/search?q=AMD+Eng+Sample

Edit : This review has Richland scoring much better
than in Geekbench database..

 

[FlanK3r]

Dont forget, Richland has +400 MHz higher clocks! Thats not few MHz...clock for clock...There is nice improvement for CPU. Remember, clock for clock simillar architecture as SB vs IB vs HS or Phenom 65nm or Phenom II or Zambezi core vs Piledriver vs Steamroller B. I can tell you, the STeamroller is in this case the best in improvement.

clock for clock:

SB vs IB (3% for IB)
IB vs HSWL (5% for HSWL)
Phenom vs Phenom II (around 5%)
Bulldozer vs Vishera (yes, there is good impact around 10%, but its Zambezi cache system was full of bugs)

 

[Sweepr]

Dont forget, Richland has +400 MHz higher clocks! Thats not few MHz...clock for clock...There is nice improvement for CPU. Remember, clock for clock simillar architecture as SB vs IB vs HS or Phenom 65nm or Phenom II or Zambezi core vs Piledriver vs Steamroller B. I can tell you, the STeamroller is in this case the best in improvement.

clock for clock:

SB vs IB (3% for IB)
IB vs HSWL (5% for HSWL)
Phenom vs Phenom II (around 5%)
Bulldozer vs Vishera (yes, there is good impact around 10%, but its Zambezi cache system was full of bugs)

IB to Haswell saw a ~8-11% IPC bump, (10-11% in apps and 8-9% in games according to Hardware.fr). Anand averaged at ~9%. Double digits on top of an already very fast Ivy Bridge might be a little more difficult than beating sub Phenom II levels of IPC (@ MT tasks) by the same (or bigger) margin.

 

[SPBHM]

SB, IB and HW had an IPC increase + higher/same clock, it's not the case with Kaveri vs Richland,

 

[FlanK3r]

depends at test of course . In this test is Vishera clear winner after. Average clock to clock + 8% in CPU performance and + 13% in games against Bulldozer.

Kaveri si nice APU, only it need more gains in MHz for better impact against boosted Richland

 

[Sweepr]

SB, IB and HW had an IPC increase + higher/same clock, it's not the case with Kaveri vs Richland,

At the end of the day what really matters is per core performance, and even with ''up to'' 20% better perf/clock than Richland (yet to be proven) if they are running at 10% lower clocks they might not beat IB -> Haswell performance increase (around ~10% for similarly priced products, sometimes more due to slightly increased clock), thus not really closing the gap on x86 performance with Intel.

 

[Abwx]

Dont forget, Richland has +400 MHz higher clocks! Thats not few MHz...clock for clock...There is nice improvement for CPU. Remember, clock for clock simillar architecture as SB vs IB vs HS or Phenom 65nm or Phenom II or Zambezi core vs Piledriver vs Steamroller B. I can tell you, the STeamroller is in this case the best in improvement.

clock for clock:

SB vs IB (3% for IB)
IB vs HSWL (5% for HSWL)
Phenom vs Phenom II (around 5%)
Bulldozer vs Vishera (yes, there is good impact around 10%, but its Zambezi cache system was full of bugs)

Clock for clock it wouldnt matter much if single thread FP performance doesnt increase , the 10% or so in MThread seen on Cinebench 1.5 would be good enough overall but on the X86/Integer department expectations,
at least by me, are about 15% on average between ST and MT , so far
the few integer benchs yield variable results, for FritzChess it s 12.4%,
Sandra integer Arithmetic is 34% but it s purely synthetic as is Geekbench
wich show odd numbers , Wprime get 43% better perf but i suspect
that it s a special case due to Steamroller s new hardware radix 8
divider and roots extractor, well , as usual we have to wait for extensive
reviews from the usual suspects sites...


Edit : Nice CPU herd in your sig...

 

[mikk]

http://browser.primatelabs.com/geekbench3/search?q=AMD+Eng+Sample

Edit : This review has Richland scoring much better
than in Geekbench database..

They might have tested the 64bit version of Geekbench which is faster.

 

[Homeles]

dont must be...different process. Kaveri is at 99% 28nm SHP process. There is better optimalization for more tranzistors.

...what?

 

[Abwx]

They might have tested the 64bit version of Geekbench which is faster.

The Geekbench list say 32bit but once you click
on a system you ll see W7 64bit as OS , quite a mess...

 

[FlanK3r]

Clock for clock it wouldnt matter much if single thread FP performance doesnt increase , the 10% or so in MThread seen on Cinebench 1.5 would be good enough overall but on the X86/Integer department expectations,
at least by me, are about 15% on average between ST and MT , so far
the few integer benchs yield variable results, for FritzChess it s 12.4%,
Sandra integer Arithmetic is 34% but it s purely synthetic as is Geekbench
wich show odd numbers , Wprime get 43% better perf but i suspect
that it s a special case due to Steamroller s new hardware radix 8
divider and roots extractor, well , as usual we have to wait for extensive
reviews from the usual suspects sites...


Edit : Nice CPU herd in your sig...

thx. I know the true. I have one 7850K

 

[mikk]

The Geekbench list say 32bit but once you click
on a system you ll see W7 64bit as OS , quite a mess...

When it says Geekbench 3.x.xTryout for Windows x86 (32-bit) it is running on 32bit Geekbench which is the case for all A10-6800k systems. I didn't found a x64 result in the list.

 

[JDG1980]

So based on what we've seen so far with the various leaks, it looks like integer IPC has increased a good bit (15%-30%), while floating-point IPC has stayed about the same, which means that single-threaded floating-point applications may do a bit worse due to the lower clock speeds. Meanwhile, virtually all multi-threaded applications should do better, due to the reduction/elimination of the CMT penalty.

Presumably as the 28nm process matures, clock speeds for Kaveri will catch up to Richland, making it superior across the board. We'll probably have to wait for Excavator for better FPU performance, though.

Once again the WSA causes trouble for AMD. If it weren't for that albatross of a contract, AMD could have had these APUs fabricated on TSMC's 28nm process (which is fully mature by now) and gotten substantially better performance.

 

[Abwx]

When it says Geekbench 3.x.xTryout for Windows x86 (32-bit) it is running on 32bit Geekbench which is the case for all A10-6800k systems. I didn't found a x64 result in the list.

First system on the linked list :

http://browser.primatelabs.com/geekbench3/223737

Operating System Microsoft Windows 7 Ultimate N (64-bit)

Same for the second and so on...

 

[SPBHM]

First system on the linked list :

http://browser.primatelabs.com/geekbench3/223737



Same for the second and so on...

that's 32bit

"Geekbench 3.1.2 Tryout for Windows x86 (32-bit)"

geekbench free is 32bit, the OS supporting x64 is irrelevant,
and the 64bit version gives better results normally.

 

[inf64]

OS can be 64bit but geekbench 3 tryout is always executed in 32bit code.

 

[Homeles]

Presumably as the 28nm process matures, clock speeds for Kaveri will catch up to Richland, making it superior across the board. We'll probably have to wait for Excavator for better FPU performance, though.

Once again the WSA causes trouble for AMD. If it weren't for that albatross of a contract, AMD could have had these APUs fabricated on TSMC's 28nm process (which is fully mature by now) and gotten substantially better performance.

A refreshed Kaveri achieving clock speed parity with Richland is unlikely. Richland's simply on a better process.

 

[mrmt]

Presumably as the 28nm process matures, clock speeds for Kaveri will catch up to Richland, making it superior across the board. We'll probably have to wait for Excavator for better FPU performance, though.

How do you know that clocks went down due to GLF's process, and not because of design issues, like the bigger decoder or a shorter pipeline?

 

[NTMBK]

How do you know that clocks went down due to GLF's process, and not because of design issues, like the bigger decoder or a shorter pipeline?

 

[Homeles]

How do you know that clocks went down due to GLF's process, and not because of design issues, like the bigger decoder or a shorter pipeline?

It's pretty easy to come to that conclusion with this chart:
http://www.realworldtech.com/includes/images/articles/iedm10-10.png?53d41e

32nm PDSOI clearly performs better than TSMC 28nm. There's no data for CP 28nm, but we can assume it performs equal to or worse than TSMC's process because the Common Platform process uses gate first.

Oh, and I guess AMD flat out admitted it too. Lol.

 

[Sweepr]

A refreshed Kaveri achieving clock speed parity with Richland is unlikely. Richland's simply on a better process.

Also Richland arrived more than a year after Trinity. 1 year from now AMD should be launching EX-based Carrizo. There might be a slight clock increase for late 2014 Kaveri parts but probably not something like Richland this time.

 

[inf64]

I hope Excavator is done on 20nm SOI process which can not only help with lowering the power draw(95W->65W) but also with achieving needed clock speed boost. If they can get ~20% IPC with Excavator( much more in AVX stuff though) AND they can get north of 4Ghz while staying in 65W envelope, Carizo will be much more potent product than Kaveri. Kaveri was unlucky one to land on 28nm bulk node and according to the slide above AMD was working hard to offset this by core changes.

 

[krumme]

How do you know that clocks went down due to GLF's process, and not because of design issues, like the bigger decoder or a shorter pipeline?

Where does its say it has a bigger decoder and not just an seperate one for each core unlike earlier where the decoder was shared ? And why should a bigger decoder be slower anyway?

 

[Abwx]

OS can be 64bit but geekbench 3 tryout is always executed in 32bit code.

Ok, In that case the multicore score is underhelming given
that MT penalty was reduced , it look like frequency matters
more than throughput for theses uarch when Geekenched...

 

[rainy]

Also Richland arrived more than a year after Trinity.

You're wrong: desktop Trinity was released on October 2012 and Richland on June 2013 and mobile parts on May 2012/March 2013 - of course not all models in the same time.

https://en.wikipedia.org/wiki/List_...ors#Virgo_-_.22Trinity.22_.282012.2C_32_nm.29
https://en.wikipedia.org/wiki/List_of_AMD_APU_microprocessors#.22Richland.22_.282013.2C_32_nm.29
https://en.wikipedia.org/wiki/List_...ors#Comal_-_.22Trinity.22_.282012.2C_32_nm.29
https://en.wikipedia.org/wiki/List_of_AMD_APU_microprocessors#.22Richland.22_.282013.2C_32_nm.29_2