Kabini IPC, what generation are we talking about?

Chicken76

Senior member
Jun 10, 2013
279
69
101
I've read a bunch of reviews of the Athlon 5350, and while they do offer a lot of graphs, power consumption numbers, etc they do not give you the feel of the platform. What are we talking about here, in terms of IPC? Pentium IV, Athlon 64, Pentium M, what?
What I'm interested in is, if one were to downclock a Pentium IV or some other well known and extensively benchmarked CPU to ~2 GHz, would it approximate the single-threaded performance of the 5350? (of course using only instructions available in both chips)
 

Ancalagon44

Diamond Member
Feb 17, 2010
3,274
202
106
As far as I can remember, and I could be wrong...

The CPU before Kabini, Bobcat, beat K8 (the original Athlon 64) in IPC. It came close to K10 in some benchmarks. So, a 2GHz Bobcat would be faster than an AMD K8 Athlon 64.

Kabini is something like 20% faster at the same clocks as Bobcat, so I would think that a 2GHz is likely to be the equivalent of an AMD K10 Phenom I (but not Phenom II). Maybe even a little faster.

A Pentium IV downclocked to 2GHz will be substantially slower than a 5350.
 

sm625

Diamond Member
May 6, 2011
8,172
137
106
Athlon 5350 scores 433 on sunspider 1.0.2. This is slightly less than half of what a 2GHz haswell scores. It is 1/4 of what a $60 20th anniversary pentium with a modest overclock on the stock cooler can accomplish. In other words, it is absolutely abysmal.
 

Chicken76

Senior member
Jun 10, 2013
279
69
101
Thanks for reminding me of the Anandtech CPU bench. Completely forgot about it. Glad to see the 5350 added to the list of bechmarked CPUs.

It looks like it's close to the cut-down first-gen Core2 Duos (2MB L2, 65nm), am I right?
 

Chicken76

Senior member
Jun 10, 2013
279
69
101
Athlon 5350 scores 433 on sunspider 1.0.2. This is slightly less than half of what a 2GHz haswell scores. It is 1/4 of what a $60 20th anniversary pentium with a modest overclock on the stock cooler can accomplish. In other words, it is absolutely abysmal.

So it has about half the IPC of Ivy/Haswell. Is this a correct approximation guys?
 

coercitiv

Diamond Member
Jan 24, 2014
7,399
17,553
136
It looks like it's close to the cut-down first-gen Core2 Duos (2MB L2, 65nm), am I right?
IPC wise a Kabini core is a bit lower than a C2D core. These graphs from hardware.fr should come in handy. (keep in mind the clock speeds too)

It takes an overclock to around 2.75Ghz for Kabini to reach the same Cinebench 11.5 score as Q6600 @ 2.4Ghz.
 

SPBHM

Diamond Member
Sep 12, 2012
5,067
422
126
if you look on the anand bench the 2.05GHz Kabini scores around the same as the 2.1GHz K8 on cinebench r10 single core test.
 

TrulyUncouth

Senior member
Jul 16, 2013
213
0
76
Core 2 Celerons yes.



Less than half. Its roghly 41.5% of IB IPC in Cinebench. And some 37% or so of Haswell.

You guys are clearly much better at this than me- what about a size comparison for these parts and/or cost comparison. I would be very interested to see how much larger of a core it takes for IB and Haswell to get 2-3x the performance. Also, what is the comparison in IPC with bay trail?

Thanks
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
146
106
You guys are clearly much better at this than me- what about a size comparison for these parts and/or cost comparison. I would be very interested to see how much larger of a core it takes for IB and Haswell to get 2-3x the performance. Also, what is the comparison in IPC with bay trail?

Thanks

J1900 and 5350 is roughly the same singlethreaded Cinebench performance.

J1900 runs at 2.42Ghz and the 5350 at 2.05Ghz. So IPC for Kabini is 18% over Baytrail.

EDIT: Just saw the J1900 in Anandtech bench too, different numbers there for whatever reason. That would put Kabini 32% in front of Baytrail.
http://anandtech.com/bench/product/1223?vs=1227
 
Last edited:

TrulyUncouth

Senior member
Jul 16, 2013
213
0
76
J1900 and 5350 is roughly the same singlethreaded Cinebench performance.

J1900 runs at 2.42Ghz and the 5350 at 2.05Ghz. So IPC for Kabini is 18% over Baytrail.

EDIT: Just saw the J1900 in Anandtech bench too, different numbers there for whatever reason. That would put Kabini 32% in front of Baytrail.
http://anandtech.com/bench/product/1223?vs=1227

Thanks for that. I did a bit of digging myself and couldn't find anything on IB for some reason- however I found a thread here that has a comparison of a ton of cores. If I am not mistaken, it looks like Haswell is almost 5 times the size of Jaguar to get that 3x improvement in ipc.
2013_core_sizes_768.jpg


I was wondering if there is a better benchmark for CPU comparison on the baytrail/kabini I know AMD fans always complain about cinebench- is there some bench both sides agree on?

No luck on a cost comparison yet, but I would be interested to see it.
 

jhu

Lifer
Oct 10, 1999
11,918
9
81
Based on my testing, Piledriver is about Phenom II IPC. Bobcat is about double Pentium 4, but still about 30% lower IPC than Piledriver. Kabini is based on Steamroller and supposedly has higher IPC than Piledriver, so perhaps it's closer to Core 2 IPC.

What I find interesting is how well Core 2 does without an integrated memory controller.
 

TrulyUncouth

Senior member
Jul 16, 2013
213
0
76
Based on my testing, Piledriver is about Phenom II IPC. Bobcat is about double Pentium 4, but still about 30% lower IPC than Piledriver. Kabini is based on Steamroller and supposedly has higher IPC than Piledriver, so perhaps it's closer to Core 2 IPC.

What I find interesting is how well Core 2 does without an integrated memory controller.

I am pretty sure KAbini is based on the cat cores and not steamroller.
 

Roland00Address

Platinum Member
Dec 17, 2008
2,196
260
126
You guys are clearly much better at this than me- what about a size comparison for these parts and/or cost comparison. I would be very interested to see how much larger of a core it takes for IB and Haswell to get 2-3x the performance. Also, what is the comparison in IPC with bay trail?

Thanks

2013_core_sizes_768.jpg


Intel Cores 130nm to 45nm penryn cpu cores here

Intel 45 nm nehalem (1st gen Core i series)

Now that picture is kinda misleading for it does not tell you what the final chip size is. Don't forget you need everything else on the chip such as the cache, the gpu, the memory busses, the input and output, etc. These parts of the chip are much more than just the cpu core, for example a haswell quad core is about 1/3 cpu cores not counting cache. Knowing the cpu core size is only useful info if the maker of the cpu will have dozens of configurations based off several dies. Intel does not even do this currently having only about 5 die configurations, they will add a couple more die configurations next year when they release the haswell ex for servers but we are still talking about limited dies choices made.


AMD never announced the die size of kabini but it is about 114mm^2 on 28nm, this is a quad core with 128 stream processors Source Hans
Intel Haswell has many different configurations. A ulv dual core with the biggest graphics (40 EU) is 181 mm^2 on 22 nm(if I recall this also includes the southbridge). A higher tdp quad core with medium graphics (20 eu) is 177 mm^2 (this does not count the southbridge). A desktop or non ultrabook laptop dual core with medium graphics (20 eu) is 130 mm^2

Just for comparison the final die size of amd 40nm tmsc bobcat was about 74mm^2, intel 45nm pineview atom (the first atom) was about 87mm^2, the new 22nm atom die size has not been revealed, atoms a7 in the iphone 5s and ipad air is 102mm^2 on samsungs 28nm. Tegra 2 die size was 49 mm^2 on 40nm TMSC, anand estimated tegra 3 die size to be in 80s on TMSC 40nm, and tegra 4 die size to be once again in the 80s on TMSC 28nm
 
Last edited:

TrulyUncouth

Senior member
Jul 16, 2013
213
0
76
Wow, what an info-filled response. Still working through all your links but just wanted to respond to the main idea.

I am the first to admit I don't know a ton about CPU design and whatnot- however if we are comparing IPC of the core does the uncore really matter? The only thing I can see being relevant is the core and the cache- which if I'm not mistaken would put the Haswell even larger in size in comparison to the cat cores. Assuming things like USB controllers and the various IO are similar between small and large cores then including everything would give a natural advantage to large cores because the size of the actual core matters less.
 

Roland00Address

Platinum Member
Dec 17, 2008
2,196
260
126
Some things do not need to get bigger, they do not need to scale, when you increase the cpu core size, add more cpu cores, and/or improve the gpu.

Some things do need to get bigger when you enlarge the cpu cores, add more cpu cores, and/or improve the gpu. You need more cache to feed the cpu or your cpu will just idle, you need more memory bandwidth and thus more memory controllers to feed the gpu, etc.

That said usually with intel the other stuff does not need to increase much "other stuff" on the "mainstream" platform when they increase the number of cpu cores and gpu cores.

On other platforms such as servers the "other stuff" has to get bigger a lot so you make sure your cpu is the most efficient as possible (whatever the most efficient is depends on what you are prioritizing.), You are willing to have a big die for you are getting paid several thousands or several hundreds of dollars per chip.
 

TrulyUncouth

Senior member
Jul 16, 2013
213
0
76
Some things do not need to get bigger, they do not need to scale, when you increase the cpu core size, add more cpu cores, and/or improve the gpu.

Some things do need to get bigger when you enlarge the cpu cores, add more cpu cores, and/or improve the gpu. You need more cache to feed the cpu or your cpu will just idle, you need more memory bandwidth and thus more memory controllers to feed the gpu, etc.

That said usually with intel the other stuff does not need to increase much "other stuff" on the "mainstream" platform when they increase the number of cpu cores and gpu cores.

On other platforms such as servers the "other stuff" has to get bigger a lot so you make sure your cpu is the most efficient as possible (whatever the most efficient is depends on what you are prioritizing.), You are willing to have a big die for you are getting paid several thousands or several hundreds of dollars per chip.

I appreciate the info, truly, because I find this stuff interesting- but I feel you didn't really respond to what I was trying to say. If we are looking at IPC between 2 cores and want to know how much die space each CPU needed to achieve that IPC wouldn't it make the most sense to look at cores and maybe cache added to get that answer? Everything outside of that seems like it just muddies the waters

For example if there were a haswell version with hundreds of PCIe lanes it would be silly to add in the die space taken up by all the circuitry needed to control the lanes. Am I off-base here or does that make sense?

Edit: Also, I can't find the size of Baytrail atom cores anywhere- anyone know the size?
 
Last edited:

Roland00Address

Platinum Member
Dec 17, 2008
2,196
260
126
I appreciate the info, truly, because I find this stuff interesting- but I feel you didn't really respond to what I was trying to say. If we are looking at IPC between 2 cores and want to know how much die space each CPU needed to achieve that IPC wouldn't it make the most sense to look at cores and maybe cache added to get that answer? Everything outside of that seems like it just muddies the waters

For example if there were a haswell version with hundreds of PCIe lanes it would be silly to add in the die space taken up by all the circuitry needed to control the lanes. Am I off-base here or does that make sense?

Edit: Also, I can't find the size of Baytrail atom cores anywhere- anyone know the size?

Let me try answering it again.

1) How much cache, memory controllers, and such is still an important issue with IPC. Sometimes you gain substantial increases (10%+) by adding more cache, sometimes adding more cache means barely any improvement. Thus I think things that do increase IPC should be taken in consideration comparing cpu designs.

(part 2 I am just picking random numbers, to illustrate a point)

2) I think knowing the cpu core size and the final die size does matter. Lets say you have a 70mm^2 cpu, 15 mm^2 is the cpu cores and the cache, thus 21% of the chip is cpu while 79% is everything else. If you can get 30% more ipc and keep the same frequency and roughly the same voltage by increasing the cpu size to 25mm^2 (an increase of 66% of the cpu cores and the cache) and everything else stays the same size we now have a 80mm^2 chip. The cpu core is now 31% of the chip, and overall the die size for the entire chip went up 14%. If we can sell the chip for at least 15% more money (assuming good yields) then this design decision probably made sense. Sometimes you can sell a chip for how ever much higher the ipc goes up, sometimes you can actually get more than 30% for this is the best silicon out there and thus people will pay top dollar. It really depends on the market. Sometimes increase the cpu size barely changes the entire die size since the cpu is mostly everything else, this is a big thing in the cellphone market

3) We buy cpus as finished products. The market is not commondized enough that you can tell the manufacturer how many cpu or gpus lanes we want. Instead we just choose how much we want to spend, or whatever our priorities are and just pick the best cpu for our needs. (I am going to pick Apple for a moment, but any company can substitute here) Thus caring about things like apple A7 has a 100 mm^2 or 500 mm^2 die size really does not matter to consumers. It may matter to apple, but it is "not my circus, not my monkeys" the only way it affects us is how much profit apple makes at the end of the day whether they make more profit by shaving costs off the cpu or shaving costs of the screen cost does not really matter.

In sum I do not care about how efficient a cpu design overall is compared to another cpu design. I do not care in this type of intellectual exercise for I am more of a brass tacks type of guy. How is this going to affect me. Steve Jobs did not care how Rod Holt design the apple II power supply, he just wanted it to be quite, small, and cheap enough to make so he can put it in an enclosed area. For all Jobs cared little mice could be running on little treadmills just as long as it got the job done. Job recognized how hard his demands of Holt were but he did not need to know the details. (I may be telling the apple II story a little off, I am not that into it, but I am just trying to demonstrate a point).

I am amazed and full of wonder with what all the talented engineers have done with limited resources. I understand how hard it was to do what they did. I do not need to know "how" they did it, I find the "why and what" they did to be far more important.

-----

Just found Intel's Die Size for the original bay trail they did not publically announced it but they gave all the info you need in the publically disclosed datasheat pdf on page 351 c1 and c2. It is about 102 mm^2 since the dimensions are 9.723 by 10.477 which equals 101.87

Now I do not know what the individual core sizes are minus the cache. This will be important now that we will be start seeing custom baytrail with other designs such as powerVR graphics. If some one can get the accurate lithography of the cpu you can have someone like Hans calculate how much of the soc is dedicated to the cpu cores now that we got the die size down to the thousandth of the millimeter.
 

Lepton87

Platinum Member
Jul 28, 2009
2,544
9
81
I haven't read the whole thread yet, I will later, but it's generally agreed that Kabini offers 90% of IPC of K8.(which is coincidentally the official number from AMD) K8, on the other hand has 25% lower IPC then C2D, which has 96% higher IPC then P4. Kaveri/Puma offered 20% more IPC then Kabini so it clearly surpassed K8 and came close to the original Phenom. Saying that Kabini has the same IPC as P4 is just wrong.
 

VirtualLarry

No Lifer
Aug 25, 2001
56,587
10,225
126
I haven't read the whole thread yet, I will later, but it's generally agreed that Kabini offers 90% of IPC of K8.(which is coincidentally the official number from AMD) K8, on the other hand has 25% lower IPC then C2D, which has 96% higher IPC then P4. Kaveri/Puma offered 20% more IPC then Kabini so it clearly surpassed K8 and came close to the original Phenom. Saying that Kabini has the same IPC as P4 is just wrong.

So it would be pointless to upgrade my friend with an Athlon64 X2 4800+ S939 rig with 4x512MB DDR, to a 5350 AM1 rig then? Seeing as how his IPC and clockspeed would be going down?