Athlon X4 845 ExcavatorIPC benchmarks

[monstercameron]

Speaking as a 6800K owner, I find it sad that AMD still hasn't released an APU that can beat its single thread performance.

Keep in mind the 6800K does have a 600MHz (turbo/non-turbo) frequency advantage. Carrizo would need to have ~15% better IPC to just keep up.



If I can get one at current pre-order pricing I might get one to play around with. Could be a fun little chip.

I might pick one up too, they say never go full carrizo but I beg to differ

 

[frozentundra123456]

I am at least 90%, if not 99%, sure that English is not csbin's first language. If he's from somewhere in Europe, like most of the people who post around here who are still learning English, English is also not his second language. Forming complete, intelligible sentences is the very last step of learning a language.

Give him time, and I'm willing to bet that he'll become every bit as fluent in our native tongue as the other two people who've posted in this thread who don't speak English natively have become. They are both more fluent in English today than the average Anandtech forum member, including the ones whose native language happens to be English.

Irregardless, it is annoying to me as well the way he just posts some link or benchmark or gaming result without any kind of comment or explanation about it. I dont think that is a language problem, so much as just a posting style/philosophy. Nothing wrong with it I suppose, but it just seems strange to me.

 

[Arachnotronic]

I am at least 90%, if not 99%, sure that English is not csbin's first language. If he's from somewhere in Europe, like most of the people who post around here who are still learning English, English is also not his second language. Forming complete, intelligible sentences is the very last step of learning a language.

Give him time, and I'm willing to bet that he'll become every bit as fluent in our native tongue as the other two people who've posted in this thread who don't speak English natively have become. They are both more fluent in English today than the average Anandtech forum member, including the ones whose native language happens to be English.

There is no shame in being a non native English speaker and that was absolutely not the intent of my comment. Where I take issue with csbin's posts in general is that (s)he "hits and runs" with data/charts that are exclusively pro-AMD and anti-Intel/anti-NVIDIA.

I know full well that csbin is capable of producing full sentences, but I'm astonished that csbin doesn't even give fellow posters here the courtsy of full sentences in replies explaining what (s)he is trying to get across. Instead it's stuff like, "LOL IVB" with a chart cut and pasted.

 

[ElFenix]

i don't know what you guys are even trying to prove anymore.

i just request that we have 1 XV thread for whatever this is, so that people who need actual help with something don't have to wade through garbage like this thread to find actual information.

 

[frozentundra123456]

I know what you mean, but that is just the way people are. I used to have a co-worker who would argue with the boss and then get into an argument about whether they were arguing or not. It was actually hilarious and painful at the same time. That said, I have always found the forums very helpful if I asked for help with a specific hardware or software problem. I do admit that I miss the technical input of some of the old posters like IDC, guskline, and a few others.

 

[Dresdenboy]

[CB ST performance chart removed]

40% better than this would still be slower than Intel's architectures from recent years.

The construction core FPUs have longer instruction latencies, and branch misprediction penalty is high. Both things aren't that good for Cinebench and other raytracers.

And keep in mind, that this 40% number very likely is an average over very different workloads. The average return of trades might be 30%, which doesn't mean, that one had only winning trades or only 30% ones ...

 

[looncraz]

Hmm, I guess the CPU-Z benchmark isn't that great then.

Anybody around here planning to buy an XV to do more tests?

I've already bought one, it has been shipped, should be here after the weekend. I am already preparing my benchmark suite.

I will be testing the following systems for direct comparison, all at 3GHz.

Phenom II X4 955, 3GHz, 4GB DDR2-1066, Windows 7
Core 2 Duo (? model), 3GHz, Windows 7, 4GB DDR2-1066
Kaveri A8-7600, 3GHz, 8GB DDR3-1600 & 8GB DDR3-2133
Carrizo x4 845, 3Ghz, same RAM/system as Kaveri

I will be testing a large suite of benchmarks, and some games, everything fixed at 3GHz and all power management/turbo disabled.

I will write up a short and dirty article with the results, methods, some picture maybe, and a few charts. Probably won't be available until next Friday, assuming the 845 gets here by Wednesday. I have emergency car work to do that takes priority sometime next week when the parts arrive, so that costs me a day, and I haven't booted the Core 2 Duo system in a year, so I don't even know if it's gonna boot

 

[jhu]

Because there are other 3D content creation applications that BD architecture is performing a lot better vs Intel.

Like what? Because it's definitely not Blender or Povray.

BTW, don't be fooled by Povray's test scene since it doesn't test commonly used features like global illumination or focal blur. Add those two and Intel pulls quite a bit ahead in terms of IPC.

 

[IllogicalGlory]

If Zen delivers a "greater than 40%" IPC improvement as indicated by Lisa Su, they'll be basically spot on with Haswell performance. Clocking an i7-4790K to 3.8 and disabling turbo boost gives me a score of 1725 in CPU-Z. 44% faster than the bench in the OP.

 

[looncraz]

The 40% is moot in this case when it comes to estimate how it will perform in CB 11.5 ST, a better approach is to assume that a Zen core will have as much or more throughput (in MT = 1C/2T) than a EXV module since they use the same FPU or so.

From here one can estimate the expected IPC improvement in ST, so unless Zen SMT scale at 100% and has the same ST perf as EXV, wich is unlikely, the ST IPC in CB 11.5 will be increased by significantly more than 40%..

Precisely. Each Zen core has exclusive access to its FPU, and you don't get allocation conflicts between the cores. That's worth an easy 15% for FPU performance, before adding in the extra pipeline, the improved FMAC arrangement, the faster caches, and other factors.

Determining floating point performance increases, though, from what information we have is effectively impossible. Except that legacy SIMD should, in theory, perform much better on Zen even than Haswell when using mmxadd, sseiadd, or fdiv. SSE logic should also be better.

 

[richierich1212]

Please keep us posted with test results, thanks loon

 

[DrMrLordX]

Why are people still looking at the single-threaded results from programs like CineBench when run on XV CPUs and then trying to extrapolate Zen's performance from that? "40% of that will be at Sandy Bridge levels" good grief.

Try using the MT results instead?

 

[Sweepr]

Why are people still looking at the single-threaded results from programs like CineBench when run on XV CPUs and then trying to extrapolate Zen's performance from that? "40% of that will be at Sandy Bridge levels" good grief.

Try using the MT results instead?

As you wish:

The unlocked Core i5s operate at similar clockspeeds to this Athlon X4 so we can compare:

- Sandy Bridge is ~40.8% faster
- Ivy Bridge is ~64.2% faster
- Haswell is ~73.8% faster
- Skylake is almost 2x as fast (~95.6%)

 

[jpiniero]

Yeah, which is why people are saying Sandy Bridge IPC for Zen. Something like an 8 core Zen with a 3-3.2 Ghz base really would not be that bad for games given that games are starting to use 6 threads pretty effectively. Zen's appeal beyond gamers is really limited though.

 

[SPBHM]

Yeah, which is why people are saying Sandy Bridge IPC for Zen. Something like an 8 core Zen with a 3-3.2 Ghz base really would not be that bad for games given that games are starting to use 6 threads pretty effectively. Zen's appeal beyond gamers is really limited though.

I think if the IPC is close to sandy bridge the clock better be closer to 4GHz than 3GHz and 6+ threads to make it really interesting for gaming (with competitive pricing)

 

[USER8000]

The Cinebench 11.5 single threaded score is very close clock to clock to a Phenom II X4
980BE.

 

[JDG1980]

40% better than this would still be slower than Intel's architectures from recent years.

This assumes that Zen shows 40% improvement across the board on everything. If it's an average, then some applications would show less improvements than that, and some would show more. I'd expect AMD to be focusing most on shoring up their weak spots, so there's a good chance we'll see considerably less than 40% improvement in the integer-focused workloads where AMD already does well (e.g. 7-Zip, x264) and considerably more than 40% improvement in the floating point applications where AMD has the most trouble currently.

The_Stilt mentioned that Cinebench 11.5 and 15 use a variety of FPU instructions, including MMX and the various iterations of SSE. It's worth pointing out in this context that Steamroller actually represented a potential regression from Piledriver in MMX performance, as it now shares hardware with the 128-bit FMAC pipes. I suspect the same is true of Excavator.

 

[JDG1980]

The Cinebench 11.5 single threaded score is very close clock to clock to a Phenom II X4
980BE.

And this shows just how much Bulldozer hurt AMD. Phenom II X4 is part of the Deneb architecture, which was released in February 2009. Seven years later, and AMD still hasn't pushed IPC beyond that level. This is a damning indictment of the construction cores, and demonstrates why Zen can't come soon enough.

 

[Face2Face]

I've already bought one, it has been shipped, should be here after the weekend. I am already preparing my benchmark suite.

I will be testing the following systems for direct comparison, all at 3GHz.

Phenom II X4 955, 3GHz, 4GB DDR2-1066, Windows 7
Core 2 Duo (? model), 3GHz, Windows 7, 4GB DDR2-1066
Kaveri A8-7600, 3GHz, 8GB DDR3-1600 & 8GB DDR3-2133
Carrizo x4 845, 3Ghz, same RAM/system as Kaveri

I will be testing a large suite of benchmarks, and some games, everything fixed at 3GHz and all power management/turbo disabled.

I will write up a short and dirty article with the results, methods, some picture maybe, and a few charts. Probably won't be available until next Friday, assuming the 845 gets here by Wednesday. I have emergency car work to do that takes priority sometime next week when the parts arrive, so that costs me a day, and I haven't booted the Core 2 Duo system in a year, so I don't even know if it's gonna boot

Looking forward to it :thumbsup:

 

[deasd]

More here:
http://forums.anandtech.com/showpost.php?p=37645739&postcount=193

Very thanks, dresdenboy. Looks like I misunderstand, the x64 version of CPUZ is executing SIMD instead of pure integer, and using VC++. This is still a bench which is fair enough between different vendors and architecture.
I wonder if x32 version which use both x86 & x87 result is any differ than x64. (x32 ver is a kind of bench similar to superpi though......)

 

[frozentundra123456]

This assumes that Zen shows 40% improvement across the board on everything. If it's an average, then some applications would show less improvements than that, and some would show more. I'd expect AMD to be focusing most on shoring up their weak spots, so there's a good chance we'll see considerably less than 40% improvement in the integer-focused workloads where AMD already does well (e.g. 7-Zip, x264) and considerably more than 40% improvement in the floating point applications where AMD has the most trouble currently.

The_Stilt mentioned that Cinebench 11.5 and 15 use a variety of FPU instructions, including MMX and the various iterations of SSE. It's worth pointing out in this context that Steamroller actually represented a potential regression from Piledriver in MMX performance, as it now shares hardware with the 128-bit FMAC pipes. I suspect the same is true of Excavator.

Or lets face it 40% could be best case for a cherry picked benchmark.

 

[DrMrLordX]

As you wish:

The unlocked Core i5s

How about the i3s instead?

 

[TheELF]

How about the i3s instead?

What about them? Nobody is going to consider buying an I3 for rendering work.

 

[Dresdenboy]

As you wish:

The unlocked Core i5s operate at similar clockspeeds to this Athlon X4 so we can compare:

- Sandy Bridge is ~40.8% faster
- Ivy Bridge is ~64.2% faster
- Haswell is ~73.8% faster
- Skylake is almost 2x as fast (~95.6%)

AMD CON quad cores have 2 FPUs.

2T/2M vs. 2T/2C or 4T/2M vs. 4T/2C should give a better impression of the FPU performance.

 

[USER8000]

And this shows just how much Bulldozer hurt AMD. Phenom II X4 is part of the Deneb architecture, which was released in February 2009. Seven years later, and AMD still hasn't pushed IPC beyond that level. This is a damning indictment of the construction cores, and demonstrates why Zen can't come soon enough.

True,but the Phenom II launched at 3GHZ and it took until May 2011 for the 3.7GHZ Phenom II X4 980BE to be launched. So that is just under 5 years ago.December 2011 the last Llano SKU the A8 3870K which was launched on 32NM which was only clocked to 3GHZ and has similar IPC,and AMD could never really get an SKU clock high enough.

Since then we had only one 28NM bulk process which took ages to match 32NM SOI in clockspeeds.

So in that sense I consider AMD more in a holding pattern for the last 4 years or so,and Global Foundries has not helped.

Remember, Kaveri is only getting the clockspeeds which it was kind of expected to launch with two years ago.

To put it in context Intel has had three full node shrinks to use in the time AMD has one and a bit,and Intel has had no issue ramping clockspeeds.