Intel Sandy Bridge Performance

[aegisofrime]

Hi, I have here what is probably an unusual request.

I'm contemplating upgrading my current Phenom II 955 PC to the Intel Core i7-2600K. My PC is mainly used for deinterlacing DVD videos with QTGMC, and encoding that with x264.

Thus, I would really like to know what kind of performance improvement I can expect with Sandy Bridge. I know that plenty of websites utilize x264 as a benchmark, however I would like to know how QTGMC scales with Sandy Bridge.

My usual settings are as follows:
QTGMC: QTGMC(Preset="Slower")
x264: --crf 23 --preset slower --tune film --aq-mode 2

If anyone of you guys happen to have a Sandy Bridge CPU and could let me know what FPS your encoding process runs at, it would really help me make my purchasing decision. Thanks! :)

[burfadel]

the phenom II 955 isn't too bad... all things considered. You may not get the advantage others have, and comparing fps is useless as they may be using different filters, versions, overclocking, and of course the source material makes a massive difference! Also there's the cpu utilisation...

The next gen AMD's are coming out later in the year, significantly different architecture and also with the new instructions the Sandy Bridge has, plus some others. Not sure if x264 could advantage from FMA4 or not...

Might be worth waiting...

[aegisofrime]

Thanks for the reply burfadel. I understand those concerns, which is why I posted the settings that my workflow uses. It will not completely replicate my workflow, but hopefully it should come close enough.

With regards to Bulldozer, I do not have confidence that it will come close to Sandy Bridge, for the following reasons:

1) At this point, more than ever AMD needs consumer confidence. Sandy Bridge is available and Bulldozer will only launch in Q2 at the earliest.

2) Engineering silicon has already been produced, yet there are no leaked benchmarks.

3) If Bulldozer is competitive to Sandy Bridge, AMD would release benchmarks to show that its competitive to Sandy Bridge. At this point there are people like me who is contemplating an upgrade, and positive Bulldozer benchmarks will go a long way to convincing people like us to wait for Bulldozer.

4) AMD CEO Dirk Meyer, who was regarded by the industry as being a pretty good CEO, recently resigned. Is it because Bulldozer is not up to expectations? It's possible as AMD is counting on it to fight Intel.

Some of my reasoning may come across as absurd, yet we must remember this is the age of viral marketing... Look at how Anandtech got hold of Sandy Bridge benchmarks 6 months before it's release...

[mariush]

If the engineer samples exist, doesn't necessarily mean benchmarks are possible. As you say the architecture may be different, several new x86 instructions exist which are not used by the applications (the benchmark software), maybe the new cpu is maybe 12 core or 16 cores at low frequency (like the existing 8 core 1.9 Ghz ones or 12 cores x 2.1 Ghz for example) and the benchmark software / applications are optimized for 2-4 high frequency cores...

It would be a somewhat unfair benchmark for an engineer sample (which may have issue like power leaks, overheating, throttling cores and so on... if any of the above holds true.

[CruNcher]

I don't think it will be that much ahead Speed wise if @ all if you look into AMDs release cycle Bulldozer is in research some time now since the Release of the A64 actually the CPU im currently on K8 and since then they moved slowly and Intel made very huge steps they came into financial problems and where restructured Intel even drove farer away of them in that time.
So you could expect that they might be a little faster at the same Watt usage as Sandy Bridge but not the difference that Intel reached with the Core architecture change back then and now Sandy Bridge drove again 1 Generation in front and improved the Power/Speed ratio if you compare current AMD CPUs the newest 6 Core currently you see a power draw of 190 Watt AMD @ load on the whole system and 130 Watt for Intel (and with that lower consumption being faster most of the time), that is a huge difference in Efficiency Bulldozer most probably be able to close that gap but more Powerfull is still questionable even with a new Architecture, but AMD doesn't really would need that they just need a nicer Price/Watt/Speed balanced Platform again and their Position would be save :P
But myself i decided to not longer wait even im not happy how it turned out to be with the GPU on the Sandy not being able to be used as easily in combination with a discrete GPU to bundle its power (Multiple Decoding, Encoding,Preprocess Balancing) still hurdles here but some solutions exist already for that which look not bad :)
But waiting for most probably reaching the same Power Consumption with Bulldozer then Sandy @ mostl roughly the same Speed or expected a little faster no not for me and im surely wouldn't cry if Bulldozer comes out, because Sandy would be also future proof for the next years and currently it would be a 3,5x Compute improvement for me (changing from 2 to 4 cores with better Per core performance even) and using Multiple GPUs & DSPs @ the same time for different tasks would improve that even significantly more depending on the balancing :)

[aegisofrime]

Er... It looks like I might need a Sandy Bridge CPU in my brain to decipher your reply Cruncher. But thanks for replying anyway, I will have a go at trying to comprehend it tonight :p

[popper]

If the engineer samples exist, doesn't necessarily mean benchmarks are possible.

As you say the architecture may be different, several new x86 instructions exist which are not used by the applications (the benchmark software), maybe the new cpu is maybe 12 core or 16 cores at low frequency (like the existing 8 core 1.9 Ghz ones or 12 cores x 2.1 Ghz for example) and the benchmark software / applications are optimized for 2-4 high frequency cores...

It would be a somewhat unfair benchmark for an engineer sample (which may have issue like power leaks, overheating, throttling cores and so on... if any of the above holds true.

what give's you that idea ?

if you compile x264 from git
a simple

make checkasm;./checkasm

checkasm –bench

should work fine and be far more useful than some/many benchmarks at least for the instructions that are currently in the code, true no AVX yet i dont think but that can change quick enough.

http://www.techpowerup.com/138328/Bulldozer-50-Faster-than-Core-i7-and-Phenom-II.html
according to that AMD reportedly gave out performance figures ,an 8-core Bulldozer can get 50% higher performance than existing processors such as the Core i7 950 (4 cores, 8 threads)...

that's saying a real 8-core AMD CPU is 50% faster than a real 4 core Intel and 4 virtual cores that can do upto 50% of the real cores work , Or if you prefer

a brand New AMD Bulldozer 8-core can do the same work as that (Old,non SB ?) Intel Core i7 950 clock for clock... doesn't seem so good when you think of it like that.

but a Real life checkasm –bench would probably show IF AMD can finally run assembly as fast as Intel at the same clock speed ,or as as it seems so far on all AMD CPU's slower as usual.

[mariush]

Yes, it should work fine. But the asm code using the new instructions on the AMD processor doesn't write itself, someone has to write it. So you can compile it on the new processor, but the code will treat it like a regular processor and as the code automatically checks what instructions are supported by the CPU and makes adjustments in some cases the performance may not be as big as the processor can do.

For example - and I'm just making this up - a certain SSE3 function works fast on Intel but on AMD the SSE2 version of same function works faster than the SSE3 one so it's selected automatically during runtime but now on the new AMD CPU the SSE3 version is faster than the SSE2 version.

[burfadel]

I'm assuming by an 8 core Bulldozer, they are actually really referring to a physical 4 core CPU that has 8 integer cores and 4 FPU cores. I think this terminology will trick a lot of people, because in reality, it is really only a quad processor that happens to have 2 integer units per core :). The extra cost of the extra integer unit per core was apparently very minimal, so cost wise the Bulldozer should be comparable to the Sandy-Bridge, just (supposedly) a lot faster! In addition to both having AVX instructions, the Bulldozer has the additional benefit of having what was once known as SSE5, but is now FMA4, CVT16 etc, although I'm not sure whether they will be beneficial for encoding.

[popper]

Yes, it should work fine. But the asm code using the new instructions on the AMD processor doesn't write itself, someone has to write it. So you can compile it on the new processor, but the code will treat it like a regular processor and as the code automatically checks what instructions are supported by the CPU and makes adjustments in some cases the performance may not be as big as the processor can do.

For example - and I'm just making this up - a certain SSE3 function works fast on Intel but on AMD the SSE2 version of same function works faster than the SSE3 one so it's selected automatically during runtime but now on the new AMD CPU the SSE3 version is faster than the SSE2 version.

well AMD Could cure that very simply , (Intel couldn't be bothered apparently though,and it had to be pointed out to them directly) they could get the core x264 dev's (and their new assembly Minions, or is it just a single minion) access to these working samples ASAP and So have better benchmarks on official release.

[burfadel]

I forgot about the XOP instruction set earlier... Intel has added AVX with Sandy Bridge, and will add FMA3 with Ivy Bridge. Bulldozer will have added the AVX XOP, FMA4, CVT16 instruction sets.

Wonder which will be the main instruction set out of FMA3, or FMA4, since they are apparently incomaptible. Intel was FMA4 then went to FMA3, and AMD FMA3 and went to FMA4... lol. FMA4 sounds more powerful, but FMA3 is Intel, who no doubt went FMA3 for Ivy Bridge as apparently its easier for hardware implementation (cheaper?)

I'm curious as to what potential benefits (especially) FMA4, XOP, maybe CVT16, and AVX bring to x264?

[LoRd_MuldeR]

I'm curious as to what potential benefits (especially) FMA4, XOP, maybe CVT16, and AVX bring to x264?

AVX is useless for x264, because it's FP-only. Don't know about the others...

See also:
http://doom10.org/index.php?topic=514.msg3536#msg3536

[Sharktooth]

maybe FMA3/4. but it will depend on the execution speed.

[LoRd_MuldeR]

maybe FMA4. but it will depend on the execution speed.

Any info whether FMA3 (Intel) and FMA4 (AMD) will support Integer instructions or Floating-point instructions or both? And how does XOP fit in? :confused:

[Sharktooth]

i'll ask and hope to come back with an answer.
btw some more specific info on the new instructions set supported by bulldozer: http://www.atomicmpc.com.au/News/144746,inside-amds-latest-cpu-instruction-sets.aspx
however, AFAIK FMA3/4 are FP.

[burfadel]

Well, according to Wikipedia (if you can trust this info):

The XOP instructions include:

* Integer vector multiply-accumulate instructions
* Integer vector horizontal addition
* Integer vector compare
* Integer vector shift and rotate instructions
* Vector byte permutation
* Vector conditional move instructions
* Floating point fraction extraction
http://en.wikipedia.org/wiki/XOP_instruction_set

The FMA3 {Intel} and FMA4 {AMD} instruction sets have almost identical functionality but are not mutually compatible. Both contain fused multiply-and-add instructions for floating point scalar and SIMD operations.http://en.wikipedia.org/wiki/FMA4_instruction_set

So, completely as a guess as I really don't have ani idea :D it sounds like XOP should be quite beneficial to the motion estimation.

[Sharktooth]

FMA could be too, but still it depends on speed...

[aegisofrime]

Alright, you guys have convinced me to wait a few more months for Bulldozer. Can't hurt I guess. I just hope that RAM prices don't shoot up in these few months; They are at rock bottom prices now, at least in my country.

[Sharktooth]

you can get the RAM now and wait for the rest of the components... ;)

edit: oh, Bulldozer is an AM3+ CPU and has DDR3 support.

[Sharktooth]

According to various websites, a 4 modules Bulldozer seems to be 50% faster than core i7 950 and a 8 modules Bulldozer is twice as fast as a core i7 980x in 3d rendering benchmarks.
If that is true (and considering i7 2600k is about 20% faster than i7 950 on 3d rendering benchmarks) Bulldozer is faster than Sandy Bridge.

[ronnylov]

And now when all the chipsets for Sandy Bridge have been recalled by intel because of some problem with the SATA ports there is no alternative to wait anyway. When Sandy Bridge is back on the market it may not be long further wait for the Bulldozer. But who knows, maybe Bulldozer have some hidden faults too. You always take a risk when buying something completely new on the market.

[aegisofrime]

Yes, Intel has kindly helped me to make the decision to wait for Bulldozer. Although truth to be told, the fault isn't really that bad. All you need to do is to hook up your HDDs onto the first two SATA ports and the DVDs on the rest. The ports aren't supposed to die that quickly.

From what I have heard, the FMA instructions on Bulldozer is likely to help, x264 wise. So too is the 2 to 1 integer to floating point unit ratio, since supposedly x264 is mainly integer operations. At least that's what I gathered from Dark Shikari's comment on AVX. "fp only, so a pile of <insert dung related word here>"

In any case, April's the time when we will slowly get fixed SB boards, and April's around the time Bulldozer is supposed to be released. What a coincidence.

[Sharktooth]

Yes, Intel has kindly helped me to make the decision to wait for Bulldozer. Although truth to be told, the fault isn't really that bad. All you need to do is to hook up your HDDs onto the first two SATA ports and the DVDs on the rest. The ports aren't supposed to die that quickly.
not possible on laptops... ;)

From what I have heard, the FMA instructions on Bulldozer is likely to help, x264 wise. So too is the 2 to 1 integer to floating point unit ratio, since supposedly x264 is mainly integer operations. At least that's what I gathered from Dark Shikari's comment on AVX. "fp only, so a pile of <insert dung related word here>"

In any case, April's the time when we will slowly get fixed SB boards, and April's around the time Bulldozer is supposed to be released. What a coincidence.
what x264 needs is int SIMDs. so int pipelines will be usefull but no soooo much. on the other side the SIMD units are pretty fast and have separated schedulers. so a bulldozer module can execute int separately from fp and SIMD intructions. however FP and SIMD can be executed in a "smart" way as described here: http://blogs.amd.com/work/2010/10/25/the-new-flex-fp/
thus giving a major flexibility and speed boost.
about FMA, they works on floating points. it's still to be determined how much usefull FMA will be since FP instructions are usually slower than int ones. XOP are interesting though.

[CruNcher]

Ok im on Sandy Bridge and i have to say it wows so far in terms of power consumption and speed though i cant use anything currently of the Intel HD 2000 Graphics based Multimedia improvements as the whole Media SDK 2.0 ISvs are working with is based on DXVA2 and so only works either on Vista/7 nothing of the interesting stuff works on XP on both sides.
So keep in mind if you want to use the full Potential of Sandy Bridge you will be per se forced to move/adapt to Vista/7 at least no way arround that (if you want full hardware multimedia functions) Intel seems to not further want to support Windows XP and DXVA1 anymore, only MPEG2 is supported on XP currently :)
Also see in the other thread with a discrete card in the system you need to have 2 monitors connected to make any of the Media SDK 2 functions (Applications written for it) to work properly currently.

[Sharktooth]

the same (or more) SB limitations affect linux too...

[wlee15]

In a Bulldozer module each core has it's own load/store system, is that beneficial to x264?

[Sharktooth]

each core has its own scheduler too. in theory it's good for parallelism but the performance are unknown until there will be some engineering samples for testing.

[Sharktooth]

intel will start shipping the fixed chipsets in few days. that means manufacturers will receive the new chips before the end of february. considering the manufacturers production rates everything is still on track for april (as expected).