LGA-1156 (new Lynnfield i7/i5 on p55 chipset and beyond)

[valis]

So the reviews are out all over for p55 chipset (LGA-1156 initial platform) and Lynnfield cpu's (45nm generation for LGA-1156), and there seems little reason for LGA-1366 era Nehelem/i7 for most users now (only the $999 i7's still have any performance advantage for single socket systems but that's certainly not the 'sweet spot' in pricing for most).

I can hazard a guess that there's a lot of potential for audio performance (when is this not true?) but right now there are some 'interesting' benchmark results that point at possible motherboard trace/BIOS/driver issues. Some of these are related to the PCIe bus's changes I think. For the adventurous ones who do want to try to learn the Scope 'adjustments' needed on those boards I encourage your posts but as an investment I would perhaps wait until volume ramps and you can get rev2/rev3 motherboards (november/december.) I'm sure the benchmark errata that shows odd results atm will resolve by then, meaning whatever architecture specific tunings need to be done will be in place (let's just hope this isn't another generation plagued by graphics card drivers interfering with low latency audio & dsp needs--ie the on-cpu PCIe controller starving the cpu for data.)

In the end the changes this brings to the boards ought to get this generation's pricing of hardware back down to where Core2 was as well, and that's great news imo. Dual channel ram means it's cheaper to populate your board, ie, with 4 sockets instead of 3 since you don't lose the 3rd channel with all 4 slots when you only have dual channels to begin with. And dual channel 1333 (validated) with Lynnfield seems competitive to the triple channel i7 for everything but dual GPU users (who lose out on 2 16x channels for gpu alone) so there doesn't seem to be much loss.

That last part is perhaps also relevant to audiousers with LGA-1156/Lynnfield. This is because there is only 1 16x controller that gets split evenly for dual gpu support when the user wants that. The way the support for 2 8x lanes is enabled (using simple switches) I would wonder if the gpu will happily split the direct cpu connected PCIe bus with anything else, leaving you with only PCIe 1x slots available for anythign else (though there are 8 PCI 1x lanes now with the P-55 chip instead of 6 on the ICH10 'south' chip for X58/LGA-1366/Nehelem.) X58/Nehelem do have up to 36 lanes connected to the one of the QPI busses which enables motherboard makers to offer multiple arrangements for PCIe slots higher than 1x. Now most existing PCIe audio products that I know of *are* PCIe 1x. Though I can't recall offhand if Xite differs from that I would guess it probably doesn't since RME's HDSPe MADI card can do 64in/64out with a single 1x slot.

Intel states that a 1x slot has a peak bandwidth of 500MB/s with the DMI connection between the cpu & P55 chipset is 2GB/s, which is the same as between older ICH10 & X58/X48. But now that PCIe cards are emerging for audio (and higher samplerates are becoming more common) I wonder how soon that DMI interconnect will show saturation issues. Especially if there's a bug with the current DMI>CPU connection this might be an issue for those who use multidrive arrays or SSD's along with multiple PCIe dsp cards and a high i/o count PCIe card. (As an example on early PCIx implementations in Intel Xeon chipsets, the PCIx bus could stall the PCI 'subbuss' if it was getting hit hard enough; something that a few Xeon users with dsp cards may have noticed if they used onboard SCSI on the PCIx bus hard enough *while* using Scope dsp cards or other makes.) As I mused in the first paragraph though, hopefully a few board revisions & BIOS tweaks will solve anything show-stopper, and for a midrange user there's still PLENTY of usable bandwidth available on these systems forthe near future.

Plus the good news is that power consumption for Lynnfield is unbeatable as of yet (aside from cpu's that are specifically targeted at low-power).

Thoughts?

[petal]

Thanks for that thorough analasys Valis!

My thoughts, this looks like my next platform.

I'll be looking forward to hear the tales from the brave souls willing to be pioneers and test out these new products with a couple of "the old" pci scope cards installed...
Come to think of it, one of those brave soul might even be me - my old trusted system is starting to show it's age and I've been looking into a suitable replacement for a while.

So what do you guys think, can we trust the P55 chip set - will it work with 3-4 Scope pci cards installed? (yes I have four I'd like to get into my next machine )

[valis]

There do seem to be a few boards with 4 slots, a few more with 3 but many boards have only 1 PCI slot (or some even have none.) Also be sure to check PCI 2.1 compatibility (5volt keying)

• Gigabyte has SEVERAL models already with 4 PCI slots on their P55 page
• Asus P7P55D does seem to have a 3 slot variant but the models I see listed on review sites seem to vary in layout for the same model # right now (there are several variants and it's unclear right now which is the one with 3 pci slots since Asus's main site seems to lack p55 boards right now)
• MSI P55-CD53 seems to be MSI's only model with 3 PCI slots right now
• Same with Foxconn, the P55A-S listed on newegg seems to have 3 slots but both of their P55 models listed on their website have only 1 PCI slot
• Newegg also lists EVGA's P55 LE 123-LF-E653-KR as having 3 slots but checking EVGA's website only shows models intended for gaming with 1 PCI slot right now
• AsRock also has the ASRock P55 Deluxe LGA 1156 Intel P55 board listed on Newegg with 3 PCI slots
• DFI's p55 Lanparty board listed on their site only has 2 PCI slots
• Supermicro currently only lists socket 1156 'xeon' machines that lack any PCI slots or have 1 at the most

[garyb]

nice work v!

[Sounddesigner]

From the benchmark i seen the new i7 870 does not look like a bargain. It's retail price is $560 iirc and the older i7 920 outperforms it in that particular audio benchmark and its price is only $280. Below is the benchmark from ADK.

dawbench.jpg (104.46 KiB) Viewed 4435 times

PS. The i5 750 seem to be a bargain tho.

[valis]

Interesting. But think about that graph for a second. Besides the fact that it makes no mention of system configuration or motherboard make/model/bios revision...and this is very early in the P55 game so that info is (imo) quite relevant. Also the i7 920 price ($280) are intended to be versus the i5 750 ($210) and i7 860 ($299) and once tuning issues settle out I think the performance will reflect that (and that's before even accounting for overall system cost).

In the more general computer benchmarks, the cache-coherent (on-core) multithreading & single threading tests (like povray/cinebench/blender) and the memory bandwidth tests show that the p55 platform is capable and performs as you'd expect based on pricing compared to the i7 nehelems. It would be really useful if ADK Pro Audio (who made the graph you posted) could do more testing and isolate exactly why you're seeing such reduced performance in that benchmark versus the other benchmarks out there right now.

The new cores bring higher frequencies at the same price points as the previous i7, and the cores haven't been cut down beyond removing the QPI connects. Trading the QPI & triple channel ddr3 for dual channel at lower latency & higher (validated) ddr3 speeds wouldn't show the disparity alone (shown in the graph above). And the cache sizes etc are unchanged though the new core has had a few specific long instruction eatures added and other core optimizations. Also it seems that DDR3-1600 is running fine on the p55/LGA-1156 boards from several makers (though beyond intel's spec) and OC'ers are getting DDR3 speeds up to 2000mhz (unnecessary for a DAW imo) so the DDR3 support has matured fast compared to X58's launch (which you would expect since mobo makers have been making X58 for a year now.)

So I would check back in 1-2 months and see if the benchmarks aren't changing. I would suspect those benchmarks are showing a BIOS tuning issue or something with the motherboard/chipset needing to be debugged. I remember that the mobo makers had a hell of a time with the x58 motherboards at launch too (though off the time of my head I can't recall if that was affecting audio benches as I just recall it affecting some RAID/SSD users.) That was actually close to 12 months ago now so you would certainly expect the x58/LGA 1366 to show its maturity.

Once you factor in the cost of an x58 motherboard & a usable amount of ram I believe you'll see prices on more par with building a quad Core2 'midrange' system, which is what I see most audio users targeting these days. When Intel brought out nehelem/i7 they made a break from the core/core2 line to address some shortcomings in the server/high end workstation markets. Lynnfield is again making a break from Nehelem to reduce motherboard complexity (quite a lot in fact) and reduce featureset for their middle of the market offerings.

So for users who don't need the top machine available at the time of purchase, I do still think that Lynnfield will scale quite nicely and be the modern alternative to Core2Quad for most. Just a question of finding the bugs (as always)

[dawman]

Intel was smart and really helped their partners with the X58 motherboards and triple channel RAM. Everyone made money and AMD had no answer.
Triple Channel RAM won't even show a difference until 6 and 8 core CPU's are realeased.
The 1156 Supermicro boards are designed for 1U's. The 16MB DDR2 Matrox chip and the Xeon CPU have got my attention. 32MB's of RAM is a big plus also. This means 16GB's using 2GB DIMM's. The Matrox chip seems small, but it doesn't use system RAM and my first Scope DAW used a Matrox G450 w/16MB's of RAM, so this could be a great 1U choice for XITE-1 guys.
One thing is certain....AMD CPU's and AM3 motherboards are forced to drop their prices to compete. 2010 is going to be a great year to build or buy.

[valis]

I would expect this generation is less about you Jimmy, and more about the people who likely will never even read this thread

No die shrink yet, Intel decided this year instead of doing the 'tock' as a die shrink they were just going to do more than the normal core stepping but at the same process (Lynnfield is 45nm still, and the only real changes are a few extra instructions for media/encryption and the uncore changes to remove QPI for the native PCIe controller and simplifiied mem controller.) Also Harpertown is out last year, I've got them in this Xeon rig. I think you meant "Westmere" which is Clarkdale/Gulftown (the desktop/Xeon monikers for 32nm Westmere), and yes that should be the next BIG performance jump. Intel is deviating the year from tick-tock, by doing tick-tick with the Westmere tock coming after. Hence the lack of major performance gains, this is closer to a core stepping with some uncore changes.

The real question I have for Westmere (desktop) is what socket they'll use, and if it's one of the 2 current ones will the voltages & motherboards out now be compatibile?

Anyway the real reason I think that P55/Lynnfield is the answer to Core2Quad for the future is because of the prices for X58 motherboards. This is the first generation that I've seen prices on the desktop motherboards at price parity or HIGHER than the Xeon/workstation boards (which are a bit cheaper this generation due to needing less onboard features.) Previously the only boards that were more expensive were bleeding edge boards targeting niches (triple SLI nforce tuned for OC'ers or the silly Skulltrail crap that was really just repackaged Xeons with more limited RAM layouts.)

And again I think that's the current problem too, but considering this is 1 week after launch that's certainly not unexpected. Yet another HUGE change in the way that motherboard makers & BIOS vendors need to address & layout the hardware, and only 12 months after the X58...

[Neutron]

Supposedley "just a bios update" for current LGA-1366 Motherboards,
but i have heard that song before.

The first version (6 core 32nm) will be Extremely expensive edition for a while anyways. and i somehow doubt they will ever release one at as good a price as the 920.

By the way that graph is almost the same when comparing this dual Xeon 5430here at work with i7 920at home for 3ds MAX rendering time. the dual xeon took 360 seconds to render a frame of my scene and the i7 took 280 (there were no bitmap textures, all procedural, so texture loading time does not count for anything)

Jimmy using a PCI matrox card seems a bit silly with scope cards, even iof it is just sipping on its own memory when it is displaying, how do you think it gets told what to display? through the PCI bus.

even a crappy $40 ATI on the PCIe ought to be better these days. the bus they are sharing is like the 8 lane highway compared to the dirt track donkey road that is PCI.

[valis]

Agree about the PCI vs. a cheap PCIe card, the only issues PCIe graphics has ever had were due to improper motherboard/chipset (specific motherboards or chipsets) tuning or the VERY early days of PCIe. I've got a big GTX285 plonked into my slot and haven't ever bothered with PCI latency adjustments for it to increase audio performance.

Also I've got a Xeon 5430 as well, the reason these bench below the i7 is of course the FB-DIMM ddr2 ram + MCH latency (higher latency than ddr2 on a 'cheap' core2duo/quad board even) versus the QPI direct connection to ram, as well as the architecture overall that Nehelem is in general for SMP (beefed up oncore/pipeline.)

[dawman]

I just awoken from a bad dream.
And actually it is of interest to me because of 1U possibilities.
I do this a couple times a year and get jager'd up.
Dude that one was so bad it needed erasure....
Westmere is the version to shrink and add the GPU on die....that's right.
Thanks for the slap.

[valis]

Larabee (Intel's upcoming GPU) is a while off still, but this conversation is still interesting.

First I'll say 2 things.

I get no financial kickback for anything said here, and am not trying to viscously defend some imaginary stance on a p55 motherboard that I don't even own etc I'm just curious to hear input. I'm certainly not trying to argue that Lynnfield is the definitive answer to our audio woes. That never applies for longer than a few months anyway...what I *am* interested in is the path this is going.

The reason I am keeping an eye on things again (to the degree that I am atm) mostly because I'm watching for Clarksfield (mobile i7 in older fpga format) for an upcoming laptop purchase, and I'm also watching to see if I should buy into my next desktop with Gulftown/Clarkdale (from "Westmere") or wait another generation & see what Sandy Bridge brings. This is where I 'come from' so to speak

Part of my decision for the next desktop is to potentially abandon Xeon for the first time ever (having used duals since before the PPro/P6) and I'm really interested in seeing where Intel's sockets go as one of the things that would get me to go to a single socket board again is upgrade path. The only reason that I've stuck with Xeons over the years is for my 3d & graphics needs, and I have been thinking even desktops are strong enough now that both my audio and graphical needs are just not requiring a Xeon (the longer upgrade cycle is nice but the latency hits & weird cache contention or NUMA related issues that crop up don't make Xeons a win for every situation.)

As for "BIOS issues", the main reason I included that in my initial post was because there are some benchmarks showing errata with PCIe performance under Win7 and even just general multithreaded performance under Win7, errata that doesn't show up for nehelem i7 which uses the same HT/TURBO arrangement for its cores. Vista & Xp32 didn't show the same issues, so the consensus was that it's between MS/Nvidia/ATI to get DirectX titles where they should be and possibly some BIOS tuning to account for the new PCIe integration into the core. Balancing & timing the load so to speak. But noone here cares about gaming performance when we're talking audio benchmarks so just stating there is some errata seemed enough.

So the other errata we have introduced to this converation is that graph above, and honestly it's the first benchmark besides the ones I just mentioned where I've seen the 870 underperforming based on its pricepoint & configuration. The only real 'losses' that Lynnfield has shown against 920's are either memory bound tasks (expected with the reduction of 1 channel and mitigated quite a bit by the increase UNCORE speeds of the newer cpu's and the stabler performance with high speed ddr3,) or GPU scaling (especially for CUDA/OpenCL on nvidia). So I would really be curious to see why 870 takes such a hit in that particular benchmark. I did send an email to ADK Pro Audio to see if they had any explanation or more figured to go from, will report back if I hear anything.

Basically I only see 1 non-Xeon LGA-1366 (nehelem's socket) cpu listed on Intel's roadmap right now for Westmere, and it's the Gulftown cpu which should be branded i9. It looks great but it's currently listed as $1500, intending to be the 'Core2Extreme' pricepoint for that generation. Of course other Gulftowns will be Xeons, but where are the lower pricepoints for i9/i7? Intel seems to be saying the p55 introduction of the LGA-1156 socket is the new desktop socket from any info I've seen so far? So the questions I was posing are: will current p55 boards handle Clarkdale (upgrading to next gen i5) in terms of cpu support/voltage etc and will Clarkdale scale enough in performance to actually be an upgrade?

BIOS issues are much easier to answer in a few weeks time, the upgrade path is really up to Intel to flesh out. Either way, I guess we'll see...

By the way that graph is almost the same when comparing this dual Xeon 5430here at work with i7 920at home for 3ds MAX rendering time. the dual xeon took 360 seconds to render a frame of my scene and the i7 took 280 (there were no bitmap textures, all procedural, so texture loading time does not count for anything)

I also meant to say that the graph does represent the disparity between Harpertown 5400's and i7, but the i870 is positioned in an odd place versus any other benchmarks I've seen. Specifically for 3ds MAX & other 3d benchmarks, they all see to look something like this:



Memory benchmarks of course still show i7 leading, but most of those aren't stressing the multiple PCIe lanes either. Stick a PCIe SAS card in RAID 10 (or RAID 0 with SSD) on there and try to do workstation graphics at the same time and you'll see why there's so much QPI bandwidth around. Lynnfield doesn't need that (and the Uncore speed bump along with the lower latency memory accesses without the hub attached to the QPI) seem to be doing fine with the platform in most less synthetic benchmarks (ie, besides the ones that only test a given subsystem.) Except that multiband compressor benchmark of course...

[Roman A.]

Thanks for this thread, valis.

I am relieved to have found a place where these benchmarks are properly analyzed! I searched nearly every other forum only to find very generalised information, and it seems that more specialzed forums like DAWbench or the benchmark forum at Cubendo are not accessible for ordinary people.

I too am currently planning on building a new DAW and also find the lynnfield's lackluster audio performance in the ADK benchmark surprising, as it can generally compete with the bloomfield i7 920 in every other setup. If not for this single benchmark, i probably would have bought an i7 860 already.

If Scott answers to your questions, please let us know. I am hoping to see more audio benchmarks with the lynnfields in the wild the next few weeks.

[siriusbliss]

great analysis.

Looks like I'll be waiting til next year.

Greg

[Sounddesigner]

@Valis, Scott from ADK did state " FYI my 870 numbers may be a tad low due to the use of Beta Intel board they sent. now that we have retail release we will be retesting."

So i gather that the i7 870 numbers may improve a bit the next test.

The i5 750 will be added to the benchmark soon.

[dawman]

ADK does benchmarks for apps I use like Kontakt and also is one of the reasons I was more interested in his AMD benching.
Many local guys ( Memphis, Nashville & St.Louis ) have a great rapport with ADK and hardly need any support..
I am only interested in which platform will stream well, with low heat in a 1U without the need for a video card...

Brotha' Man Neutron, I was only commenting about the Graphics VGA chip from Matrox that comes on the Supermicro board.
In the past they have had 8MB sized chips which were not enough to run 2 or 3 apps on a display. From using the old Matrox card in the AGP4x slot, I know 16MB's was O.K. for Giga/Scope, but I use VST hosting and multiple apps now.
So far the Sideport RAM motherboards are the only solutions I am seeing that work, but these AMD CPU's get damn hot in a 1U.
The 1156 w/ dual channel RAM would be fine, as long as the 1156 motherboards from Supermicro and the 16MB Matrox chip can run the spps without screen redraw troubles, and a sticky mouse.
Sharing system resources is no longer an issue as I have seen the 785G's bypass the Sideport RAM, and use the system RAM w/ zero conflicts, and the XITE-1.
My problem with AMD is the heat, even the 95W CPU's are hot, the one I want to use is the X4 at 140Watts.............this isn't a very comfortable feeling, as even in a 4U it has noticable heat coming from the fans of the PSU which I have never felt before.

So I am looking at the 1156 and wish I could see it in action before springing for the Xeon version, and mobo / RAM.
I really never would use Cubase or Reaper w/ so many plugs, etc. I bought the XITE-1 for those chores. So the 1156 is something I will watch because of the low power, and slower speeds are fine as this will be for the PCI-e 1X connectrion and RAM streaming.
These racks are a major PITA and I must downsize as soon as possible. Laptops are not an option because of the extreme pricing for 8GB models.

http://www.supermicro.com/products/moth ... 3000/#1156
http://www.supermicro.com/products/chas ... 2F-280.cfm


These guys are interested in frames per second but stable long life solutions in cramped spaces.

[valis]

That makes more sense. I essentially eliminated SuperMicro from my little bulleted list because of the lack of PCI slots, but they do all have at least 1 PCI & 1 or 2 PCIe slots so you would still be able to get at least 1 audio card or the Xite card, eh? I also hadn't realized that Matrox had moved onboard like that, my board has a disabled ati rage xl which has been on most boards for server use since I can remember...

Good to share the mixture of opinions then since we certainly have different needs that result in different 'takes' on the platform.

[dawman]

I am interested in the Xeon L3426 @ 1.86 with 4 cores/8 hyperthreading with 8MB cache for $284. The fact it is 45Watts is what I am interested in. How will that even get hot...?
Streaming only with this might be enough, but the other Xeons are awfully cheap this time around too.

http://www.siliconmadness.com/2009/09/i ... -xeon.html

This should be a fine streamer using the 3.5" Cooling enclosures w/ 2 x SSD's.

http://www.supermicro.com/products/syst ... 016I-T.cfm

[Sounddesigner]

I stole this benchmark from KVR. It's said to be done by dawbench.com but TAFKAT has not posted it in any forums most of us visit yet so whether this benchmark is legit or not is'nt fully verified yet. The benchmark shows the i5 750 competing well with a i7 920 without HT enabled wich is the performance ADK said it is capable of.

[valis]

It's too bad they didnt' included i860/i870 since they're faster versions of that i5 with HT & TURBO core support enabled.