Sign-in / Sign-up
Your question

x-bit labs' Intel V-8 Review-Good+Bad.

Tags:
  • CPUs
  • Monster
  • Intel
  • DIMM
Last response: in CPUs
May 19, 2007 11:44:02 AM

X-bit labs puts Intel's V8 WS-class monster through it's paces, but all is not great. The article reaffirms what most of us predicted(FSB limitations,+ FB-DIMM sucks), yet there are a few surprise performances. All-in-all, probably out of most people's price bracket. :wink:

Intel V8 Review

Comments? :) 

More about : bit labs intel review good bad

May 19, 2007 12:05:56 PM

Nice review, I don't like the board desigh though; the CPUs are placed to close each other and with the heat each of them generates... I don't think they've thought too much about it. The CPU interconnection bottlenecks are often pretty serious and if AMDs 4x4 move was a bit reactive, I think this is kind of a foolish move by intel; it still uses expensive server RAM and they have to wait for CSI to get really effective on dual desktop server like AMDs HTT is doing.
May 19, 2007 12:57:12 PM

I agree... what's here to differentiate this system from a standard workstation, really? Similarly, I get the feeling that by the time games can take advantage of 4 cores, well, people in this price bracket will be running single-socket quad core computers. If 8 cores can be used well, then in time octo-core CPUs will be the next step. Putting all kinds of money into being ahead of the curve isn't a new idea, but frankly, there are so many drawbacks of the dual-socket systems that you aren't really ahead of the curve. If the kind of work you're doing benefits from dual-socket, then chances are you should just be getting a workstation anyway.
Related resources
May 19, 2007 1:04:47 PM

K10 isnt looking a killer on the desktop but its sure looking good on 2+ sockets servers.

8 clovertown cores are badly limited by the FSB in many applications (including my own).

Intel needs nehalem (+IMC) to really get back into servers in a big way.

AFAIK Intel will be releasing 1600 FSB mobos with the server Penryns. That may help somewhat.
May 19, 2007 1:08:03 PM

Quote:
Nice review, I don't like the board desigh though; the CPUs are placed to close each other and with the heat each of them generates... I don't think they've thought too much about it. The CPU interconnection bottlenecks are often pretty serious and if AMDs 4x4 move was a bit reactive, I think this is kind of a foolish move by intel; it still uses expensive server RAM and they have to wait for CSI to get really effective on dual desktop server like AMDs HTT is doing.
I must say, that with a TDP of 150W 8O vs. the FX-74's 125W(and they say that Intels TDP's are low),i was surprised that the 4x4 still drew more wattage at load. It looks like the DIB is helping a bit, but they surely need CSI for 4+ cores. They might actually sell a few of them, if they could get SLI/X-Fire configured properly.
The system definetely has some number-crunching cojonas though(in apps that can use all 8 cores)...the POVray bench was Phenominal :oops: , as well as 3dsMax8, Excel 2007 :o  (4 times as fast as 4x4)...and Fritz9 Chess. They've absolutely gotta get rid of the damn FB-DIMM's, or they have no hope in the enthusiast sector(or at least until FB-DIMM's mature). Gaming sucks!!, but that's sort of to be expected, til more games are mutli-threaded. I would like to have seen what a configuration like that would cost. :?
May 19, 2007 1:08:19 PM

Certainly Anand's article on the P35 chipset memory performance indicated that the increase from 1066 to 1333 gives an appreciable memory subsystem performance boost. It should equally boost other communications if indeed you're having bottlenecks. I imagine this extends well to the 1600 FSB speed.
May 19, 2007 1:14:19 PM

Quote:
I must say, that with a TDP of 150W...


I thought it was odd that CPU-Z detected the processors as X5343. The frequency is still 3 GHz though.

May 19, 2007 1:17:48 PM

Quote:
I must say, that with a TDP of 150W...


I thought it was odd that CPU-Z detected the processors as X5343. The frequency is still 3 GHz though.

Yeah i found that a little weird too. Thoughts of FUDzilla went through my head, at first. :D 
May 19, 2007 1:41:37 PM

Quote:
X-bit labs puts Intel's V8 WS-class monster through it's paces, but all is not great. The article reaffirms what most of us predicted(FSB limitations,+ FB-DIMM sucks), yet there are a few surprise performances. All-in-all, probably out of most people's price bracket. :wink:

Intel V8 Review

Comments? :) 



I think this shows that making a two socket enthusiast platform is rather difficult.
May 19, 2007 2:17:32 PM

This system really has a place, a niche market not your usual gamer or benchmark addict. As a serious user of 3dS Max, Adobe Production Studio and others, I would really benefit from the V8. Right now I have 3 C2D's and 3 X2's set up in a rendering farm along with my main machine. Be nice to have one machine do the work of four. And would be faster, reduced communication over the network. But as I am on a budget, as soon as the June price cuts come, the C2D's are going to swapped for Q6600's and I'll pick up some budget boards to add to the farm.
It's nice to dream though.
May 19, 2007 2:33:59 PM

Quote:
This system really has a place, a niche market not your usual gamer or benchmark addict. As a serious user of 3dS Max, Adobe Production Studio and others, I would really benefit from the V8. Right now I have 3 C2D's and 3 X2's set up in a rendering farm along with my main machine. Be nice to have one machine do the work of four. And would be faster, reduced communication over the network. But as I am on a budget, as soon as the June price cuts come, the C2D's are going to swapped for Q6600's and I'll pick up some budget boards to add to the farm.
It's nice to dream though.

You got it; that's what I am thinking about for myself, but until now, I'd rather wait for the K10 chips, because that is a real solution, while with it's FB-DIMMs and lack of interconnect architecture, the V8 is more or less a failure; a bigger, architectural mistake done by Intel, following AMDs timing mistake.
May 19, 2007 3:06:30 PM

I use the equivalent of a "rendering" farm but for my scientific calcs. I specifically use C2Ds in the "render" farm as opposed to quads because of FSB scaling problems.

I get perfect scaling going from 1 to 2 cores in a C2D but only 150% performance improvement going from 2 to 4 cores in a C2Q i.e. I effectively only use 3 of the 4 available cores.

After July 22 I will be buying another 10 C2D boxes to expand my farm.

Video/animation rendering is embarasingly parrallel so you shouldnt experiance much FSB sensitivity.

The bottom line is you have to benchmark your own software on any multicore platform you intend buying into.

PS: I also recently found that Gigabyte 965 DS3Ps were 15% slower than the Intel 975 XBX2s I currently use in my farm i.e. the premium I pay for the 975 chipset is well worth it in my case.
May 19, 2007 3:21:36 PM

Quote:
X-bit labs puts Intel's V8 WS-class monster through it's paces, but all is not great. The article reaffirms what most of us predicted(FSB limitations,+ FB-DIMM sucks), yet there are a few surprise performances. All-in-all, probably out of most people's price bracket. :wink:

Intel V8 Review

Comments? :) 



I think this shows that making a two socket enthusiast platform is rather difficult.

Precisely, good point Baron ---

A different platform (i.e. Intel instead of AMD), workstation class (i.e. a server/wrkstation CPU combination), running desktop/consumer level software simply shows a lot about how software is optimized for HW for different applications.

Intel is going to fail on this product line, FSN8 will blow it out.

JackYep...they should either drop the idea til Nehalem, or at least drop the FB-DIMM's. The combination of FB-DIMM's, and FSB saturation is more than tons of cache can even handle. :wink: I didn't really think Intel was going to pursue this avenue(when they first demoed V8 )..i figured it was more of a "we can do that too"/proof of concept as well. :?
May 19, 2007 3:25:00 PM

Cache above 2M/core is irrelevant even for their actual architecture and considering that a memory controller makes it more insignificant, it will really have nothing to do with performance.
May 19, 2007 3:25:08 PM

Quote:

I think the reason that the 4x4 system drew more power @ load was due to the 2 power hungry Nividia chipsets that it uses.


Correct. A V8 and 4x4 draw +- the same power but the V8 is twice as fast.

Its fine to criticise the V8s FBDIMMs increased power draw. Fact is the 4X4s chipset powerdraw makes the FBDIMMs look economical power wise.
May 19, 2007 3:30:55 PM

Quote:
Nice review, I don't like the board desigh though; the CPUs are placed to close each other and with the heat each of them generates... I don't think they've thought too much about it. The CPU interconnection bottlenecks are often pretty serious and if AMDs 4x4 move was a bit reactive, I think this is kind of a foolish move by intel; it still uses expensive server RAM and they have to wait for CSI to get really effective on dual desktop server like AMDs HTT is doing.
I must say, that with a TDP of 150W 8O vs. the FX-74's 125W(and they say that Intels TDP's are low),i was surprised that the 4x4 still drew more wattage at load. It looks like the DIB is helping a bit, but they surely need CSI for 4+ cores. They might actually sell a few of them, if they could get SLI/X-Fire configured properly.
The system definetely has some number-crunching cojonas though(in apps that can use all 8 cores)...the POVray bench was Phenominal :oops: , as well as 3dsMax8, Excel 2007 :o  (4 times as fast as 4x4)...and Fritz9 Chess. They've absolutely gotta get rid of the damn FB-DIMM's, or they have no hope in the enthusiast sector(or at least until FB-DIMM's mature). Gaming sucks!!, but that's sort of to be expected, til more games are mutli-threaded. I would like to have seen what a configuration like that would cost. :?

I think the reason that the 4x4 system drew more power @ load was due to the 2 power hungry Nividia chipsets that it uses.Possibly, but i would have thought the power-hungry FB-DIMM's should close the gap, somewhat. Intel really needs to rework their EIST. I've underclocked my E6600 all the way down to~850MHz(94x9)for sh!ts and giggles, and it still works okay, but at 1.2GHz and up it starts putting out quite acceptable performance for mundane tasks(and the temp drops were nice...and that was without dropping vCore). :) 
May 19, 2007 3:53:23 PM

Only burns 15% more under idle with 8 cores and FBDIMM vs 4 cores and desktop memory. Quad FX burns 17% more under load.

In AutoGK/Xvid, it is 45% faster.
In H.264/QuickTime the FX takes 44% longer (or V8 is 30% faster).
In Photoshop, FX takes 44% longer (V8 30% faster).
In Premiere, FX is 20% slower, V8 17% faster.
In 3ds Max 8 rendering, V8 is 69% faster.
In POV ray, V8 is 131.5% faster.
In Cinebench, V8 is 55% faster.
In Excel, FX is 288% slower, V8 74% faster.
In 7-zip, V8 is 25% faster.
In Quake, FX is 10% faster (though QX is 15% faster than that) and V8 is 9% slower.
Then there are some synthetics. Overall, good performance/watt compared to FX. performance/price is a different story.
May 19, 2007 4:11:27 PM

Quote:
X-bit labs puts Intel's V8 WS-class monster through it's paces, but all is not great. The article reaffirms what most of us predicted(FSB limitations,+ FB-DIMM sucks), yet there are a few surprise performances. All-in-all, probably out of most people's price bracket. :wink:

Intel V8 Review

Comments? :) 



I think this shows that making a two socket enthusiast platform is rather difficult.

Precisely, good point Baron ---

A different platform (i.e. Intel instead of AMD), workstation class (i.e. a server/wrkstation CPU combination), running desktop/consumer level software simply shows a lot about how software is optimized for HW for different applications.

Intel is going to fail on this product line, FSN8 will blow it out.

Jack

Man this is just creepy!!!!!!!!!
May 19, 2007 4:17:06 PM

Quote:
Yea tell me about it. Jack agreeing with Baron???? 8O I am about ready to bring out the frozen hell sign again. :lol: 
Yeah, that's a few times...lately. The planets must be aligned this week. :D 
May 19, 2007 4:35:48 PM

Quote:
Yep...they should either drop the idea til Nehalem, or at least drop the FB-DIMM's. The combination of FB-DIMM's, and FSB saturation is more than tons of cache can even handle. :wink: I didn't really think Intel was going to pursue this avenue(when they first demoed V8 )..i figured it was more of a "we can do that too"/proof of concept as well. :?


It would function better and be more appealing if they went straight DDR2 or DDR3, but Intel does not make a chipset that supports both a Dual FSB and DDR2. The FB-DIMMs are a server memory technology plain and simple, it does not mix well in consumer level platforms....

This is a bad idea... at first I thought the computex demo was more a joke or showmanship more than anything, I am a little surprised that Intel actually intends to market a product.

I am wondering if these reviews on the V8 are nothing more than rev 0 platforms, and that a variant of the x38 or an nVidia chipset does not pop up supporting Xeon processors but with using DDR2.... otherwise this is a no-go and not even a close competitor to FASN8.

I am more and more disappointed in Xbitlabs ability to comprehend data though, there is no FSB bottlnecking in the data, they have not run the right experiments to demonstrate that affect. The fact that it 'fails to scale as expected' in some apps is app dependent and not bottlenecking, they need to rethink their concept of what an FSB bottlneck is...

You will see the same affect on FSN8 where some apps (multithreaded apps) do not speed up as expected.

JackSo, what do you think would be bottlenecking the Xvid, Adobe suite, and their multi-tasking scenario? If it's not the FSB, that would only leave RAM latency/bandwidth, or HDD. I would expect the Adobe apps to properly utilize the extra cores.
May 19, 2007 6:44:44 PM

Quote:
I use the equivalent of a "rendering" farm but for my scientific calcs. I specifically use C2Ds in the "render" farm as opposed to quads because of FSB scaling problems.

I get perfect scaling going from 1 to 2 cores in a C2D but only 150% performance improvement going from 2 to 4 cores in a C2Q i.e. I effectively only use 3 of the 4 available cores.

After July 22 I will be buying another 10 C2D boxes to expand my farm.

Video/animation rendering is embarasingly parrallel so you shouldnt experiance much FSB sensitivity.

The bottom line is you have to benchmark your own software on any multicore platform you intend buying into.

PS: I also recently found that Gigabyte 965 DS3Ps were 15% slower than the Intel 975 XBX2s I currently use in my farm i.e. the premium I pay for the 975 chipset is well worth it in my case.


It still seems that economically, C2Q is still better for you. I don't know what kind of machines you are using, so I'll just use a couple examples.

You can buy a reasonably equipped Dell Precision 390 with an E6600 for around $2000. A Q6600 is $420 upgrade. You would get 50% more performance for 21% more money. If you were to buy 10 DC workstations, it would be $20k. For $19,360, you could have 8 QC workstations. Not only would they cost less, but they would do the work of 12 DC workstations.

If you were building E6600 machines for say $1000 each, a Q6600 is about $320 more at retail. This is the same 50% gain in performance for 32% more money. You could have 10 DC machines at $10k or 7 QC machines at $9,240. In this case it is closer, but the QC machines equal 10.5 DC machines for less money.

The advantage the Q6600 has will widen come July 22nd because of the price drops, the cost difference will be less, but the performance advantage will stay the same.

Not to mention, none of this factors in software licensing. The savings can be much greater if you have fewer software licenses to purchase and have less hardware to maintain and repair.

Maybe I'm missing something here, but if not hopefully this helps you get the most bang for your buck come July 22nd.

Ryan
May 19, 2007 7:20:40 PM

Q6600 based systems would cost me 15% more for the same summed GFLOPS. Doesnt sound like much, but it adds up if you buy 10 machines.

I will have to reevaluate the pricing after July 22 but I still expect to go with e6600s OCed to 3.2 GHZ. Yes the Q6600s will drop but so will the E6600s naturally.

If the C2Q offered near linear 2 to 4 core scaling the C2Q would obviously offer the best price / performance ratio.
May 19, 2007 7:28:35 PM

What are you energy costs, and what lifespan do you expect to keep these machines working? I think the total cost of ownership, rather than capital expense, might push more favour on the quad-core.

It certainly is interesting to read about how you've done the analysis and dual-core came out on top. I remember the original thread where you posted your graphs. The conclusion certainly isn't what one would think to be the obvious choice before the analysis.
May 19, 2007 8:15:28 PM

dual athlon beats it in quake 4 ? :?
May 19, 2007 8:20:11 PM

Quote:
K10 isnt looking a killer on the desktop but its sure looking good on 2+ sockets servers.

8 clovertown cores are badly limited by the FSB in many applications (including my own).

Intel needs nehalem (+IMC) to really get back into servers in a big way.

AFAIK Intel will be releasing 1600 FSB mobos with the server Penryns. That may help somewhat.

There is no K10, never was a codename K10. The new chip is just called the Star core.
May 19, 2007 8:46:52 PM

Quote:

I am more and more disappointed in Xbitlabs ability to comprehend data though, there is no FSB bottlnecking in the data, they have not run the right experiments to demonstrate that affect. The fact that it 'fails to scale as expected' in some apps is app dependent and not bottlenecking, they need to rethink their concept of what an FSB bottlneck is...
Jack


Yeah, there is no FSB bottleneck... FBDIMMs performance is terrible. Also, while the snoop filter can increase memory bandwidth (in synthetic tests anyway) it seems to hurt performance under certain workloads. The first thing I noticed when I got my new Xeon system setup was the poor memory performance... in dual channel mode its 36% slower than my E4300 with DDR400 (CL3).
May 19, 2007 8:52:46 PM

Quote:
Q6600 based systems would cost me 15% more for the same summed GFLOPS. Doesnt sound like much, but it adds up if you buy 10 machines.

I will have to reevaluate the pricing after July 22 but I still expect to go with e6600s OCed to 3.2 GHZ. Yes the Q6600s will drop but so will the E6600s naturally.

If the C2Q offered near linear 2 to 4 core scaling the C2Q would obviously offer the best price / performance ratio.


I can't blame you. If it was 5%, with the kind of money you're looking at, I would do it. Heck, a large corporation when buying if all else is equal and they can save 1% on thousands of machines, they would make the same call.

Are you building E6600 machines that are less than $640? That is the only way I can figure that the 50% increase in performance does not outweigh the >50% increase in cost. (640+320 for Q6600) At $640 for an E6600 machine, the Q6600 machine at $960 provides the same GFLOP/dollar. So 2 Q6600s could do the same work as 3 E6600s.

I'm interested because I have recently been getting into Maya and have found that rendering scenes in HD takes a lot of work. Maya licenses are so expensive, it is much better to keep the number of machines down, and the number of cores up. Does your software not require additional licenses for additional machines?

Ryan
May 19, 2007 9:18:06 PM

Quote:
What are you energy costs, and what lifespan do you expect to keep these machines working?


Very expensive power.
3 - 6 months of compute time (not wallclock time).
May 19, 2007 9:21:50 PM

Quote:

Are you building E6600 machines that are less than $640?

Does your software not require additional licenses for additional machines?



I write my own sw. It runs on http://en.wikipedia.org/wiki/Message_Passing_Interface and http://www.rocksclusters.org/

Here is my costing analysis:




Costs are in British pounds. Should be self explanatory. "4500 MHZ" for the Q6600 is just the equivalent performance Ive measured in benchmarks i.e. 3 GHZ OC x3 cores.

I have 6 test machines now and will get another 10 after July 22. Its going to be cool to walk in and hit and space bar and listen to 16 boxes spin up 8)
May 19, 2007 9:46:48 PM

What do we expect the cost of the V8 motherboard(s) to be? I wouldn't be so against the 4x4 if the mobo wasn't bloody expensive. Still, I think the 4x4 (and probably V8) are good deals in terms of workstations - definitely not gaming machines.
May 19, 2007 10:18:14 PM

Your links to the cost analysis spreadsheet appear to be incorrect. They go right to the imageshack homepage. Could you correct them? I'm interested to see your analysis.

Ryan
May 19, 2007 10:21:05 PM

Quote:
What do we expect the cost of the V8 motherboard(s) to be? I wouldn't be so against the 4x4 if the mobo wasn't bloody expensive. Still, I think the 4x4 (and probably V8) are good deals in terms of workstations - definitely not gaming machines.


There available now. Just look for a board from Intel, Tyan, or Supermicro with a 5000X chipset. I've seen them for as little as $350.

Ryan
May 19, 2007 10:22:51 PM

You're absolutely right Jack; it has nothing to do with FSB bottlenecking. This thing has much more to do with the poor CPU to CPU communication due to the FSB but it's not related directly to it's bandwidth, no way. It might also be connected to cache conflicting, or more stupidly, with the fact that that motherboard has not been well designed but just rushed.
May 19, 2007 10:45:39 PM

I see, you're building bare bones boxes yourself and overclocking them. It will be interesting to see if that changes when the prices drop in July.

Ryan
May 19, 2007 10:55:45 PM

Ive seen shared memory molecular dynamics applications where having the snoop filter on drops performance by 15%.
May 19, 2007 11:06:56 PM

Quote:
Ive seen shared memory molecular dynamics applications where having the snoop filter on drops performance by 15%.


Yeah, not too surprising...

Also, noticed you changed up your signature file some.... I am wondering what food chain that would be?

The only thing it makes me think of is a book; "The Time Machine" :D 
May 19, 2007 11:10:41 PM

Quote:
[
Also, noticed you changed up your signature file some.... I am wondering what food chain that would be?


Baron needs to keep his crap short and sweet so I can fit more of it in. He's been rambling lately ...
May 20, 2007 12:55:29 AM

This doesn't look like the official 'V8' platform Intel announced, called Skulltrail.

For starters, Skulltrail had Intel's Common System Interface replacing the FSB and had multi GPU (most likely CrossFire) compatibility.

EDIT: Just had a look see at the motherboard, it's just a plain old motherboard based on the 5000X chipset. These boards have been out for quite a while.

Nothing new here. :/ 
May 20, 2007 1:45:33 AM

"V8" is simply a marketing initiative to reminde people that you can buy an 8 core machine now. Skulltrail will be more than a rebadged server, SLI/CF, regular DDR2 and whatnot.
May 20, 2007 1:52:05 AM

V8 sux as much as 4x4.
WTF are Intel doing?
May 20, 2007 3:41:00 AM

Quote:
"V8" is simply a marketing initiative to reminde people that you can buy an 8 core machine now. Skulltrail will be more than a rebadged server, SLI/CF, regular DDR2 and whatnot.


Sorry to disappoint you Rock, but the Skulltrail only takes FBDimm's. As for SLi, yes it will be the first desktop motherboard that officially supports SLi.
May 20, 2007 7:02:11 AM

It could have been worse... I don't really think its performance is bad. Unless you want to use it for gaming... but for heavily multi-threaded apps and multitasking it looks pretty good. Plus it has lower power consumption than 4x4 at load. Idle power consumption isn't very good... although with only a single processor its only 8.8 watts more than QX6800 setup.
May 20, 2007 7:16:17 AM

Quote:
"V8" is simply a marketing initiative to reminde people that you can buy an 8 core machine now. Skulltrail will be more than a rebadged server, SLI/CF, regular DDR2 and whatnot.


Sorry to disappoint you Rock, but the Skulltrail only takes FBDimm's. As for SLi, yes it will be the first desktop motherboard that officially supports SLi. You sure about SLI support? Currently nvidia does not support the 5000x chipset for SLI. Although Im sure it could work... maybe with a driver hack?
May 20, 2007 7:22:13 AM

Quote:
V8 sux as much as 4x4.
WTF are Intel doing?
What were you expecting? Intel to produce a new chipset for such a low volume product? Its not like performance characteristics for this kind of setup were unknown.
May 20, 2007 7:31:37 AM

Intel is using some special BGA parts that Nvidia sells. I guess because Nvidia has no products in this category they feel free to let Intel to support SLi on it.
May 20, 2007 7:45:01 AM

Quote:
Intel is using some special BGA parts that Nvidia sells. I guess because Nvidia has no products in this category they feel free to let Intel to support SLi on it.
Thats interesting. Do you have a link to any info on that?
May 20, 2007 7:50:48 AM

Nope, just what my friend at Intel has told me. I asked him how much PS you would need if you fully loaded out a system. He just laughed. I took that to mean it goes beyond anything you would think of. The board will have 4 x16 PEG slots. Take two Extreme Edition processors and I will let you add up the cost in Power. I don't really see this becoming the ultimate game machine unless you can get DDR2 800 FBDIMM's in the system. Even then, I see a Skull Trail system as the ultimate workstation.

It should have full overclocking capabilities to match what is on the Bad Axe 2 if not more.
May 20, 2007 8:06:21 AM

Thats the most Ive heard about skulltrail so far...
May 20, 2007 8:14:02 AM

Quote:
I don't really see this becoming the ultimate game machine unless you can get DDR2 800 FBDIMM's in the system.
Memory performance will most likely still be a problem. I think the bottleneck is either the AMB or the chipset. When I overclocked my Xeon from 266MHz FSB to 333MHz FSB I only got an extra ~4% memory bandwidth... Of course thats with only one processor and 2 FBDIMM channels.
      • 1 / 2
      • 2
      • Newest