Intel Core i7-4960X Preview: Ivy Bridge-E, Benchmarked

[SNA3]

So far there are not even 4 core aplications out there, so Intel dosent bother. Just improves things a bit each now and then, and charges the same price. Ar this rate, they will start to see their income drop fast.

Untill we get games using 4+ cores, i dont see this moving at all.

correction : no Games.

there ARE Applications that use every single core you have even if you have 32 cores

3D rendering and movie Editing and AutoCAD and all that stuff.

 

[robertoivs]

Hi to all,

in this review I have found several errors in the analysis. The author says that Core i7 3820 is not enough competitive with core i7 4770 or 3770. This claim is absolutely wrong.
In a real work enviroment using any NLE software like Premiere, Vegas, Edius, ivsEdits etc.working with HD footage (1920x1080) when large amount of data have to be moved in memory (It is not very difficult it is enough to set some PIPs) the i7 3820 Quad channel controller offers exactly the double of performance than dual channel memory i7 Cpus. It is easy to check that in many situations i7 3820 allows to mix double Timeline layers than 3770. The author should know that a quadchannel memory controller allows up to 4 cores can read/write the memory while as it is obvious in a dual channel system only two cores can access to memory. Moreover really not only the cpu cores are affected by memory controller channels also PCI-E DMA accesses can lock some cores to receive data. This means that all 4 cores of corei7 3820 have better probabilities to work on data on most available time while with i7 4770/3770 two cores have to wait while the either 2 cores are accessing to memory. With current Tomshardware benchmarks this huge difference not appears very evident but If they benchmark the same test opening two concurrent sessions the sound will appear completely different.

Some other few words on PCI-E 3. there is a tiny benchmark that measures CPU COPYMEM performance and raw PCI-E transfers speed. We have not fond yet any difference between PCI-E 16x 2.0/2.1/3.0 so many words spent on 8GT/s are only marketing numbers.
http://www.ivs.it/public/ivsEdits_Folder/Setup_ivsEdits_SpeedTester.exe

Plese note : running the test with AMD Catalyst 13.6 Beta2 all RADEON 7xxx GPUS dont are able to switch the clock from Idle to work phase so the result read in 1920x1080 FPS is totally disappointing. AMD is already informed about this iussue.

 

[Someone Somewhere]

*smacks head into desk*

The author should know that a quadchannel memory controller allows up to 4 cores can read/write the memory while as it is obvious in a dual channel system only two cores can access to memory.

The poster should know that this is entirely wrong, and that bandwidth is shared across all cores. I guess all those that buy chips for servers with quad-channel RAM and 8 cores are stupid?

With current Tomshardware benchmarks this huge difference not appears very evident

Because it's nearly non-existent. Very few programs are bottlenecked by RAM bandwidth.

Some other few words on PCI-E 3. there is a tiny benchmark that measures CPU COPYMEM performance and raw PCI-E transfers speed. We have not fond yet any difference between PCI-E 16x 2.0/2.1/3.0 so many words spent on 8GT/s are only marketing numbers.

Evidently you haven't tested SAS cards, which can max out a PCIe2.0x8 slot, with fast SSDs.

 

[slicedtoad]

If it OCs well, it may prove interesting to a few of the richer enthusiasts.

 

[ThermalV]

You cannot complain about incremental performance improvements, there is simply no need for massively powerful CPU's because current ones already do their jobs damn well.
And for those comparing the speed of GPU performance to the slowness of CPU development need to appreciate that GPU has become increasingly interesting as applications take advantage of the parallelism they provide. More interest = Greater development budgets = Faster advances in tech.

 

[mapesdhs]

You cannot complain about incremental performance improvements, there is simply no need for massively powerful CPU's because current ones already do their jobs damn well. ...

That depends on what one is doing. Those who work with rendering systems, complex compilations, transaction
processing, QCD, CFD, and grud knows what else would I am sure thoroughly disagree.

I know a guy doing medical research who wants as many cores as he can get. His computations
run for hours or days. Just because you don't need more performance doesn't mean that applies
to everyone. Those doing pro work always want more speed.

Ian.

 

[Someone Somewhere]

Those people are likely to be able to get work to spring for Xeons (which also have ECC etc). IB-EP should have up to 12 cores IIRC, and you'll be able to do dual-socket builds.

 

[mapesdhs]

... there ARE Applications that use every single core you have even if you have 32 cores

3D rendering and movie Editing and AutoCAD and all that stuff.

Talking to someone at the Cosmos facility, apparently there's a lot of work going on atm to get
certain codes scaling beyond 512 cores, so they can take better advantage of the larger systems
that are now available, whether it's fine-grained codes that benefit from big shared memory
systems, or course-grained codes for running on large clusters. The Cosmos admin mentioned
galaxy simulations as one example, but there are dozens of other fields in science & industry,
and of course as you say film rendering.

CFD applications can already exploit more than 3000 cores though. It varies from one app to another.

Here's an extreme example of the compute demands required today:

http://www.sgi.com/company_info/newsroom/press_releases/2013/june/spirit.html

ThermalV, that system has more than 73000 cores. Feel free to tell the DoD they don't need them.

Ian.

 

[J_E_D_70]

Hi to all,

in this review I have found several errors in the analysis. The author says that Core i7 3820 is not enough competitive with core i7 4770 or 3770. This claim is absolutely wrong.
In a real work enviroment using any NLE software like Premiere, Vegas, Edius, ivsEdits etc.working with HD footage (1920x1080) when large amount of data have to be moved in memory (It is not very difficult it is enough to set some PIPs) the i7 3820 Quad channel controller offers exactly the double of performance than dual channel memory i7 Cpus. It is easy to check that in many situations i7 3820 allows to mix double Timeline layers than 3770. The author should know that a quadchannel memory controller allows up to 4 cores can read/write the memory while as it is obvious in a dual channel system only two cores can access to memory. Moreover really not only the cpu cores are affected by memory controller channels also PCI-E DMA accesses can lock some cores to receive data. This means that all 4 cores of corei7 3820 have better probabilities to work on data on most available time while with i7 4770/3770 two cores have to wait while the either 2 cores are accessing to memory. With current Tomshardware benchmarks this huge difference not appears very evident but If they benchmark the same test opening two concurrent sessions the sound will appear.

That's great, but the focus of the majority of reviews on this site is gaming. Should be apparent since they review gaming rigs not graphics workstations and the monthly "for the money" articles for cpu and gpu are "gaming."

 

[mapesdhs]

Those people are likely to be able to get work to spring for Xeons (which also have ECC etc). IB-EP should have up to 12 cores IIRC, and you'll be able to do dual-socket builds.

You'd be amazed how many solo professionals and smaller companies would love to use XEON-class
hardware, but just can't afford them.

It's ironic that certain consumer boards support XEONs, because with the lower clocks and locked
ratios, they're slower for threaded apps than an oc'd i7. The 8-core E5 vs. oc'd 3930K is typical.

Now if Intel released an unlocked E5, that would be a whole different ball game, doubly so if one
could put 2 such chips on a board and oc the hell out of them.

As for 12-core IB-EP, I expect their usefulness will depend on the imposed TDPs and hence the
resulting clock levels.

Am I right in thinking that atm there isn't any consumer board that can take more than one 6-core i7?
Not built for it AFAIK. 2 XEONs, sure, but the locked ratios spoil the gains.

Ian.

 

[Someone Somewhere]

Multi-socketing is exclusive to chips with it enabled; Intel disables it on consumer chips, and some of the cheaper xeons.

 

[mapesdhs]

Multi-socketing is exclusive to chips with it enabled; Intel disables it on consumer chips, and some of the cheaper xeons.

Thought that was the case (wasn't 100% sure).

Beats me why they don't offer even one unlocked E5. It would offer so much more flexibility
for 2-CPU pro desktops.

Ian.

 

[ThermalV]

You cannot complain about incremental performance improvements, there is simply no need for massively powerful CPU's because current ones already do their jobs damn well. ...

That depends on what one is doing. Those who work with rendering systems, complex compilations, transaction
processing, QCD, CFD, and grud knows what else would I am sure thoroughly disagree.

I know a guy doing medical research who wants as many cores as he can get. His computations
run for hours or days. Just because you don't need more performance doesn't mean that applies
to everyone. Those doing pro work always want more speed.

Ian.

These are essentially consumer chips, if you want pure grunt and stability you go xeon.

 

[ThermalV]

Now if Intel released an unlocked E5, that would be a whole different ball game, doubly so if one
could put 2 such chips on a board and oc the hell out of them.

I don't think you quite appreciate the point of a xeon chip. They are designed to be the pinnacle of stability for mission critical tasks, overclocking is typically not a great idea for stability as it can be unstable.

 

[mapesdhs]

These are essentially consumer chips, if you want pure grunt and stability you go xeon.

Not everyone who would wish to use them can afford XEONs.

And btw, I know pro artists at big movie companies who would love to build some oc'd consumer
setups with pro cards shoved in them, because many apps like Maya are often single-threaded in
various aspects of their interactive functionality. An oc'd 3930K is a darn sight better for running
Maya than a dual 4-core XEON. Alas such companies tend to be somewhat inflexible in what they
can do, as they often have large budgets, purchasing policies, etc.

Ian.

 

[mapesdhs]

I don't think you quite appreciate the point of a xeon chip. They are designed to be the pinnacle of stability for mission critical tasks, overclocking is typically not a great idea for stability as it can be unstable.

I know exactly what a XEON is for. I have several myself, eg. Dell T7500.

Of course overclocking can be unstable, but Intel's chips do it anyway, it's called Turbo. What I'm saying
is I know of pro users who would like the choice to experiment with this, but they can't. Stability
depends on so many things; as I'm sure you know, without a voltage increase, one can already do a lot
with most Intel CPUs. XEONs are used in plenty of non-mission-critical tasks aswell btw.

Ian.

 

[Someone Somewhere]

Maybe a quad-core E5 at 4.xGHz would do well.

 

[robertoivs]

"Evidently you haven't tested SAS cards, which can max out a PCIe2.0x8 slot, with fast SSDs."

Today you are unlucky, on my personal computer there are 8 HDDs in Raid 6 with a Promise PCI-E 8x SAS card. However as System Integrator I have worked with so many exotic hw that this is not so relevant.


"The poster should know that this is entirely wrong, and that bandwidth is shared across all cores. I guess all those that buy chips for servers with quad-channel RAM and 8 cores are stupid?"

The bandwith is not shared between cores. I can help to understand it. My target was not to attempt to ridicolize you but to try to improve the level of Tom's reviews. I will try to write you privately.


best regards
Roberto Castellano
Interactive Video Systems

 

[Mike Stewart]

still rocking out on the 2700K and at 4.8ghz it will destroy 3770K and a 4770K as well...
as for 4960E if that thing goes over 800$ , its pure fail

 

[dextermat]

After reading ivy bridge and Haswell review, I am really please my Sandy bridge is still one of the greatest chip so far.

To bad intel has notice it been too good and decided to put bad thermal paste and slash in the overclocking capability of cpu's .

I might Haswell keep it for a while! hahaha!

 

[Shawna593767]

Heck I just want an I5- 4xxxX with 6 non threaded cores around 300

 

[Someone Somewhere]

"Evidently you haven't tested SAS cards, which can max out a PCIe2.0x8 slot, with fast SSDs."

Today you are unlucky, on my personal computer there are 8 HDDs in Raid 6 with a Promise PCI-E 8x SAS card. However as System Integrator I have worked with so many exotic hw that this is not so relevant.


"The poster should know that this is entirely wrong, and that bandwidth is shared across all cores. I guess all those that buy chips for servers with quad-channel RAM and 8 cores are stupid?"

The bandwith is not shared between cores. I can help to understand it. My target was not to attempt to ridicolize you but to try to improve the level of Tom's reviews. I will try to write you privately.


best regards
Roberto Castellano
Interactive Video Systems

Yes, the bandwidth is shared between cores. There is no magic link that appears between one memory controller and a core. The core sends a request into a queue, which get accessed by a memory controller, which then puts the response into the ring bus, where it goes around to the L2 cache of that core. If a core sends multiple requests, they may be acted on by multiple memory controllers.

If not, you couldn't use all available memory bandwidth in single threaded tasks. Which you can.

 

[Shawna593767]

Heck I just want an I5- 4xxxX with 6 non threaded cores around 300

 

[cangelini]

Shouldn't that be Broadwell?

No, that's the Haswell refresh.

 

[cangelini]

No overclocking?
No de-liding?
A little disappointed. Guess it will come eventually.

Little point overclocking a pre-production CPU when Intel could turn around and go from solder to paste again, as with Ivy Bridge. Better to wait for boxed processors before we start looking at tuning the chip.