Expected Ivy Bridge performance

[billyb0b]

i'm very satisfied with my i7 975 and beefy asus sabertooth x58 mobo. dual 16x pci-e. plenty!

 

[Torquemada]

What's the consensus on expected Ivy Bridge performance? Any chance we'll get something worth upgrading a i7 2600 over? I keep hoping we will (if you see my graphics card post, I've got a line of friends I can easily sell "old" hardware to). I'm holding out hope, but my guess is, "probably not without some serious OC'ing".

Current bets seem to be, nothing serious in IPC or clockspeed, but lower temps/noise. And these are still bets. Good bets IMO though, many have noticed ringing ears that need not be, including Intel product planners not the least.

BTW. Down the thread there was mention of pipeline length. Which pipeline?

Amusingly, we're coming somewhat back to uniformity -- one of the things that SB did very right was simplify retirement by basing dispatch on instruction latency rather than type. [I'm trusting holy Agner Fog here.]

 

[CHADBOGA]

Being right for all the wrong reasons is the uninteresting result because it says nothing of your capability to be right in the future. :\

In Dec 2010 not even AMD's own people knew how bulldozer was going to perform in reality, claiming you somehow "knew" is just a lucky guess masquerading as a fact, don't fool yourself into thinking people are so easily duped or swayed :|

I would say that had a pretty good idea, less than one year out from what turned out to be yet another delayed launch in 2011.

 

[IntelUser2000]

Ivy Bridge: http://en.expreview.com/2012/02/19/ivy-bridge-core-i5-3570k-engineering-sample-test/21214.html
Sandy Bridge: http://www.xbitlabs.com/articles/cpu/display/sandy-bridge-microarchitecture_6.html#sect0

They are clocked differently but it can always be normalized to cycles.

L1/L2: Identical at 4 and 12 cycles for both chips
L3: 14 cycles on Ivy Bridge, 18 cycles on Sandy Bridge(may even be higher at 19-20 cycles)

According to this very benchmark: http://cdn.techiser.com/wp-content/uploads/2011/12/Intel-Core-i7-3770-Compared-To-2600.jpg, the claim was that Excel's gains were due to "faster" cache and "larger" cache.

We now know the cache is faster, but how is it larger? One thing that most people don't know is that the comparisons are iGPU vs iGPU comparisons. It's possible Sandy Bridge is doing fixed L3 cache allocation for the iGPU, and Ivy Bridge becomes dynamic. IVB's GPU also has its dedicated L3 cache, further lowering the need to use the L3 cache.

That also inflates the CPU gains. Dedicated GPU comparisons are going to show lower gains.

 

[aaksheytalwar]

IMO Ivy Bridge will be a worth it upgrade, just like Conroe to Wolfdale.

1. Going from P67/Z68 to Z77 and 2600k to 3770k will increase real time FPS/benchies by 5 to 15% approx in games, more like nearly 10% across the board for a constant core clock.

2. Almost certainly, Ivy will be a lot more overclockable, so 4.5 avg on SB would be like 5+ avg on Ivy. So expect an additional 10% performance boost there.

So you will get at least 20-25%+ performance boost real time after overclocking assuming both setups are decently overclocked.

And this difference will be even greater for setups which require PCIe 3.0 like 7900 CF etc, especially after 6+ months and on Windows 8.

Also, from personal experience, I can tell you that a better CPU makes the experience a lot lot lot smoother and overall better even if the FPS only translate to 10-20% extra. Even with nearly the same FPS, often a powerful CPU can result in a better gaming experience and I have noticed this from my P4 times, so noticed it with almost every CPU I had, including going from an oced i7 860 to stock 2600k, the stock 2600k definitely made the experience better despite marginal FPS improvements.

IMO, Ivy will be a must have if you want the best tech and are willing to pay reasonably well for it.

Even a 3570k will be better than a 2600k in most of the stuff I guess.

I will be buying a Ivy 3770k + MSI Z77 GD80/65 at launch, and this is when I have a 2600k @ 4.3 + P67A UD4 B3.

 

[bronxzv]

L3: 14 cycles on Ivy Bridge, 18 cycles on Sandy Bridge(may even be higher at 19-20 cycles)

What is your source for the 14 clocks L3 latency on Ivy Bridge ?, this looks like a rather incredible value

For Sandy Bridge Anand talk about "26 - 31 cycles" here http://www.anandtech.com/show/3922/intels-sandy-bridge-architecture-exposed/4
maybe AIDA is not very reliable for this test ?

 

[IntelUser2000]

What is your source for the 14 clocks L3 latency on Ivy Bridge ?, this looks like a rather incredible value

For Sandy Bridge Anand talk about "26 - 31 cycles" here http://www.anandtech.com/show/3922/intels-sandy-bridge-architecture-exposed/4
maybe AIDA is not very reliable for this test ?

Both can be explained. Latency calculations can be derived by diving the absolute time by inverse of frequency.

5.4ns/(1/3.4GHz) = 18 cycles

Aida also tests single threads. That's why the latency is lower than multi-thread tests as it adds ring hops. But I believe the final results will be 4-6 cycles lower too.

 

[bronxzv]

Both can be explained. Latency calculations can be derived by diving the absolute time by inverse of frequency.

sure but I can't find any Ivy Bridge L3 latency test at your links, what did I miss ?

That's why the latency is lower than multi-thread tests as it adds ring hops.

If the test use the whole LLC capacity it should endure extra hops for remote L3 slices, also for single thread code

 

[exar333]

What are you talking about? PCIE 3.0 might be less useful on a setup with 40 lanes but who in their right mind is going to "upgrade" from a 2011 to an IB rig. 1155 (16 lanes) has been shown to bottleneck top end cards on x8 x8 so anyone running SLI/crossfire with the top cards that come out this year will relish using a PCIE 3.0 setup. Likewise anyone running quad SLI with 580's or above will be bottlenecking their cards on 2011 so i'm sure they will be waiting for IBE to drop. Just because you don't need it the way you have your rig set up doesn't mean you can call it a gimmick.

I agree 100%. PCIE 3.0 is a HUGE plus for IB.

Folks who are using 2x top-end GPUs (GTX 580 or 7950/7970) will see decent performance gains compared to their SB rigs that are using x8 x8 setups today. For those using 3 or 4 GPUs, 1155 never made much sense anyway, so thats probably not an issue. For that much $$$, they are either still using X58 rigs with 980x or 990x or have moved to SB-E.

Socket 2011 with 40 3.0 lanes will have HUGE bandwidth down the road. Its a little disappointing that AMD will not be adopting PCIE 3.0 for a significant amount of time. IB puts Intel on-par with bandwidth for PCIE-3.0 devices (x8x8 3.0 vs. x16 x16 2.0). You could also argue that Intel really needed 3.0 support to get their mainstream lanes 'up to snuff' and AMD needs this less, and that would also be accurate.

 

[exar333]

What is your source for the 14 clocks L3 latency on Ivy Bridge ?, this looks like a rather incredible value

For Sandy Bridge Anand talk about "26 - 31 cycles" here http://www.anandtech.com/show/3922/intels-sandy-bridge-architecture-exposed/4
maybe AIDA is not very reliable for this test ?

Agreed. That is most impressive.

 

[bronxzv]

Agreed. That is most impressive.

probably an issue with AIDA, maybe it doesn't test the whole L3 and/or it's mislead by some L3 to L2 prefetcher because its access pattern isn't random enough ?

these values look more credible :
http://www.lostcircuits.com/mambo//...ask=view&id=98&Itemid=42&limit=1&limitstart=7

 

[IntelUser2000]

sure but I can't find any Ivy Bridge L3 latency test at your links, what did I miss ?


If the test use the whole LLC capacity it should endure extra hops for remote L3 slices, also for single thread code

Damnit! I pasted the wrong link.

http://wccftech.com/intel-22nm-core...erclocked-46ghz-11v-msi-z77agd65-motherboard/

3.1ns @ 4.6GHz

The Sandra resuts are too high but what's sure is that the relative values are lot lower. Sandy Bridge @ 5GHz only achieves 3.8ns latency.

http://imageshack.us/f/717/5ghz2100cas6.png/

 

[bronxzv]

Damnit! I pasted the wrong link.

http://wccftech.com/intel-22nm-core...erclocked-46ghz-11v-msi-z77agd65-motherboard/

3.1ns @ 4.6GHz

The Sandra resuts are too high but what's sure is that the relative values are lot lower. Sandy Bridge @ 5GHz only achieves 3.8ns latency.

http://imageshack.us/f/717/5ghz2100cas6.png/

Thank you much, very interesting

 

[Joseph F]

Damnit! I pasted the wrong link.

http://wccftech.com/intel-22nm-core...erclocked-46ghz-11v-msi-z77agd65-motherboard/

3.1ns @ 4.6GHz

The Sandra resuts are too high but what's sure is that the relative values are lot lower. Sandy Bridge @ 5GHz only achieves 3.8ns latency.

http://imageshack.us/f/717/5ghz2100cas6.png/

So clock-for-clock, the cache is about 30% faster?

 

[IntelUser2000]

Also, I am assuming the reason single threaded latency results are lower is because multi-thread accesses come with conflicts. When Intel went from Pentium M "Dothan" to Core Duo the L2 latency went from 10 to 14 cycles. The change in latency might be to accomodate for the extra core. Core Duo doubles L2 cache bandwidth when both cores are accessing it.

The Ivy Bridge 3770K latency results are consistent with ES results clocked at 2.2GHz. Will link later, gotta go for now!

Update: http://www.coolaler.com.tw/coolalercbb/ivybridge/2G/2.gif

5.7ns @ 2.4GHz = ~14 cycles

Update2:

So clock-for-clock, the cache is about 30% faster?

Yes, at least for single threaded accesses. Multi thread might show a little less difference but still 4-6 cycles faster.

 

[gramboh]

IMO Ivy Bridge will be a worth it upgrade, just like Conroe to Wolfdale.

1. Going from P67/Z68 to Z77 and 2600k to 3770k will increase real time FPS/benchies by 5 to 15% approx in games, more like nearly 10% across the board for a constant core clock.

My speculation is there will be zero perceivable impact on most game titles at 1080p/1200p max details with a single 7970, clock for clock between SB and IB.

2. Almost certainly, Ivy will be a lot more overclockable, so 4.5 avg on SB would be like 5+ avg on Ivy. So expect an additional 10% performance boost there.

So you will get at least 20-25%+ performance boost real time after overclocking assuming both setups are decently overclocked.

And this difference will be even greater for setups which require PCIe 3.0 like 7900 CF etc, especially after 6+ months and on Windows 8.

I dunno, some ES leaks have shown clocking to not be much better than SB, while others have shown 5ish GHz on 1.1-1.2v, so we will see. I'd be happy with 5GHz at stock volts.

Also, from personal experience, I can tell you that a better CPU makes the experience a lot lot lot smoother and overall better even if the FPS only translate to 10-20% extra. Even with nearly the same FPS, often a powerful CPU can result in a better gaming experience and I have noticed this from my P4 times, so noticed it with almost every CPU I had, including going from an oced i7 860 to stock 2600k, the stock 2600k definitely made the experience better despite marginal FPS improvements.

IMO, Ivy will be a must have if you want the best tech and are willing to pay reasonably well for it.

Even a 3570k will be better than a 2600k in most of the stuff I guess.

I will be buying a Ivy 3770k + MSI Z77 GD80/65 at launch, and this is when I have a 2600k @ 4.3 + P67A UD4 B3.

Personally, I don't think there will be a perceivable improvement in games going 2600k @ 4.3 to 3770k @ 5 with the same GPU setup, but if you enjoy the hardware and can afford it, why not.

 

[rgallant]

http://tof.canardpc.com/view/bb759213-85ff-4999-85f3-714496c20213.jpg
here's one ib @ 5ghz -1.27v
-they could blow up at 1.3v , but who knows for the tri gates.

-part of post 140
http://www.xtremesystems.org/forums/showthread.php?277242-Early-Ivy-Bridge-Sample-Tested/page6

 

[gmaster456]

Any word on what Ivy Bridge offerings there will be on the server side of things?

 

[Anarchist420]

So IB doesn't have AVX2?

I bet once intel is so far ahead of AMD/nv on process nodes (or whatever the jargon is), they could beat them just by using the newest AVX to emulate ROPs and adding texture addressing/ filtering units.

 

[aaksheytalwar]

Acc to people even a i7 860 at 3.4 to 2600k stock with a single 6970 isn't an upgrade, and although fps inc only a bit, the real world experience in terms of smoothness and snappiness is much better, definitely worth an upgrade. I have done it so I know what i am talking

Similarly

A64 2.4 ghz single core
1900xtx
1280x1024
Gave me average 40-50+ fps in NSF mw
But no smoothness, stutters, crap experience

Swapped a e4500 and fps only went up by 10-15 above what was supposed to anyway be very playable, but smoothness came in even with similar fps
So huge upgrade

Just an example

Ivy will be a big upgrade, no matter what anyone says

Not as much as core 2 duo, but perhaps almost as big as 9xx to 2xxx

 

[IntelUser2000]

http://translate.google.ca/translat...?q=heise.de&hl=en&biw=1280&bih=920&prmd=imvns

-160mm2 die size, 1.4 billion transistors for the largest variant.
-Four seperate dies are to be available: Quad cores + GT1/Quad cores + GT2/Dual cores + GT1/Dual Cores + GT2. Currently, Sandy Bridge uses same dies for HD 2000 and HD 3000
-DDR3 support extends to DDR3L with 1.35V.

 

[Dadofamunky]

Pipeline limitations on clockspeed are not a given.

If Intel did a straight shrink of Sandy Bridge (not dummy shrink, but a reasonable straight shrink) and dialed in the voltage such that the power consumption was normalized between Sandy Bridge and a hypothetical 22nm Sandy Bridge then I would expect the 22nm version to clock roughly 15-20% higher than the 32nm microarchitectural equivalent based on the pipeline-limited-clockspeed argument you are taking.

Yeah... like the man said. Let's actually take the time to look at what's actually being DONE by Intel instead of throwing crap against the wall, eh? Like I said, a 10% boost is a very conservative estimate.

I just hope my P67 board can take an IB chip. Don't want to do a board swap if I can help it.

I agree 100%. PCIE 3.0 is a HUGE plus for IB.

Folks who are using 2x top-end GPUs (GTX 580 or 7950/7970) will see decent performance gains compared to their SB rigs that are using x8 x8 setups today.

It'll be interesting to see just how much of a difference true two-way x16 makes. I would wager a lot of people will be surprised by how little diff it does make. I could be wrong.

 

[Joepublic2]

. delete me plz wrong thread

 

[AtenRa]

Acc to people even a i7 860 at 3.4 to 2600k stock with a single 6970 isn't an upgrade, and although fps inc only a bit, the real world experience in terms of smoothness and snappiness is much better, definitely worth an upgrade. I have done it so I know what i am talking

Similarly

A64 2.4 ghz single core
1900xtx
1280x1024
Gave me average 40-50+ fps in NSF mw
But no smoothness, stutters, crap experience

Swapped a e4500 and fps only went up by 10-15 above what was supposed to anyway be very playable, but smoothness came in even with similar fps
So huge upgrade

Just an example

Ivy will be a big upgrade, no matter what anyone says

Not as much as core 2 duo, but perhaps almost as big as 9xx to 2xxx

From 40fps to 55fps = 37,5% more
From 50fps to 65fps = 30% more

You will not see that performance from SB to IV on the CPU side in games.

You will not understand any difference (not even in numbers) from an SB at 4.5GHz to an Ivy at 5GHz in the majority of the games (1080p and above) with the same graphics card.

 

[aaksheytalwar]

I didn't say 40 to 55 FPS jump.

I said I used to get 40-55 FPS with an AMD single core and that went up to 50-65 FPS using E4500.

Even earlier I used to average around 50 FPS. Later I averaged around 60 FPS with the new CPU.

And you still didn't get the point.

Going from 50 FPS to 60 FPS wasn't what caused the smoothness, it was the newer CPU. With the newer processor, if I increased the settings to sub 50 FPS it was still smoother than the performance I got on the old CPU with the same FPS.

Of course, AMD A64 to C2D was a much larger jump.

But I am expecting a jump similar to going from a i7 860 @ 3.4 to 2600k @ 4 GHz, I did that myself and the difference is well worth $500-700+ in my country. It was a very significant performance jump, even to a stock 2600k it was worth it.

I hope I am clear now.