Was Sandy Bridge a let down?

[Purpose]

Gulftown was not just a process shrunk Nehalem. There were small architectural changes in GT that lead to the performance enhancement. One of the primary differences is the decoupling of the Uncore with relation to Memory speed. Nehalem required the Uncore to be 2x Memory speed, while GT only requires 1.5x. This allows for higher memory frequencies before being bound by the Uncore itself not being able to go faster.

That said, SB only has about an 11% IPC increase over GT. Not 20% as Nemesis above says(LOL @ delusional people). I'm pretty sure 11% is the figure given here at AT in the SB launch article. GT also beats the pants off of SB if you're doing something that can use all available threads.

All in all... Like any other cpu release... If you own a high end system with technology from just one generation prior... It makes no sense to upgrade. Anyone that tries to tell you that getting 66.6fps vs 60fps is a worthwhile upgrade, is absurd.

As others mentioned though, the increased IPC isn't the whole story. Where I think SB really shines is in the mobile space. Since the introduction of Lynnfield and SB, we've seen battery run times skyrocket. And you still get a very capable processor.

 

[exar333]

If you define a 'letdown' by releasing a productin around $200 that is comperable in most tasks to the previous $1000 flagship, uses WAY less power, overclocks much better, and has very affordable motherboards.

Just look around, have you seen one person on the forum with a 2500k in their signature?

 

[DaveSimmons]

Derp and trololol?

Better IPC, lower power consumption, decent onboard graphics that will even play some 3D games like Torchlight and Half-Life 2 (I've done it with HD2000 on an i3-2100).

The dual-cores are great general use and HTPC chips, the quad-cores are great for performance and gaming use.

Unlike AMD, every aspect I care about was improved with SB CPUs. All steps forward, no lurches back.

 

[LOL_Wut_Axel]

Gulftown was not just a process shrunk Nehalem. There were small architectural changes in GT that lead to the performance enhancement. One of the primary differences is the decoupling of the Uncore with relation to Memory speed. Nehalem required the Uncore to be 2x Memory speed, while GT only requires 1.5x. This allows for higher memory frequencies before being bound by the Uncore itself not being able to go faster.

That said, SB only has about an 11% IPC increase over GT. Not 20% as Nemesis above says(LOL @ delusional people). I'm pretty sure 11% is the figure given here at AT in the SB launch article. GT also beats the pants off of SB if you're doing something that can use all available threads.

All in all... Like any other cpu release... If you own a high end system with technology from just one generation prior... It makes no sense to upgrade. Anyone that tries to tell you that getting 66.6fps vs 60fps is a worthwhile upgrade, is absurd.

As others mentioned though, the increased IPC isn't the whole story. Where I think SB really shines is in the mobile space. Since the introduction of Lynnfield and SB, we've seen battery run times skyrocket. And you still get a very capable processor.

The difference between Bloomfield and Gulftown in IPC was actually very small. Clarkdale seems noticeably slower than both, though.

I agree on everything else.

 

[Ferzerp]

Agreed. I liked it better when they released high end first then trickle down to mainstream.

I wonder if it's a technical reason, or a business reason that they are going that route? I would guess technical, because from a business standpoint, I would imagine that many SB purchases would have been SB-E purchases if it were available. Now by the time SB-E is available, why not wait for IB?

 

[BrightCandle]

Every CPU release since the Pentium 4 should really be considered a letdown. The core 2 duo was the first set of chips to be forced to favour IPC improvements over clock speed thanks to the physics involved. Since then we have barely had a 50% improvement in performance over a 6 year period. In the Pentium/P2/P3/P4 days we could look forward to a doubling of performance in 2 years. All modern CPUs are severely limited improvements on the previous ones.

In the past a process shrink gave you an increase in clock speed, a decrease in power consumption, more transistors to play with and a small die making it cheaper. We are still getting the increase in transistors in a similar way to before but the other effects have decreased dramatically as the number of atoms in a transistor shrinks. You certainly don't get IPC for free as you shrink the process, it should just make the same CPU smaller. IPC comes from using the additional transistors to improve some aspect of the core which returns an increase in performance.

Cores don't equal performance unfortunately. We are already seeing basic scale problems with modern multithreaded software and we have barely begun increasing the numbers of cores. They are however a simple way to spend the ever increasing transistor budget, just like increasing the cache sizes. There is a limited amount of obvious ways to spend the transistor budget to get decent improvements in performance beyond more cores and the cache to compensate.

I doubt I'll ever be happy with a new CPU release again. The 10-15% we get these days is disappointing, but I also understand why it is we wont likely ever see fast progress again.

 

[sm625]

every aspect I care about was improved with SB CPUs. All steps forward, no lurches back.

That isnt true. You can buy a last generation i3 and overclock it to beyond i3-2100 level performance. In fact if you bought the very first i3-530 when it released you'd gain nothing and lose quite a bit by "upgrading" to an i3-2100.

This was even covered on anandtech:

http://www.anandtech.com/show/2921/5

Huge disappointment. I dont know why people are fawning all over intel for doing this.

 

[dawks]

Every CPU release since the Pentium 4 should really be considered a letdown. The core 2 duo was the first set of chips to be forced to favour IPC improvements over clock speed thanks to the physics involved. Since then we have barely had a 50% improvement in performance over a 6 year period. In the Pentium/P2/P3/P4 days we could look forward to a doubling of performance in 2 years. All modern CPUs are severely limited improvements on the previous ones.

In the past a process shrink gave you an increase in clock speed, a decrease in power consumption, more transistors to play with and a small die making it cheaper. We are still getting the increase in transistors in a similar way to before but the other effects have decreased dramatically as the number of atoms in a transistor shrinks. You certainly don't get IPC for free as you shrink the process, it should just make the same CPU smaller. IPC comes from using the additional transistors to improve some aspect of the core which returns an increase in performance.

Cores don't equal performance unfortunately. We are already seeing basic scale problems with modern multithreaded software and we have barely begun increasing the numbers of cores. They are however a simple way to spend the ever increasing transistor budget, just like increasing the cache sizes. There is a limited amount of obvious ways to spend the transistor budget to get decent improvements in performance beyond more cores and the cache to compensate.

I doubt I'll ever be happy with a new CPU release again. The 10-15% we get these days is disappointing, but I also understand why it is we wont likely ever see fast progress again.

</troll>

Wow this has to be one of the saddest and most delusional posts I have ever read on the forums.. Unbelievable. For anyone to think the netburst architecture was good at any level is ridiculous.

Netburst was Intels marketing betting on 'slowing down' a CPU in favor of hitting higher clock speeds. They figured they'd sell more chips having a computer on the shelf of CompUSA with a higher clock speed, and that mattered more than actual performance.

Conroe/Merom/Woodcrest were much more efficient processors. It seems like we've basically hit a wall in terms of clock speed, so the engineers are looking at other ways of driving efficiency. And to say we've "barely had a 50% improvement in performance over a 6 year period" is just silly. Moores law is still holding as far as I know.

Labelling someone a troll when they are clearly trolling is a matter of free speech, but I must've forgotten, this is not Anandtech, its STILL Banandtech. Sad really. OTOH, calling that poster a douche bag would be disrespectful, since I dont know if the poster is really a douche or not.. </edit> <back on topic>



Folks, there is a sticky regarding personal attacks and insults not being acceptable. Labeling a post or poster as a "troll" is just insulting, there's no reason to do it, period.

Please reconsider this demeanor. We can do better, as a community.

Administrator Idontcare

 

[DaveSimmons]

That isnt true. You can buy a last generation i3 and overclock it to beyond i3-2100 level performance. In fact if you bought the very first i3-530 when it released you'd gain nothing and lose quite a bit by "upgrading" to an i3-2100.

^ I don't overclock, so it is true for me. I agree it's a step back for people wanting to overclock the cheaper CPUs.

But even for OC intel made overclocking easier for the K chips, and all of the i5 and i7 CPUs now automatically and safely do Turbo Boost overclocking with guaranteed 100&#37; stability.

 

[sm625]

Turbo boost is a smack in the face after being able to extract 50&#37; more performance out of a $70 chip.

 

[Hulk]

Hmm let me see, my 2500k is about the same cost as my E6400 was and...

it's about 5 times faster video editing
half the power draw at idle
easier to overclock

No I don't think SB is a letdown!
The question is kind of silly actually. Sandy Bridge is actually quite a marvel.

 

[IntelUser2000]

There was no real IPC changes in Gulftown over Bloomfield except that the cache got bigger due to having 2 more cores. What did that do, 1-2&#37;? If it was a quad core chip there would have been zero difference.

The IPC increase is actually around 11%, so not really much, but the higher overclocking potential makes the difference 20-25% in total. Overall, not bad at all.

http://www.computerbase.de/artikel/...-sandy-bridge/47/#abschnitt_skalierungsrating

15-17% over i7 930 at same clock.

Bloomfield was a good chip but it was more oriented at adressing Core 2's weaknesses. Sandy Bridge on the other hand, has a very consistent gain over the predecessor.

Bloomfield vs. predecessor:
-Bloomfield vs Yorkfield at the same price point when it was introduced: http://www.anandtech.com/bench/Product/46?vs=49

Pay attention to low threaded scenarios and games

-Power consumption increased over Yorkfield
-Motherboards were typically at $250+
-3-channel setups back then were expensive

Sandy Bridge vs. predecessor:
-Sandy Bridge vs. Bloomfield at same price point(assume its i7 960 not 965)
http://www.anandtech.com/bench/Product/287?vs=45
-Power consumption decreased
-Motherboards at the mainstream point
-Only needed regular 2-channel setup

Look for example, iTunes, and and Lame, which are 1-thread applications largely insensitive to memory BW: http://www.tomshardware.com/reviews/processor-architecture-benchmark,2974-11.html

iTunes
-Gulftown: 4% faster than Penryn
-Sandy Bridge: 16% faster than Gulftown

Lame
-Gulftown: 2.4% faster than Penryn
-Sandy Bridge: 15.7% faster than Gulftown

Where Nehalem makes up to make it look good in are memory BW intensive apps.

 

[Rvenger]

I was an AMD fanboy (haha) until I bought my 2500k. This processor got me back into being interested in computers again.

 

[IntelUser2000]

And to say we've "barely had a 50&#37; improvement in performance over a 6 year period" is just silly. Moores law is still holding as far as I know.

Probably not enough to be called a Troll as no one is going to listen to what he says but I did some calculations based on Anand Bench.

Overall gain per core Sandy Bridge had over Conroe is in the 50-60% range. That's good as having 2/3 more cores just by architectural changes. And in few of the apps, the gains are good as getting 2x the cores. Not to mention we had 32% clock speed increase, and 2x more cores.

55% gain in performance per core
32% clock speed increase, say the gain is 25%
2x more cores = ~80%

In 5 years, the mainstream CPU has gotten 3.5x faster, for an annual average gain of 30%.

 

[sm625]

Probably not enough to be called a Troll as no one is going to listen to what he says but I did some calculations based on Anand Bench.

Overall gain per core Sandy Bridge had over Conroe is in the 50-60&#37; range. That's good as having 2/3 more cores just by architectural changes. And in few of the apps, the gains are good as getting 2x the cores. Not to mention we had 32% clock speed increase, and 2x more cores.

Are you including the overclock that can be done with ease on any ol cheap conroe? I seriously doubt the performance delta of an i3-2000 vs an overclocked E6320 or 6420 is more than 30%. It is really pretty goddawful for 6 years...

 

[IntelUser2000]

Are you including the overclock that can be done with ease on any ol cheap conroe? I seriously doubt the performance delta of an i3-2000 vs an overclocked E6320 or 6420 is more than 30&#37;. It is really pretty goddawful for 6 years...

That's a strawman argument because the OC limitation is an artificial barrier that will be improved in Ivy Bridge, and further done in Haswell(and apparently Intel admitted it was a mistake on the desktop processors).

Intel preparing i3-2120K:
http://www.nordichardware.com/news/...eparing-unlocked-core-i3-2120k-processor.html

 

[sm625]

Until that chip (i3-2120K) appears on ark.intel.com it is just rumor.

 

[exar333]

That's a strawman argument because the OC limitation is an artificial barrier that will be improved in Ivy Bridge, and further done in Haswell(and apparently Intel admitted it was a mistake on the desktop processors).

Intel preparing i3-2120K:
http://www.nordichardware.com/news/...eparing-unlocked-core-i3-2120k-processor.html

Can't wait for the MC deals for the 2120k! If you can pick this up for ~$100, it will be a steal.

 

[IntelUser2000]

Until that chip (i3-2120K) appears on ark.intel.com it is just rumor.

Rumor or not you are arguing on something that could easily change versus architecture, which isn't easily changed.

 

[podspi]

</troll>

Wow this has to be one of the saddest and most delusional posts I have ever read on the forums.. Unbelievable. For anyone to think the netburst architecture was good at any level is ridiculous.

Netburst was Intels marketing betting on 'slowing down' a CPU in favor of hitting higher clock speeds. They figured they'd sell more chips having a computer on the shelf of CompUSA with a higher clock speed, and that mattered more than actual performance.

Conroe/Merom/Woodcrest were much more efficient processors. It seems like we've basically hit a wall in terms of clock speed, so the engineers are looking at other ways of driving efficiency. And to say we've "barely had a 50% improvement in performance over a 6 year period" is just silly. Moores law is still holding as far as I know.

I have to disagree. I don't think the P4 was designed to fool people, I think Intel vastly over-estimated (somehow) the clock-speeds they were going to be able to hit. After the P4, was saw Intel go back towards more conservative "safer" designs because they didn't want another debacle like the P4, and I think that was probably a smart move.

AMD again tried a very risky design with BD, and we all know how that turned out. While AMD is going all out w/ marketing (fastest processor + moar cores), I believe again they designed it for an actual non-marketing reason. It just didn't work out... yet.

 

[Smartazz]

The IPC increase is actually around 11%, so not really much, but the higher overclocking potential makes the difference 20-25% in total. Overall, not bad at all.

But how much of that overclocking potential was added by the 32nm process? It'd be interesting to see how a quad core Gulftown 32nm overclocks compared to quad core Sandy Bridge.

 

[Joseph F]

The architecture change was significant in its focus on performance per watt.
At 80W, the 2600K @ 3.4GHz totally rocks the 130W i7 960 or 95W i7 880 @ 3.06GHz.

The i7 2600k is 95W, not 80W.

 

[Smartazz]

The i7 2600k is 95W, not 80W.

While you're right, you can't discount the fact that the 2600K has a GPU on board that is included in that 95W TDP. ~80W seems like a good estimate as to what the CPU-only TDP would have been without the integrated GPU.

 

[Smartazz]

Don't get me wrong, Sandy Bridge offered a great upgrade path and offers similar performance to Gulftown for a lot less money. However, enthusiasts never got to buy 32nm quad cores before Sandy Bridge. I think that Sandy Bridge would have been far less appealing if Intel didn't skip 32nm quad cores last generation.

 

[Arachnotronic]

80W.

http://www.newegg.com/Product/Produc...82E16819115081