Anand Sandy Bridge performance preview is up

[Cogman]

I think you're asking for too much. Intel has been making x86 processors for over 30 years now, so you're really not going to have many drastic changes at this point.

In that context I do think that graphics on die is a fairly large step, and they also added 10% improvement per clock.

I agree, the next big leap will likely come from the death of the x86. Heck, even if they removed legacy instructions like enter and leave, I bet we would see some significant power savings from the reduced complexity.

 

[IntelUser2000]

From what I hear, Haswell's fundamental architecture will still look like Core uArch.

It's a fine line to determine what's a new architecture and what's not. I've read some comments in other forums where the people think Bulldozer took a regular Phenom II, deleted the extra ALU, and doubled it.

How should they determine what a new architecture is? Pentium III added absolutely nothing with Katmai, just the SSE extensions yet received a whole new brand. Prescott got a substantial architectural overhaul but was labeled as a "extended Pentium 4".

With Core i7 we might say Intel moved calling the "CPU" the architecture to a "Platform". They have been pushing the platform for quite a while now. While all of us might have liked consistent boosts, Nehalem did bring big benefits to everyone. Although I'll say Sandy Bridge will probably have more "core" changes than Nehalem did.

I guess a brand-spanking new architecture is exciting to geeks like us. Pentium 4 was exciting, and so was Rock. See where I am trying to go here?

About the post earlier by someone saying "Intel doesn't reveal enough architectural details unlike AMD". Finding out how Bulldozer performs with its brand new architecture with all the details is no easier than finding out how Sandy Bridge performs with all its secrets, but with a more conventional circuitry.

 

[Idontcare]

There's only so much you can get done in 2 years. It's a little saddening to see people sum up Core i7 as HT + IMC and that's it.

My apologies, sincerely, I was intentionally being a little over-dramatic there just to press the point...having been on that other side of the fence I assure you I have no doubt that the tireless 10-12hr days logged in week after week and year after year leading up to the creation of Nehalem (and Sandy) were in fact required and invested for good reason and the products would have suffered if such efforts weren't undertaken.

Please understand my comments regarding tick-tock are made entirely at the superficial level...just as the marketing aspirations of labeling a tick is a rather superficial glossing-over of the efforts and sacrifices made by the engineers working on the tick.

I merely aim to say that from a consumer perspective, which is about all I have been reduced to being, this whole "every tock is a new architecture" marketing dogma has really failed to deliver on the impressions of what a new architecture implies to me based on the history of new architectures from Intel.

There was little question whether Pentium was a new architecture in comparison to the 486. Likewise with Itanium. But Nehalem? Nehalem is phenomenal, I've made posts assessing the data and extolling my praise as such, but was it really a tock - a new architecture? Or was it core architecture extended and iterated once again?

Its all subjective, as you know, as any design engineer knows, the notion of new architecture is purely a creation of marketing. Any IC, be it designed for a new node or iterated on the existing node, is really a whole new beasty at the engineering level.

But I wasn't aiming my critique at the engineering level, I'm going for the superficial marketing level...give me a fricking tock already. Give me another Cedar Mill -> Core architecture transition.

 

[Nemesis 1]

I have to agree here. They stated every tock was going to be a "new" architecture...but the reality is we don't really get a tick-tock...we just get a tick-tick-tick-tick-tick progression with a node transition and some mild ISA extensions thrown in here and there.

This is the very reason I am typing this on a Q6600 desktop (that I use for a lot of computationally bound apps)...nearly 4yrs old but to "upgrade" to at best 2x the performance I'd have to drop a shizerload of cash.

From Intel we are getting drip-drop-drip-drop evolution instead of tick-tock revolution.

Well can ya tell me what AMD has done since hammer. Intel went from x% slower to x% faster. That was @ 65nm . Than at 45 intel brought forward 45nm high k and latter caught AMD on inter connects. On the High K die intel innovated here were as they seen problems with gate first alignment because of the annealing process so they innovated and went gate last. AMD didn't and will lose ground this round on transistor size and widing the time gap . They really all ready have as the 32 nm AMD with be a half breed seems they be using 45nm bits and piecies. Thats is a hybreed. Talk about drip drop drip drop . AMDs faucet isn't leaking so as to be noticedable. What is the IPC differance between hammer and PHII. What The IPC dlfferance from P4 C to Nehalem. Someone is dripping and dropping alot more than the other . I would also remind you it was intel who brought the High AMD pricies back to earth . NOT AMD AMD had no choice but to cut to the bone . AS many have said the lower hanging fruit is gone . AMD doesn't seem to have a later. But I look for ATI to pull there ass from the fire with fusion . Fusion I am liking but again Intel already showed there integarted fusion with sandy. SO ANand doesn't know if that was 6 or 12 Eu. I would say that the glass is half full still. Memory speed in that test were low. the 4 core 8 threaded i72600K looks interesting . normally if a thing don't look right feel right and your knowing the people who did this are way smarter than we are than its likely we haven't seen everthing yet. I wounder what IGP will gain fromDDR3 2100 memory speed will do for performance or speeding the core cpu above 4ghz Are we going to get to O/C the GPU because that has my interest considering The last IGP was 65nm I believe. Now were at 32nm on hjgh K . So many questions so few answers. IDF is going to be the show. This is Intel thread I won't bother the server crowd in AMD threads its not my cup of tea. Servers that is,

 

[IntelUser2000]

some rant about the sounds a clock makes

Inside of every great engineer is a geeky kid wishing for his dreams to come true.

9/2/2010 IntelUser2000

 

[ilkhan]

BD 4 module 8 core 8 threads will be up against SB 8 core 16 thread than in 2011 1st qt Low end IVY will likely bitch slap amds top of the line on desktop.

I agree. If (and AFAIK AMD has confirmed this) AMD puts 2 module BD against 4 core Ivy they're going to get their heads handed to them. At that points its not core+super-HT against superior core+HT, but against *2*superior 22nm cores with HT. BD would have to be stupid fast to have a chance at that point.

I think you're asking for too much. Intel has been making x86 processors for over 30 years now, so you're really not going to have many drastic changes at this point.

Please understand my comments regarding tick-tock are made entirely at the superficial level...just as the marketing aspirations of labeling a tick is a rather superficial glossing-over of the efforts and sacrifices made by the engineers working on the tick.

I merely aim to say that from a consumer perspective, which is about all I have been reduced to being, this whole "every tock is a new architecture" marketing dogma has really failed to deliver on the impressions of what a new architecture implies to me based on the history of new architectures from Intel.

...extended and iterated once again?

...Give me another Cedar Mill -> Core architecture transition.

Tick/tock IS a marketing tool. As with Khon, I think you may be being a little bit too harsh here. The cedar mill -> core jump happened only because P4 was so damn inefficient. It helped intel at that point simply because it made core look superhuman, whether it was really was or just was where Intel should have been.

IMO, the tock thing really is about pushing the limits of a process, once the tick has given them a little bit of time to get real world results. As you've said before, a new process is already risky enough, no need to make major changes in addition to that. The tock is their chance to showcase what they can do. Extended and iterated is exactly the goal. Another thing, to me, is that tick/tock seems to be regular, *small* changes. Let AMD have bigger improvements less often, tick/tock lets Intel have a steady stream of improvements. Sure, going from Conroe to Ivy Bridge would be a big improvement, but would you really want to wait that long (AKA the wait for vista, the chevy volt, or prescott)? Halo products announced too early is a great way to create buzz when nobody can buy the goods.

 

[Edrick]

I agree, the next big leap will likely come from the death of the x86. Heck, even if they removed legacy instructions like enter and leave, I bet we would see some significant power savings from the reduced complexity.

The problem is that the x86 killed off all competition over the years. Not due to better design or better benchmarcks, but due to the monster called Intel. (Alpha, MIPS, PowerPC, etc.)

Hell, even Intel x86 is killing off their own chips (Itanium) that has a different architecture. I had high hopes for this cpu back in the early 2000's.

I think we are stuck with x86 for decades to come. Thanks to Intel and Microsoft. So I will be happy with 10%-15% performance increases every year or so.

And yes, I would call Nehalem a tock. L3 cache, IMC, intergrated PCIe (Lynnfield), HT, QPI, branch predictor upgrade, etc.

 

[Nemesis 1]

if you say so. I'll wait for the testing.it's also a shame we are stuck with poor performing os's that can't capitolize on the treading.I still see amd coming out swinging but you can think what you want.

I can't I agree more ON cpu AMD is coming out swinging and its going to be 1-2-3 strike your out. Were is all this performance coming from 2 weak cores in a module . AMD lengthening the pipe line out to P4C I like that 18 pipes I believe. Should allow AMD to get to 6 ghz on air LOL. AMD saving grace this round will be fusion. and I like the sounds of bobcat very much/ Hay Intel wanted to get away from x86 long ago but AMD and MS worked to gether to bloody intel. SO thank AMD for the fact x86 is still around.

 

[Idontcare]

Well can ya tell me what AMD has done since hammer. Intel went from x% slower to x% faster. That was @ 65nm . Than at 45 intel brought forward 45nm high k and latter caught AMD on inter connects. On the High K die intel innovated here were as they seen problems with gate first alignment because of the annealing process so they innovated and went gate last. AMD didn't and will lose ground this round on transistor size and widing the time gap . They really all ready have as the 32 nm AMD with be a half breed seems they be using 45nm bits and piecies. Thats is a hybreed. Talk about drip drop drip drop . AMDs faucet isn't leaking so as to be noticedable. What is the IPC differance between hammer and PHII. What The IPC dlfferance from P4 C to Nehalem. Someone is dripping and dropping alot more than the other . I would also remind you it was intel who brought the High AMD pricies back to earth . NOT AMD AMD had no choice but to cut to the bone . AS many have said the lower hanging fruit is gone . AMD doesn't seem to have a later. But I look for ATI to pull there ass from the fire with fusion . Fusion I am liking but again Intel already showed there integarted fusion with sandy. SO ANand doesn't know if that was 6 or 12 Eu. I would say that the glass is half full still. Memory speed in that test were low. the 4 core 8 threaded i72600K looks interesting . normally if a thing don't look right feel right and your knowing the people who did this are way smarter than we are than its likely we haven't seen everthing yet. I wounder what IGP will gain fromDDR3 2100 memory speed will do for performance or speeding the core cpu above 4ghz Are we going to get to O/C the GPU because that has my interest considering The last IGP was 65nm I believe. Now were at 32nm on hjgh K . So many questions so few answers. IDF is going to be the show. This is Intel thread I won't bother the server crowd in AMD threads its not my cup of tea. Servers that is,

Well since you inquired, and I seem to be on a roll today endearing myself to all manner of industry professionals...I personally view the entire train of K7/K8/K10 skus out of AMD from the K7 Athlon up through Thuban Phenom II as nothing more than a series of iterations to the original K7 architecture.

Bulldozer, from my uneducated ill-informed misconceived ignorant arrogant viewpoint, will be the first *new* architecture from AMD in over a decade (since the K7 was introduced).

That's not to say there hasn't been innovation nor to say that those iterations haven't delivered impressive (over the course of a decade) IPC improvements...just saying from a high-level architectural viewpoint I look at a Thuban and I see 6 super-tweaked and optimized K7 Athlon cores in there. When I look at a K6 I don't see an optimized K7. When I look at a Conroe core I don't see a tweaked P4. When I look at a Nehalem core I see a tweaked Penryn.

But I readily admit and acknowledge that my opinion, plus 50 cents, still leaves me short about $3 for getting a cup of coffee at Starbucks.

Inside of every great engineer is a geeky kid wishing for his dreams to come true.

9/2/2010 IntelUser2000

lol, xactly, true dat.

 

[IntelUser2000]

Before the Pentium Pro, the RISC parts had such a big advantage that they could put take out enough performance enhancing circuitry to make it price competitive, put it on an binary translator, and still be competitive performance wise. They did exactly that with the Alpha PC.

That is until the Pentium Pro, when its integer performance went up high enough to briefly eclipse the RISC chips running its code natively. Combined with intense competitive pressure from AMD, x86 dashed ever far away from RISC ever hoping to have a chance of securing the marketplace.

 

[Nemesis 1]

My apologies, sincerely, I was intentionally being a little over-dramatic there just to press the point...having been on that other side of the fence I assure you I have no doubt that the tireless 10-12hr days logged in week after week and year after year leading up to the creation of Nehalem (and Sandy) were in fact required and invested for good reason and the products would have suffered if such efforts weren't undertaken.

Please understand my comments regarding tick-tock are made entirely at the superficial level...just as the marketing aspirations of labeling a tick is a rather superficial glossing-over of the efforts and sacrifices made by the engineers working on the tick.

I merely aim to say that from a consumer perspective, which is about all I have been reduced to being, this whole "every tock is a new architecture" marketing dogma has really failed to deliver on the impressions of what a new architecture implies to me based on the history of new architectures from Intel.

There was little question whether Pentium was a new architecture in comparison to the 486. Likewise with Itanium. But Nehalem? Nehalem is phenomenal, I've made posts assessing the data and extolling my praise as such, but was it really a tock - a new architecture? Or was it core architecture extended and iterated once again?

Its all subjective, as you know, as any design engineer knows, the notion of new architecture is purely a creation of marketing. Any IC, be it designed for a new node or iterated on the existing node, is really a whole new beasty at the engineering level.

But I wasn't aiming my critique at the engineering level, I'm going for the superficial marketing level...give me a fricking tock already. Give me another Cedar Mill -> Core architecture transition.

Look at it like a brain new car its new only until ya drive it off the lot . Than try selling it for what ya paid for it . It hurts man. Cpu and gpus and the rest after a year really aren't that old . But whats intel to do. Stop producing for new buyers and upgrades. You don't have to upgrade all the time. Thats ego nothing more.

 

[Idontcare]

As with Khon, I think you may be being a little bit too harsh here. The cedar mill -> core jump happened only because P4 was so damn inefficient. It helped intel at that point simply because it made core look superhuman, whether it was really was or just was where Intel should have been.

I think if we are going to head down into the rabbit hole that is "IDC's meager opinion" then we ought to make the distinction (or at least agree that I wish for it to be known that I feel the distinction is merited) that I am merely/simply talking architecture...how many decoders, ALU's, where the cache is located, blah blah blah...and am not talking performance.

P4 willamette performance may have sucked the proverbial balls off of the preceding P3 performance-wise, but architecture-wise there was no mistaking a P4 for a P3.

Nehalem...if you were merely looking at the core layout and weren't made aware of the fact that they put a memory controller next to it on the same slab of silicon then you might be easily convinced you were looking at a tweaked penryn core and not a "new" architecture.

At the risk of sounding petty or shallow, I must cut to the quick and acknowledge that I prefer we not delve any deeper into this analysis because I know at the engineering level I couldn't be more wrong (as I stated in response above in my post to TuxDave). Don't go down this rabbit hole, it is full of rabbit crap.

 

[TuxDave]

Tick/tock IS a marketing tool.

Technically not 100% true. It does have some design implications. A tock is our one shot to dump as many new instructions/features and get all the performance we want out of it because a tock is when software is most receptive to recompile and to take advantage of all of it. If a new feature isn't worth using during a tock, no one is going to use it if we made it better on a tick because no one is going to recompile when the next tock is coming around with another huge list of enhancements that they'll have to recompile for.

IMO, what's a "IDC tock". I think IDC is looking for something that doesn't resemble its predecessor. So I guess you should call Atom a tock. And Larabee a tock. Both are drastic changes to a market segment. But it doesn't come for free and that's where the nice backup plan comes around. P4 --> Core was a tock but if Core failed there was Tejas. If Atom failed there would probably be some single core Westmere or something. I dunno. That's how I personally feel when asked about "where's the next massive revamp". We are adding tons of features but to the common man, it's too compilcated to explain (either that or it's "secret sauce").

 

[Cogman]

The problem is that the x86 killed off all competition over the years. Not due to better design or better benchmarcks, but due to the monster called Intel. (Alpha, MIPS, PowerPC, etc.)

Hell, even Intel x86 is killing off their own chips (Itanium) that has a different architecture. I had high hopes for this cpu back in the early 2000's.

I think we are stuck with x86 for decades to come. Thanks to Intel and Microsoft. So I will be happy with 10%-15% performance increases every year or so.

And yes, I would call Nehalem a tock. L3 cache, IMC, intergrated PCIe (Lynnfield), HT, QPI, branch predictor upgrade, etc.

sad but true. Intel, in some ways, became their own worst enemy with the x86 architecture. I have little doubt that they would love to drop loads of their legacy support (what engineer wouldn't?) but it has become "too big to fail" They can't even take out something that was nearly imminently made obsolete (MMX)

 

[evolucion8]

When I look at a Conroe core I don't see a tweaked P4. When I look at a Nehalem core I see a tweaked Penryn.

Conroe is basically a tweaked Dothan which is a tweaked Tualatin.

 

[Riek]

I agree. If (and AFAIK AMD has confirmed this) AMD puts 2 module BD against 4 core Ivy they're going to get their heads handed to them. At that points its not core+super-HT against superior core+HT, but against *2*superior 22nm cores with HT. BD would have to be stupid fast to have a chance at that point.

So thats why Amd top model will be 4module/8core while at the same time intel will only have 4core/8thread?. Amd top model 16 core will compete with the 8-10core from intel. So AFAIK AMd confirmed roadmap wise en they way they compare their cores to core+HT that they indeed will offer similar cores as intel threads.

we still have to see how superior the other cores will be... If they have a 20% advantage in perf/core they might still be anhilated in in very corescaling applications. (just like a 8core Deneb would be faster for multithreaded applications then a 4core i7/SB).

 

[ydnas7]

i just read on another forum the 2500K will be $15 more than a 2500, but the 2600k will be ~$200 more than a 2600??? i don't expect it will eventuate.

 

[ydnas7]

using some of Anand's results

this started as a quick comparison between the HT of Nehalem and Sandybridge, which on average is similar, but varies widely per benchmark.

then i compared value (assuming the same price at the 2400's introduction) between the 2400 and i5-760, the 2400 without turbo (or HT) is about 23% percent faster, but again it varies considerably

then i compared IPCC between the HT enabled 2400 sample and core i7-880, the increase is about 13% but again varies considerably.

it seems that SandyB has significantly changed under the hood, it may not appear to be novel, but i suspect this little baby needed as much design validation and testing as any new Intel architecture.

 

[Acanthus]

Conroe is basically a tweaked Dothan which is a tweaked Tualatin.

which is a tweaked coppermine, which is a tweaked katmai, which is a tweaked, dixon, tonga, deschutes, klamath...


The major architectural deviations happened with the Pentium 4 and Itanium.

The pentium 4 architecture was good if the clockspeeds would have scaled as predicted.

Itanium is good, but never saw market penetration for various reasons.

 

[ilkhan]

I like how, while both start at 2.3Ghz and 45W, the i7-2820QM has 8MB of cache, HT, 12EUs and turbos to 3.4Ghz, while the i5-2500T only has 6MB, no HT, maybe only 6 EUs and turbos to only 3.3Ghz.

Also the random naming scheme. The desktop parts get a new setup ("2"+100/400/500/600) and the mobiles just get a "2" put in front of the number. The dual core desktop parts also get 2100 / 2120. We thought the i3/i5/i7/"i9" setup was wonky. WTF.

Comparison chart with what we (I, at least) know about them. Still not sure on the EU stuff, so its not in here. We dont "know" pricing yet, thats approximate based on the previous chart. I'm guessing the 2c turbo for the quad-core mobile parts will be 2.6/2.7/2.8(2.9?)Ghz, which would match the duals for the same 1c turbo speed. Again, thats just a guess, and also not on the chart because of that.

As of tomorrow I'll (finally) have a westmere laptop, so I'm not as concerned with mobile parts this time around. My desktop needs upgrades, so I'm focusing on that this tock.

 

[khon]

I think your pricing is a bit off on a few of the models.

The prices for the i3 desktops were leaked, and they're $115 and $135 respectively. Furthermore the i5-2500K is going to be more than $210. It is 200MHz faster than the i5-2400, and has the unlocked multiplier, so I'm going to guess it will be more like $240.

 

[Idontcare]

Realistically speaking, if I wanted to buy a SB laptop how long do you think I will be waiting? Summer next year? Spring?

 

[ilkhan]

I think your pricing is a bit off on a few of the models.

The prices for the i3 desktops were leaked, and they're $115 and $135 respectively. Furthermore the i5-2500K is going to be more than $210. It is 200MHz faster than the i5-2400, and has the unlocked multiplier, so I'm going to guess it will be more like $240.

I'm going exclusively off of what newegg had for equiv pricing last week when the article came out, as that's when I looked them up. Pricing for 1k units is lower. Pricing for 660 and 655k is identical (Actually it's a spread of $2). 185 is the 650. On newegg the 650 is/was $180, the whole box (760, 660, 661, 655k) was $208-210.

Again, I was actually looking at retail, despite what the chart says. 115->125 and 135->150 sounds reasonable for retail profits. Intel has pretty well defined price points, they fluctuate through the year, but generally price cuts just mean everything goes down one step on the ladder.

Realistically speaking, if I wanted to buy a SB laptop how long do you think I will be waiting? Summer next year? Spring?

wait, this is IDC asking? lol.
Anand's roadmap shows a Q1 launch (LV and ULV showing up in Q2). Consumer models soon after release, days to a couple weeks after Intel lets them loose. Business models have a longer testing schedule, sometime in Q2 or the last days of Q1 would be my guess.

 

[Idontcare]

wait, this is IDC asking? lol.
Anand's roadmap shows a Q1 launch (LV and ULV showing up in Q2). Consumer models soon after release, days to a couple weeks after Intel lets them loose. Business models have a longer testing schedule, sometime in Q2 or the last days of Q1 would be my guess.

Intel internal vs external use to make a lot of sense, now its just sphagetti to me. Trying to map one to the other, completely lost. I blame lynnfield/clarkdale.

 

[Nemesis 1]

^ your playing right . Hell your missing the real fun here. The game thats afoot is when does Apple release its SB comparred to all others.