AMD vs Intel at the high end in the future

[bfdd]

How long has there been talk about Bulldozer now? We've seen nothing? Phenom II's look pretty good, but they're still lagging behind and they're still just a little to much for what they offer. I personally don't see anything changing in AMDs future in the CPU market. Their stake in the GPU market on the other hand is a different story.

 

[alyarb]

What more did you want from Phenom II? Do you think they should be cheaper than yorkfield and bloomfield? because they are. They haven't architecturally revolutionized yet. They've only now got K10/45m sorted out to be a good sale and they have a good platform to sell it on. They don't claim to have an i7 killer, and they don't sell any $300+ desktop CPUs.

 

[Cogman]

Im sorry, but multithreading just isn't going to be very important to the consumer after 8 cores. There are a few places where we can take advantage of multiple threads, but after about 4 it starts speed boosts start to get hazy.

For a cooperation with larger severs, yeah, more then 16 cores still makes sense. However, for the average consumer it really doesn't

 

[alyarb]

we're going to start to see chips similar to bloomfield that can dynamically sleep cores or overclock them to maximize performance in a scenario without violating TDP. On the most basic end of the usage spectrum, though, I think dual core chips from 1.6 to 3.6 GHz will be available for a long time. They do a fine job and sip current.

 

[ilkhan]

I just want to see 5Ghz in a quad (or more). Maybe that would keep up with supreme commander. Multiple AIs and big unit counts ensure that nothing current can do it. Even 4Ghz bloomfields bog down eventually. One of the few games thats seriously CPU limited, albeit not for FPS. Wonder if Gulftown with HT off and on water will be able to hit 5Ghz...I thought bloomfield would be sufficient, but its not enough. Gullftown + GT300 maybe.

 

[WaitingForNehalem]

I don't want more cores, I want clockspeed! I loved the P4 days when clockspeed was king. I'm hoping for a stock 10GHz plus single core.

 

[jandlecack]

Originally posted by: WaitingForNehalemI loved the P4 days when clockspeed was king.

Funny that.

I don't remember those.

 

[IntelUser2000]

Originally posted by: Kuzi

Originally posted by: lopri
Conroe -> Nehalem is by no means 'incremental'. And remember that you can do it only once.

If you count HT, TurboBoost and server performance, then yes Nehalem is a great jump.

But what I meant was single core IPC, without HT and TurboBoost, at the same clocks as Penryn. I think the difference on average was like 7% faster than Penryn. Correct me if I'm wrong, too lazy to check now

Using the data here:

http://www.xbitlabs.com/articl...y/intel-core-i7_8.html

and here:

http://www.xbitlabs.com/articl...-core-i7_10.html#sect0

Core i7 SMT and Turbo Boost off vs QX9770 per clock is 7.74% faster(I did not include synthetics started counting from Half Life 2 fps)

Core i7 SMT on vs off gave SMT 7.97% advantage

Total average advantage of Core i7 with HT on and Turbo Boost off vs QX9770: 16.34%

Turbo Boost adds 2.58%.

I don't know about Turbo Boost, but since SMT is another feature you can't have easily I don't know why people disregard it quickly.

Phenom was really a highly tweaked x2 (K8), and with L3 cache added and 4 cores instead of two. The x2 was basically an Athlon XP (K7) but with an integrated memory controller, 64bit support, and two cores. The Phenom was about 15% faster than the x2, which was a nice bump in performance, but it is still the same architecture with many tweaks and improvements. AMD has been using the same CPU architecture for over a decade, it is time for a change

Of course some people call Core 2 a Pentium II. Phenom brought wider instruction fetch, 128 bit SSE and other improvements. It looks almost as big as Core 2 imo. Does it really matter whether the improvements are "incremental" or not as long as it brings performance?? Talking about radical changes, Pentium 4 was radical...

True Intel did it, but not all in one jump, they slowly added performance, Core->Core 2->Nehalem etc at a time when AMD was sleeping. I agree about the resources part and that's a big difference there between Intel and AMD.

The biggest thing there was Intel wasn't sitting down waiting to be pounded like last time. Say if Nehalem turned out to be second coming of Prescott and brought 0% performance, hotter and was late. The story would be different.

 

[JackyP]

When's AMD supposed to launch 32nm? There's a lot of talk of AMD 'closing the gap' or Intel losing their process tech advantage, but is there any official or reliable data to substantiate such rumours?
Does anyone know what we can expect from westmere IPC and feature wise? (there is that one leak showing good ol' SuperPI numbers) I'd expect turbo-mode to yield some nice gains in single-threaded applications if it is correctly implemented, or not?

ilkhan, Ivy brdige is a derivative of sandy bridge? Intel could go to 16C with a 32nm neha-shrink to replace Beckton. I'm sure it's feasible but is it also viable & a good strategy?

 

[Nemesis 1]

Well I can tell ya this. I said it long ago. The tooling AMD bought for 45nm scales to 32nm . with little cost. How soon AMD can implament is determined by Scaling core down . idontcare can give specifics. So ya they can catch up. Intel has the same deal tho going from 32nm to 22nm. So if AMD can get out early there in good shape if they get out late . Its pretty much game over.

I don't know about you . But I am impressed with AMDS 45nm. I do like bulk better tho.

Any way Your question about about 32nm beckton . If its as a believe that amds success at 45nm is mostly dew to the fact that Emersion process is the reason for this success. I believe Intels 32nm will knock socks off. My brother inlaw was telling me the overclocks there getting on latest batch of 2 core nehalem 32nm. I won't even say what air O/Cswere but these new numbers he gave me are crazy. Now the last time he gave me numbers this off the wall. It was when he at ATI. and he gave me aline of complete BS. I was pissed and hurt. His wife swore he never do again. But I not saying anything . I like Intel but these things can't do what he says . If they can I get rid of my 4 core and trade down for fact.

 

[Idontcare]

Originally posted by: JackyP
When's AMD supposed to launch 32nm? There's a lot of talk of AMD 'closing the gap' or Intel losing their process tech advantage, but is there any official or reliable data to substantiate such rumours?
Does anyone know what we can expect from westmere IPC and feature wise? (there is that one leak showing good ol' SuperPI numbers) I'd expect turbo-mode to yield some nice gains in single-threaded applications if it is correctly implemented, or not?

ilkhan, Ivy brdige is a derivative of sandy bridge? Intel could go to 16C with a 32nm neha-shrink to replace Beckton. I'm sure it's feasible but is it also viable & a good strategy?

Hi Jacky there are links out there to official AMD slides regarding GLobalFoundries 32nm timeline as well as AMD's Bulldozer timelines, I can dig them up if you really are keen to see them but suffice to say the earliest GlobalFoundries has said their 32nm foundry processes will be ready for HVM is Q1 2010 for bulk-Si 32nm and Q2-Q3 2010 for 32nm SOI high-performance...and AMD has said Bulldozer would be a 2H 2010 product at the earliest, basically a year (give or take a couple months) behind Intel's schedule westmere schedule and expected to slightly precede the Sandy Bridge release.

(similar to how AMD's 45nm release came out just prior to Intel's i7 release)

Intel has been extremely tight-lipped regarding Westmere performance...which is no different than how they handled Penryn and Nehalem timelines. Expect to see Anand releasing his "preview" articles no sooner than June, likely August.

 

[Kuzi]

Originally posted by: JackyP
When's AMD supposed to launch 32nm? There's a lot of talk of AMD 'closing the gap' or Intel losing their process tech advantage, but is there any official or reliable data to substantiate such rumours?

I think GlobalFoundries (AMD) can get bulk 32nm (for GPUs) ready in a year from now. But for high performance 32nm parts for CPUs, I dont think it can be before Q3 2010.

Does anyone know what we can expect from westmere IPC and feature wise? (there is that one leak showing good ol' SuperPI numbers) I'd expect turbo-mode to yield some nice gains in single-threaded applications if it is correctly implemented, or not?

Westmere is most likely just an i7 with higher clocks and more L3 cache.


PS: I saw the post IDC made after posting this.

 

[Idontcare]

Originally posted by: Kuzi
PS: I saw the post IDC made after posting this.

It's cool, we basically are in agreement with our recollection of the timelines which means that while neither of us gave links to support our claims we are both probably drawing on memory referencing the same slides out there on the web.

I wish I could remember which website was hosting them so I could link them for JackyP...drawing a total mindblank at the moment.

 

[JackyP]

If we know how well westmere fares we might be able to better extrapolate what AMD will be up against, as - correct me if I'm wrong - sandy bridge will be a heavily tweaked westmere chip.

Originally posted by: Kuzi
Westmere is most likely just an i7 with higher clocks and more L3 cache.

I'm expecting them to deliver the 5-10% IPC increase that we came to expect from Intel die shrinks, but I'm fascinated by the possibilities of turbo-mode @32nm.

Clarksdale (4MB) @2.4GHz does SuperPI 1M in 18.125s. How well does this compare to Nehalem and how poorly will superPI performance translate to other single-threaded code? (AFAIK superPI is making use of some pretty strange and outdated code; I'm assuming superPI is always leaked first because it's a useless app)

 

[soonerproud]

Originally posted by: Idontcare

Originally posted by: Kuzi
PS: I saw the post IDC made after posting this.

It's cool, we basically are in agreement with our recollection of the timelines which means that while neither of us gave links to support our claims we are both probably drawing on memory referencing the same slides out there on the web.

I wish I could remember which website was hosting them so I could link them for JackyP...drawing a total mindblank at the moment.

I'm certain JackyP could use google if he really wanted to verify if you are correct. Your time-line sounds about right from what I remember reading on the slides.

 

[Kuzi]

Originally posted by: JackyP

Originally posted by: Kuzi
Westmere is most likely just an i7 with higher clocks and more L3 cache.

I'm expecting them to deliver the 5-10% IPC increase that we came to expect from Intel die shrinks, but I'm fascinated by the possibilities of turbo-mode @32nm.

Probably a 5% IPC increase for Westmere over Nehalem, mainly from a larger L3 cache and some minor tweaks (if any).

But if as IDC mentioned Intel can get Westmere @ 4GHz and higher, then that's where most of the gain comes from. Though I don't see Intel releasing Westmere at such high speeds, at least not initially.

 

[Idontcare]

Originally posted by: Kuzi
But if as IDC mentioned Intel can get Westmere @ 4GHz and higher, then that's where most of the gain comes from. Though I don't see Intel releasing Westmere at such high speeds, at least not initially.

I agree, actual release clockspeeds will be determined solely based on gross margins for the SKU's.

If AMD does not have a product to put GM pressure on the low(er) clocked westmere SKU's then we'll see lower-clocked SKU's instead of higher-clocked ones, even if the underlying silicon can stretch its legs all the way to 5GHz.

Same thing we've seen since summer 2006 and 65nm.

 

[IntelUser2000]

Originally posted by: Kuzi

Originally posted by: JackyP

Originally posted by: Kuzi
Westmere is most likely just an i7 with higher clocks and more L3 cache.

I'm expecting them to deliver the 5-10% IPC increase that we came to expect from Intel die shrinks, but I'm fascinated by the possibilities of turbo-mode @32nm.

Probably a 5% IPC increase for Westmere over Nehalem, mainly from a larger L3 cache and some minor tweaks (if any).

But if as IDC mentioned Intel can get Westmere @ 4GHz and higher, then that's where most of the gain comes from. Though I don't see Intel releasing Westmere at such high speeds, at least not initially.

L3 cache sizes aren't increasing though, at least not per core. Gulftown will get to 12MB but it will also have 6 cores. Clarkdale/Arrandale will have 4MB which isn't different from Havendale/Auburndale. They'll have to rely on pure architectural changes.

Westmere-EX is 12 cores.

 

[Nemesis 1]

Isn't Westmere The old Nehalem C. If so I Think you maybe under estamating Intel 32nm process. As Nehalem was designed for the 32nm. process. Same as IVy is for 22nm. process.

This is going to be interesting. Some think Intel is playing games with AMD on AVX. It may be. Or it could be Intel is tring to save AMD from itself. Stopping AMD from following there lead.

Lets look at what we know.

Sandy is designed for 4-5 operands. Intel announced using 4 operand with FMA. and a fifth I believe used like in larrabee. As I understood it. Intel changed plan Why . TO screw with AMD . Not logical.

SO intel decides on 3 operand sandy. Well Intel is working on Larrabee. Did Intel find a performance Gain . Maybe . Could be lots of things. But I believe Larrabee is clue. If Larrabee can perform as advertized on X86. and do recompiles.

I see zero reason for intel to stay with X86 on IVy. Ivy will have Vector units on die like Larrabee. Ivy will have larrabee . So Ivy doesn't need to be x86. Larrabee is.

So if I am correct . If AMD followed Intel down thier path. When Intel switches to IVy. Amd be setting there holding empty bag. Because after intel sheds X86. AMD goes with X86. I doubt AMD is getting anypart of a new Intel none x86cpu.

Keep in mind AVX goes like this . Sandy AVX. Sandy 22 AVX. IVY no AVX instead its Vector units on die and the appearance of Intels FMA. 2 years of AVX only. I always said and ya know its true . I want to see AMD or Intel Do FMA on X86 and make it work correct . Amd is going to show us. Intel said we pass. I believe Intel has it right FMA on IVY no x86. Intel has its side kick for X86 and it must work fast or intel wouldn't have changed backends, on Ivy. I have Haswell and Ivy backwards . Haswell gets FMA and Vector unit.

 

[ilkhan]

Originally posted by: Nemesis 1
I see zero reason for intel to stay with X86 on IVy. Ivy will have Vector units on die like Larrabee. Ivy will have larrabee . So Ivy doesn't need to be x86. Larrabee is.

Did you really just suggest that a super wide (larrabee GPU) implementation of x86 could replace a fast and narrow x86 implementation (CPUs) at x86?

You didn't have much credability to me before, but that might just take the cake. CPU code isn't nearly parallel enough to work that way.

 

[Idontcare]

Originally posted by: IntelUser2000
L3 cache sizes aren't increasing though, at least not per core. Gulftown will get to 12MB but it will also have 6 cores. Clarkdale/Arrandale will have 4MB which isn't different from Havendale/Auburndale. They'll have to rely on pure architectural changes.

Westmere-EX is 12 cores.

L2$ increasese though, doesn't it? I thought they were going to 512KB L2$. Or was that just a rumor that never was to come true? I thought Anand was saying L2$ increase was coming...in one of those blogs where he talked about his views on the cache sizes of Nehalem.

 

[IntelUser2000]

It's actually not a good idea for the Nehalem generation CPUs if the L2 cache increases while L3 cache remains same per core. In Nehalem since it makes L3 caches inclusive it would effectively mean less L3 cache.

Nope, at least not in Westmere generation. Maybe they'll ramp up base clock speeds as AMD is planning to release faster clocked Phenom IIs as the year passes. See I think if they planned for larger per core caches we would have known it by now, at least a hint. Everything points out that contrary to conventional wisdom, no cache increases are happening, at least not per core.

 

[evolucion8]

Originally posted by: Kuzi

Originally posted by: ExarKun333
So do you call the shift from X2 to Phenom incremental as well? It was a whole new architecture.

Phenom was really a highly tweaked x2 (K8), and with L3 cache added and 4 cores instead of two. The x2 was basically an Athlon XP (K7) but with an integrated memory controller, 64bit support, and two cores.

The Phenom was about 15% faster than the x2, which was a nice bump in performance, but it is still the same architecture with many tweaks and improvements. AMD has been using the same CPU architecture for over a decade, it is time for a change

Intel has used the same P6 architecture from a decade too! Pentium 2, P3, Pentium M, core duo, core 2 duo and Nehalem uses the same architecture, but just improved, the Pentium 4 was a brand new architecture which was radical at it's time. AMD K8 was tweaked considerably compared to the K7, original Phenom was slighly tweaked, but Phenom 2 was more tweaked compared to the original Phenom.

Probably AMD will try to fit more cores in a CPU to remain competitive against Intel, since their cores are smaller, kinda like is ATi currently doing with their cards, lots of stream processors to keep up with the nVidia bigger ones but much fewer in amount

 

[IntelUser2000]

Originally posted by: evolucion8

Intel has used the same P6 architecture from a decade too! Pentium 2, P3, Pentium M, core duo, core 2 duo and Nehalem uses the same architecture, but just improved, the Pentium 4 was a brand new architecture which was radical at it's time. AMD K8 was tweaked considerably compared to the K7, original Phenom was slighly tweaked, but Phenom 2 was more tweaked compared to the original Phenom.

Core-wise K8 didn't have lot of tweaks. K7 to K8 is like Penryn to Nehalem. The major enhancements in those two are the memory subsystem. Up to Core Duo was a fairly similar core. Calling Core 2 Duo the same architecture pushes it a bit, as it goes borderline. In order of most "radical" enhancements I'd say the top 3 is: Pentium 4/Core 2/original Phenom

Phenom II had very small tweaks compared to Phenom. Phenom wasn't competitive thanks to low clock speeds but per clock advancements over the Athlon X2 was quite nice. Phenom to Phenom II is like Merom to Penryn.


Single threaded performance-wise, there are only two negatives for Nehalem compared to Penryn generation. They are both in caches. L1 cache size stayed same but the latency went from 3 cycles to 4 cycles. L2 cache latency decreased but the size decreased to 256KB.

I have a feeling that the increased L1 cache latency is having equal, or even greater part to holding back single thread performance on Nehalem as the smaller but faster L2 cache. L1 caches are tightly integrated into cores though, which might determine clock speeds of the architecture. Perhaps a 3 cycle L1 cache would have been a limiter in case Intel decided Nehalem needs to clock far faster to be competitive. The desire to lower power via static circuitry is probably the culprit. Now I think that will be paying off in the server market or the low power notebook chips. Probably not for desktops.

 

[Kuzi]

Originally posted by: IntelUser2000
It's actually not a good idea for the Nehalem generation CPUs if the L2 cache increases while L3 cache remains same per core. In Nehalem since it makes L3 caches inclusive it would effectively mean less L3 cache.

Nope, at least not in Westmere generation. Maybe they'll ramp up base clock speeds as AMD is planning to release faster clocked Phenom IIs as the year passes. See I think if they planned for larger per core caches we would have known it by now, at least a hint. Everything points out that contrary to conventional wisdom, no cache increases are happening, at least not per core.

It was just a guess on my part to say that the L3 cache on Westmere would increase. And about the L2 cache, at least for Nehalem, Intel found that a smaller but faster L2 cache would serve better.

Because Westmere is @ 32nm, they will have more room to add transistors, and as you mentioned, Gulftown gets 6 cores with 12MB L3 cache. Maybe that will be the right balance, while also giving Intel the ability to clock these parts higher than Nehalem (+3.6GHz).

But if that is the case, I don't see Westmere being any faster than Nehalem per core at the same clocks. Intel will save all the improvements for Sandy Brigde.