K8L?

[xtremejack]

K8L will not be up against Conroe. K8L will square off against Penryn, which is Conroe 45nm shrink with more cache and faster clocks. The fight will be tight.

 

[Viditor]

Originally posted by: xtremejack
K8L will not be up against Conroe. K8L will square off against Penryn, which is Conroe 45nm shrink with more cache and faster clocks. The fight will be tight.

At some point it will, but not at launch...K8L is scheduled for the end of H1 2007 (mid-year), while Penryn is scheduled for the end of 2007.
AMD's 45nm is scheduled for the middle of 2008 (6 months after Penryn)...

 

[Furen]

Yes, 45nm is scheduled for mid 2008, and 65nm was scheduled for late 2005/early 2006 =)

I'm a big skeptic when it comes to long-term roadmaps... especially AMD's.

 

[ahock]

I sincerely doubt AMD can will be able to release 45nm 18 months after their 65nm. I just dont know wheter they have that enough resources though we know IBM and AMD are working together on this. We haven't seen any 65nm working....

 

[Viditor]

Originally posted by: Furen
Yes, 45nm is scheduled for mid 2008, and 65nm was scheduled for late 2005/early 2006 =)

I'm a big skeptic when it comes to long-term roadmaps... especially AMD's.

Fair enough...of course keep in mind that the 65nm change involved a whole new Fab and a 300mm process as well. Fab 36 was built for 45nm already, so the changeover should be far less traumatic. What will be interesting will be the 32nm node coming out of the new NY Fab...

 

[Viditor]

Originally posted by: ahock
I sincerely doubt AMD can will be able to release 45nm 18 months after their 65nm. I just dont know wheter they have that enough resources though we know IBM and AMD are working together on this. We haven't seen any 65nm working....

Sure we have...we just haven't seen any reviews on it. AMD has been demonstrating their 65nm chips since last October.
Keep in mind that despite the reviews and projected performance of C2D, this takes a good deal of time to become actual money and marketshare (I estimate at least 6 months, and the Intel projected margins seem to back me up). A good portion of the reason for this is that Intel must ALSO ramp up new platforms for the chip. Even though it uses S775, the specs must be changed and adopted across 3rd party chipsets and motherboards (and how many people really enjoy having a vers. 1.0.0 motherboard?).
Another thing to keep in mind is that AMD demonstrated their own 45nm SRAM just 3 months after Intel did...
If it's a case of a simple shrink, I think even 12 months is quite possible...but there's the rub. It most likely will be significantly more than just a shrink.
1. That is the same time period that K10 should be coming out
2. We still don't know when/if ZRam-cached chips will be produced
3. I think that Torrenza is vastly underestimated at this point, and we will most likely see significant changes on that front in K10 as well

With all of that, I certainly can't disagree with the skepticism...however this management team has been pretty good at keeping their promises (the early 2006 prediction for 65nm Furen spoke of was based on rumours, not AMD announcements...), and 45nm by mid 2008 is one they made in their CC this quarter.

 

[Stoneburner]

viditor, what's with torrenza anyway? I'm a little confused, is torrenza just an alternative use for the second socket in the 4x4 platform? OR is there going to be an multi purpose socket which various different chips can plug into? Far as my limited knowledge tells me if you can plug in a desginated special purpose FP processor the platform will be incredibly efficient. Also, when is torrenza going to go live? I found very little info on wikipedia..


Thanks in advance. And since i thanked you in advance you should reward my kindness with juicy insider tidbits. At least a tease ?

 

[Viditor]

Originally posted by: Stoneburner
viditor, what's with torrenza anyway? I'm a little confused, is torrenza just an alternative use for the second socket in the 4x4 platform? OR is there going to be an multi purpose socket which various different chips can plug into? Far as my limited knowledge tells me if you can plug in a desginated special purpose FP processor the platform will be incredibly efficient. Also, when is torrenza going to go live? I found very little info on wikipedia..


Thanks in advance. And since i thanked you in advance you should reward my kindness with juicy insider tidbits. At least a tease ?

No insider stuff needed, just a careful scrutiny of the specs and language from AMD...
Torrenza is not just for 4x4, in fact that's just a small piece of it. It represents a new open platform of communication for a multitude of applications. One of the biggest consequences of this is that AMD has opened up their Coherent HT links for public use (before this it was only the standard HT links).
In a nutshell, the difference here is that standard HT is like any other bus model (though more efficient)...it's a way of sending data from and to a multitude of devices without the congestion of a standard bus.
cHT (coherent HT) allows you to access the cache of each core and also system memory directly. Just a brief glance will tell you that this greatly increases speed and capability (eg much lower latency, and the option of coprocessors with their own Ram slots).
The reason that making it a totally open platform is important can best be illustrated by the differences between the Mac and the PC...
1. The Mac is a Closed Platform, which means that anyone who develops anything for a Mac must first get Mac's approval before going to market. This means testing, money, and most importantly time-to-market penalties. On the flip side of that, Mac is able to personally guarantee that it works...
2. The PC is an Open Platform, which means that anyone can develop for it, and the market itself is the final decision maker as to suitability. The advantage here is that R&D is spread over a huge number of companies and innovation is MUCH quicker and cheaper, the down-side is that there are sometimes some really poor developments so you have to pay attention to what you're buying... (WHQL is a weak attempt to alleviate some of this pain, but note how long it takes for a driver to become WHQL certified if at all)

While most people have been talking about things like GPUs and Physics coprocessors (both good apps for this), the number of possibilities is really endless...
For example (I'm borrowing this from another board), what about a co-processor that can try to optimize which textures go to which GPU, and more important, anticipate which textures and other data will be needed next. For example, one fairly common optimization in games is to have two or more sets of textures for a creature. One set is small, for when perspective puts the creature in the background, and the other set is used when the creature is closer. The drivers can start loading the large textures ahead of time if the bandwidth is available. This helps to eliminate deadly lag

 

[Stoneburner]

But who will be making these subprocessors? Coprocessors?

ANd um.. let's say you have one general processor and several special processors.... isn't that the cell model? ANd basically this allows a gamer to plug in certain coprocessors, servers plug in others, etc? When can we see this in action?


Thanks for all the info. You are god.

 

[Viditor]

Originally posted by: Stoneburner
But who will be making these subprocessors? Coprocessors?

ANd um.. let's say you have one general processor and several special processors.... isn't that the cell model? ANd basically this allows a gamer to plug in certain coprocessors, servers plug in others, etc? When can we see this in action?


Thanks for all the info. You are god.

They will be made by a whole BUNCH of people, that's the whole point of an open platform. Nvidia, ATI, Ageia, IBM, Sun, etc... plus a few hundred others that may not even be in existence yet.
A major difference between the Cell model and Torrenza is that Cell isn't really x86 compatible (there are many other differences as well, but that's a big one!).
I expect to see Torrenza in action by the end of this year...
For example (again, borrowed) Cray has been selling XD-1 systems with FPGAs, but they did not share caches. They are now selling the DRC co-processors for XT-3 (Red Storm) systems. Sun sold a system to TokyoTech with ClearSpeed accelerator boards, and without using them it is at #7 on the Top 500 list. It will be interesting to see what happens in November. So the hardware is there, and the DRC accelerators that plug into Socket 940 are really a second generation product...
In other words, demand is VERY high for this at the server level...as to 4x4 we shall have to see (my guess is that it will happen in Q1 07).

Edit: BTW, this also explains the addition of more HT links and greatly increased HT speed in K8L...

 

[Stoneburner]

So what has the performance gain been in the examples mentioned? I'm quite baffled by what intel is doing, if AMD pulls this off they will seemingly have a huge advantage that 32 general purpose cores wont be able to negate easily. An article I read on cell indicated that such a model with specialized processors and a general processor would be far more impressive than merely adding more general processors.

 

[Viditor]

Originally posted by: Stoneburner
So what has the performance gain been in the examples mentioned? I'm quite baffled by what intel is doing, if AMD pulls this off they will seemingly have a huge advantage that 32 general purpose cores wont be able to negate easily. An article I read on cell indicated that such a model with specialized processors and a general processor would be far more impressive than merely adding more general processors.

With a specialized CoProcessor, most engineers are expecting as much as a tenfold increase in performance within the area that the coprocessor is designed for.
Remember that this a peak, not an average...
But yes, this will certainly hit Intel VERY hard over the next 1-3 years. I expect that they will have their own solution when CSI is released in 2009, but I see this as a development that is every bit as important as AMD64 was...probably more so.
Unfortunately, as their are no benchmarks yet (or even very few working systems), it isn't something that we can talk about with any certainty here. But I predict that towards the end of the year, we will see a HUGE amount of discussion here at AT about this...especially as developers start to show their products.

 

[BitByBit]

I'm a little doubtful as to whether K8L will really exceed Conroe's IPC.
I was under the impression that the changes in the FPU implemented in K8L will merely bring its FP performance on par with Conroe, i.e. the ability the execute two 128-bit SSE instructions, or four double precision instructions per cycle.
One of the biggest reasons for Conroe's performance is its instruction combining ability, which there so far has been no mention of in K8L. The effect of macro-op and micro-ops fusion effectively makes the core even wider, since it is able to execute more instructions per clock than it could without it.

K8L will certaily level the playing field through the inclusion of out-of-order loads and inproved branch prediction and prefetch, but I believe its biggest strength will be its AMD/IBM manufacturing process, which will almost certainly be superior to Intel's, meaning K8L may well be able to achieve the clock speed advantage it needs.

 

[AMDrulZ]

i belive K8L will be a 4 issue core and deffinatly be able to mach conroes IPC or even out perform it by 15% and will have a much stronger FPU than conroe... K8L will be able to outperform conroe hell the 3 year old K8 is only be outperformed by 15 to 20 percent and it is only a 3 issue core...

 

[CyborgNinja117]

K8L will beat conroe!!!! yes it will!!!!!!! i know it wil!!!!!1111oneoneoneelevenelevn. but it wont beat nelahem conroes succesor

 

[zsdersw]

(AMDrulZ and CyborgNinja117) ... Your matter-of-fact statements on the matter are warrantless for a couple reasons:

1. We don't know what Conroe/Kentsfield will be like when K8L is released.

2. No one has yet done any tests on K8L to give even a ballpark estimate on what its performance will be.

 

[Viditor]

Originally posted by: BitByBit
I'm a little doubtful as to whether K8L will really exceed Conroe's IPC.
I was under the impression that the changes in the FPU implemented in K8L will merely bring its FP performance on par with Conroe, i.e. the ability the execute two 128-bit SSE instructions, or four double precision instructions per cycle.
One of the biggest reasons for Conroe's performance is its instruction combining ability, which there so far has been no mention of in K8L. The effect of macro-op and micro-ops fusion effectively makes the core even wider, since it is able to execute more instructions per clock than it could without it.

K8L will certaily level the playing field through the inclusion of out-of-order loads and inproved branch prediction and prefetch, but I believe its biggest strength will be its AMD/IBM manufacturing process, which will almost certainly be superior to Intel's, meaning K8L may well be able to achieve the clock speed advantage it needs.

At the core level, you are quite correct...it should only be equal to C2D, not superior. However, you have to remeber that the K8L core will have HT and an on-die memory controller that C2D does not! This is what will make it a bit faster than C2D at the same clockspeed (my WAG is ~10-15% based on current advantages from these).

Edit: And Z is quite correct...this is all speculation at this point.

 

[zsdersw]

Originally posted by: Viditor
With a specialized CoProcessor, most engineers are expecting as much as a tenfold increase in performance within the area that the coprocessor is designed for.
Remember that this a peak, not an average...
But yes, this will certainly hit Intel VERY hard over the next 1-3 years. I expect that they will have their own solution when CSI is released in 2009, but I see this as a development that is every bit as important as AMD64 was...probably more so.
Unfortunately, as their are no benchmarks yet (or even very few working systems), it isn't something that we can talk about with any certainty here. But I predict that towards the end of the year, we will see a HUGE amount of discussion here at AT about this...especially as developers start to show their products.

So which is it? First you say it will "certainly" hit Intel "VERY hard" over the next 1-3 years... and then you say that this isn't something we can talk about with any certainty.

 

[zsdersw]

Originally posted by: Viditor
(my WAG is ~10-15% based on current advantages from these).

Current advantages from these? Versus what? Netburst CPUs?

The memory bandwidth available to K8 already vastly exceeds that of Conroe/Merom/Woodcrest, but it doesn't appear to be in use.. as Conroe/Merom/Woodcrest excel almost across the board. The ability to use more of it that K8L will undoubtedly present would merely bring it up to Conroe/Merom/Woodcrest's level. I bet that even K8L won't fully utilize all the memory bandwidth that's available to it.

 

[Viditor]

Originally posted by: zsdersw

Originally posted by: Viditor
With a specialized CoProcessor, most engineers are expecting as much as a tenfold increase in performance within the area that the coprocessor is designed for.
Remember that this a peak, not an average...
But yes, this will certainly hit Intel VERY hard over the next 1-3 years. I expect that they will have their own solution when CSI is released in 2009, but I see this as a development that is every bit as important as AMD64 was...probably more so.
Unfortunately, as their are no benchmarks yet (or even very few working systems), it isn't something that we can talk about with any certainty here. But I predict that towards the end of the year, we will see a HUGE amount of discussion here at AT about this...especially as developers start to show their products.

So which is it? First you say it will "certainly" hit Intel "VERY hard" over the next 1-3 years... and then you say that this isn't something we can talk about with any certainty.

Pick, pick, pick...
If I change it to "IMHO, this will certainly hit Intel VERY hard over the next 1-3 years", would you be mollified?

 

[zsdersw]

I'm touched that you're concerned about me in that way

 

[coldpower27]

Originally posted by: zsdersw
(AMDrulZ and CyborgNinja117) ... Your matter-of-fact statements on the matter are warrantless for a couple reasons:

1. We don't know what Conroe/Kentsfield will be like when K8L is released.

2. No one has yet done any tests on K8L to give even a ballpark estimate on what its performance will be.

If my detection algorithms are functioning properly, I believe CyborhNinja117 was trying to be funny.

 

[zsdersw]

Originally posted by: coldpower27
If my detection algorithms are functioning properly, I believe CyborhNinja117 was trying to be funny.

Perhaps.

 

[BlindBartimaeus]

I will say that I am a AMD fanboi however I am looking at Conroe but due to the higher price...even 6300's/6400's and the absolute rediculous price of the motherboards I think Intel is making a big mistake here. They are only going to have 25% of their line being Conroe by the end of the year (their statement not my speculation) so the netburst crap will be given away and price/performance guys will more than likely be scarfing up AM2 dual cores. This will mean that Intel will still sell that 25% but if you are only going for that little then why not make them MORE expensive since the price per core for Intel is going to be much higher (JMO) than AMD even considering the 65nm 300mm process. Intel has SOOO much overhead and Fabs producing Netburst. They almost have to be losing money on their low end stuff there ESPECIALLY the 8XX dual cores. That being said this price war will benefit US ultimately and I think it will hurt Intel more than AMD. Also consider the initial K8L numbers floating around and the 4X4 platform may be real cool. Just my opinion is that both companies will feel the crunch but I think Intel will be hurt more financially and they are playing into AMD's strong suit. Granted this is all speculation till it appears but AMD said K8L will start floating END OF THIS YEAR...more mainstream early next year.

Flame suit on...come get me!

 

[Stoneburner]

It seems that while intel is now ahead in the performance game, and will at least be competitive for a while, AMD plans to change the rules of the game entirely. IM just curious to see if they can pull it off, they did with x86-64 but this is a different beast entirely.