backtop

Print E-mail del.icio.us 35 comment(s) - last by Viditor.. on Mar 15 at 2:46 AM

AMD's chief officer for marketing and sales, Henri Richard
Henri Richard talks about Intel, x86-64 and K8L

DigiTimes has posted the first of a five-part interview with AMD's chief officer for marketing and sales, Henri Richard. It appears that the interview was conducted earlier in the month before Conroe benchmarks started making the rounds given the nature of some of the responses. Here are just a few snippets from Part I.

Henri on Intel's processor development:

But frankly, they've made so many claims in the past – you know, the Netburst architecture was supposed to scale to 10GHz, and look at where we are today. Then their new-generation micro-architecture (NGMA), is, quite frankly, a quick fix on the front-side bus. I don't think that's the future of the Intel architecture. I think it's another quick fix until 2008 or later, when they're going to come out with a genuinely new architecture. So again, from a pure technology perspective, my assessment is that it's a lot of marketing – it's clever marketing, but it's not revolutionary. And calling it a new-generation microarchitecture is a little bit out of balance.

Henri on x86-64:

We've seen, at least in a couple of areas, Intel acknowledge that AMD was right. And it's important to note that finally they have x86-64 across their entire product range. And I'm happy about that because I think I can legitimately claim that if it wasn't for AMD, that wouldn't be the case...I think that that's an indirect acknowledgement of AMD's leadership – leadership in thought – because we clearly drove the industry to x86-64.


We interviewed Henri a few months ago, but with NGMA now announced (Core), Richard has a lot more to talk about. 

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
Intel is dead
By Phynaz on 3/13/06, Rating: 0
RE: Intel is dead
By AnaxagorasZeres on 3/13/2006 11:55:53 AM , Rating: 1
Care to back up that fanboyism with some fact? Conroe isn't some P4 with a band-aid, and from the looks of it, AMD doesn't have anything major up their sleeves. Seems like AMD will finally have to fight to keep the speed crown.


RE: Intel is dead
By Phynaz on 3/13/2006 12:02:34 PM , Rating: 1
Of course it's a P4 with a band aid!

They took an old P3 and stuck on the P4 bus.

AMD developed a new architecture.


RE: Intel is dead
By clnee55 on 3/13/2006 1:25:33 PM , Rating: 3
It's pathetic that a band-aided P3 can outperform a new architecture


RE: Intel is dead
By zsdersw on 3/13/2006 2:00:30 PM , Rating: 2
Obviously you know nothing about what Conroe/Merom/Woodcrest really is.


RE: Intel is dead
By Furen on 3/13/2006 4:11:46 PM , Rating: 2
I wouldn't say that Intel is dead but Merom is really a "band-aid" on the P6. It has deeper buffers, much better execution units, more capable decoders (and more of them), better branch prediction, more cache, etc. Of course all these relatively minor improvements (minor if taken individually) may lead to a huge performance increase when taken as a whole. Another thing that I should point out is that the K8 is pretty much a "band-aid" on the K7, and a much minor one since only the front-end was actually improved (and the memory controller was thrown into the die).

My point is that, basically, both "Core" and the P6 are structured the same way (conceptually, since they are physically different) while Netburst was structured very differently and, as such, could really be called a different architecture. Personally, I dont think the "Core" architecture is really all that different from the original P6 to warrant a new name but Intel kind of bridged this gap by calling Yonahs Core duo/solo.


RE: Intel is dead
By saratoga on 3/13/2006 4:34:17 PM , Rating: 2
[quote]I wouldn't say that Intel is dead but Merom is really a "band-aid" on the P6. It has deeper buffers, much better execution units, more capable decoders (and more of them), better branch prediction, more cache, etc. Of course all these relatively minor improvements (minor if taken individually) may lead to a huge performance increase when taken as a whole. Another thing that I should point out is that the K8 is pretty much a "band-aid" on the K7, and a much minor one since only the front-end was actually improved (and the memory controller was thrown into the die). [/quote]

They're similar in that they're both OOOE superscalar cores. The similarity stops their. Theres more in common between the P6 and K8.

[quote]My point is that, basically, both "Core" and the P6 are structured the same way (conceptually, since they are physically different) while Netburst was structured very differently and, as such, could really be called a different architecture.[/quote]

Well yes, they're both OOOE x86 cores. That does limit you somewhat. Other then that, I'm very curious how they're "structured the same way". Please be specific.


RE: Intel is dead
By Furen on 3/13/2006 6:00:38 PM , Rating: 2
Ok, let's look at the front-end first:

Take Yonah's front-end, add an extra decoder (for simple x86 instructions, the most common), macro op fusion and deeper buffers and you've got Merom's. The Reorder window is pretty much the same with, once again, deeper buffers.

Now, let's look at the execution units:

Take Yonah's, widen the Integer units to 64 bits (needed for 64 bit suppport) and add an extra shit/rotate ALU and you have Merom's. The FPU and SSE units were widened to 128 bits (AMD's are still 64 bits, by the way, though I think it was planning on widening them eventually) and tweaked (I think the FMUL unit is the one with the most improvements). This last thing is probably what will give the most significant performance improvement, since 128 bit SSE can now be executed in a single clock. Of course the fact that there's more units will also help

Also, the width of some of the internal buffers was widened, though the L2 cache retains the same 256-bit width (AMD's is still 128-bits but it was 64-bits in the K7).

So if you look at Merom's microarchitecture design you'll notice that, though there ARE some new things, it mostly follows the P6 blueprint.


RE: Intel is dead
By saratoga on 3/13/2006 8:11:59 PM , Rating: 2
"Take Yonah's front-end, add an extra decoder (for simple x86 instructions, the most common), macro op fusion and deeper buffers and you've got Merom's. The Reorder window is pretty much the same with, once again, deeper buffers. "

Yonah != P6. It has a reworked backend, and SSE2 support, which no P6 cores had.

Even ignoring that, you're argueing that everything has been changed, but its still similar. And you didn't mention the entirely new and highly agressive guess-and-check system for detecting memory dependencies when reordering loads and stores. I'd say the 4 OP wide decoder alone is a radical change.

"Take Yonah's, widen the Integer units to 64 bits (needed for 64 bit suppport) and add an extra shit/rotate ALU and you have Merom's. The FPU and SSE units were widened to 128 bits (AMD's are still 64 bits, by the way, though I think it was planning on widening them eventually) and tweaked (I think the FMUL unit is the one with the most improvements). This last thing is probably what will give the most significant performance improvement, since 128 bit SSE can now be executed in a single clock. "

I figured you were just making stuff up without looking at the info. Clearly I was wrong. However, I'm baffled how you can dismiss these changes (even ignoring the issue of Yonah and the P6). The Yonah looks more like the Opteron here then Core. The fully pipelined SSE units, 128 bit width and 50% increase in ALU resources are huge changes. Not to mention the entire system is now more agressively pipelined, while still being much wider.

"Also, the width of some of the internal buffers was widened, though the L2 cache retains the same 256-bit width (AMD's is still 128-bits but it was 64-bits in the K7). "

True, but the L1 is now accessible by both cores. And supposidly has a 2 cycle latency, so its likely a completely different organization, even if the cacheline size is the same.

"So if you look at Merom's microarchitecture design you'll notice that, though there ARE some new things, it mostly follows the P6 blueprint."

I think heres where we really disagree. Your interpretation of the "P6 blueprint" seems suspicously close to my definition of "modern OOO processor". If you consider changes in issue width, the addition of uop fusion, a huge improvement in OOO window, completely new cache system, ~40% increase in pipeline depth, doubled FU width, and huge changes in vector execution resouces "a band-aid", then pretty much all modern processors are just minor improvements on the P6. Which may be true, but is not very instructive.


RE: Intel is dead
By Furen on 3/14/2006 12:47:08 AM , Rating: 2
What the hell is a "modern OOO processor"? Does netburst still qualify as one? Because its frontend and execution are VERY VERY different to Yonah/Merom (trace cache, dual pumped ALUs, etc). The same can be said about the K8, its internal instruction set is actually completely different from Intel's and so is its decoding, scheduling and execution.

When I put band-aid in quotes I meant it to look mocking, though it seems this proves, once more, that text is the worst way to show sarcasm. I didn't mean to say Merom is a quick Yonah hack, which it certainly isn't, but that it IS based on the P6 underneath all the improvements. The same way the K8 is pretty much a 64-bit K7 with a longer pipeline and various improvements to its buffers, though to a greater extent, so I think Intel's "built from the ground up" claims are not completely true. Netburst, on the other hand, was a completely different animal when it came out.


RE: Intel is dead
By Furen on 3/14/2006 12:51:38 AM , Rating: 2
By the way, I am quite hyped up about Intel's memory disambiguation. I'm actually quite interested to see what Merom's IPC ends up being and how much of this can be attributed to this tech.


RE: Intel is dead
By saratoga on 3/13/2006 8:12:42 PM , Rating: 2
Above I meant to say that Yonah had a reworked front end and deeper pipeline, not backend. You get the idea though.


RE: Intel is dead
By johnsonx on 3/13/2006 8:43:23 PM , Rating: 2
quote:
an extra shit/rotate ALU


yes, I know it's just a simple typo, and I'm not a typo nazi... but damn, that made me laugh! I want a CPU with LOTS of shit/rotate ALU's!



RE: Intel is dead
By Furen on 3/14/2006 12:49:07 AM , Rating: 2
It's a new kind of integer operation that improves performance by 25x when combined applied on a bull operand.


RE: Intel is dead
By Visual on 3/14/2006 8:35:21 AM , Rating: 2
The MOD system of these comments SUCKS BALLS!
AnaxagorasZeres most probably didn't reply to the post that is above his right now, but yeah, a genius mod had to come in and make him look stupid by messing the threads.
Good thing message subjects kinda show what really happened.

Now come mod me.


Digitimes talks with Dallas Texas
By DallasTexas on 3/13/2006 10:20:55 AM , Rating: 4
Digitimes talks with tech guru, Dallas Texas:

Digitimes: What do you think of Henri's comment that Intel is just marketing, have no new architecture and Netburst was effectively a failure?

Well, clearly Henri is now eating crow since the IDF showing of Conroe. It seems Henri needs to put a fresh coat of lipstick on Opteron by tweaking up the clock and their new support of DDR2 as something they invented.

Digitimes: You forgot about the failure of Netburst?

Oh yeah, sorry. I forgot AMD still wants to talk about the 5 year old Netburst. Well, AMD would kill for the hundreds of millions of Netburst chips Intel has shipped. So, from a product success perspective, it's done quite well. Intel created Netburst to address multimedia, but clearly, it was an architecture that was not meant to scale over 5 years.

Digitimes: What do you think of Henri's suggestion that AMD invented the integrated memory controller, 64 bits and about Intel's FSB?

Henri is a marketing guy and AMD is given lots of lattitude as to what they can claim. Well, Intel integrated the memory controller in a CPU in 1997 and it was a bad idea. Basically, Intel integrates the mem controller in the chipset and AMD in the CPU. Where is the best place to do that? Well, there are pro's and con's of both. Regarding 64bits? Oh god, not that again.. Ask me something else, please.

Digitimes: Well, Henri claims Intel says they were right about 64bits
OK, I'll invent something now. I claim 128bits follow 64bits. There, I said it.

Digitimes: Stop being funny and just answere the questions, Dallas

OK, back to 64 bits.. Did Henri mention that Intel was correct back in 2003 when they said 64bits will not be mainstream for another 3-5 years? Nope. he will never say that. The fact is, 64bits is AMD's marketing scam. It's easy to sell since 64 is greater than 32 and most no which number is bigger. The real performance benefit is multiple core, not 64bits. I think AMD fears multiple core because they will have much difficult competing with Intel on multiple cores versus 64 bits because the former requires capacity, the latter requires just a sticker.

Digitimes: Thank you for your incredible insight, Dallas. What do you do now?
Your welcome. I now live in Dallas and really into photograhy now but the PC industry is sorta fun to watch.





RE: Digitimes talks with Dallas Texas
By TomZ on 3/13/2006 11:14:26 AM , Rating: 1
Absolutely brilliant - well stated.

I would add that, AMD is already taking shots at NGMA, claiming it is a "quick fix." But again, Intel will sell hundreds of millions of processors with this architecture over the next 3-5 years. So is this really a failure as Richard implies? Is he that arrogant to think that AMD's technology is so superior that NGMA will be sidelined and marginalized by AMD's developments? Oh, that's right - he's the marketing and sales officer - sorry, I forgot.


By Deliya on 3/13/2006 12:40:40 PM , Rating: 1
Oh yes! You simply forgot that you are talking about processor to be in "?" months. Nice.
And about hundreds of millions of processors: Weren`t Intel processors always pricier? And now they are cheaper? Why is that so? With your brains I would say that is because they had better processors. Because if you have better product, you cut prices, don`t you? NOT!
They cut prices because they were inferior and they were forced to do something about that. They sold so many processors because they slashed prices and because they are Intel - most people just hear the name and it is OK, no matter how good (or bad!) the product is.
Why don`t you comment on always really arrogant Intel marketing? When they brought new, supossedly superior technology in 90 nm, which was obviously complete failure, they still claimed they have "perfect processors for you!"??? How come AMD then beat them with technology "Intel used way back in 1997" and was not good enough for them? And you talk about arrogant? People please! Who`s more arrogant than Intel?
Don`t think that AMD will just sit and watch quietly. Like they haven`t proved before, that they can make good processors. Have some faith, people.
For a few years now, Intel was only talking and AMD was the one walking! It would be about time that Intel did some walking, but you`d have to be pretty stupid not to see, that for now they are still just talking.
Still, if Conroe, when it comes out, proves it is better processor for less or at least for the same money as AMD counterparts, I would still buy one. Somehow, I just don`t belive that...


RE: Digitimes talks with Dallas Texas
By