Why can't AMD compete with Intel?

[DanceswithUnix]

Since Intel decided to focus on CPU core leadership (starting with Core 2) AMD haven't really stood a chance of catching up, let alone taking a lead: Intel could probably throw AMD's entire annual revenue at R&D if they wanted! As much as I'd love to see a new, enhanced, enthusiast platform from AMD, I'd much rather see them turning in consecutive profitable quarters and ensuring their footing was stable going forward.

I think Intel have always *tried* to focus on CPU leadership. The only thing that changed was that the Israel office which was tasked with making the Pentium M did a much better job than they were expected to, so serendipity won over their original failing Pentium 4 roadmap.

[scaryjim]

I think Intel have always *tried* to focus on CPU leadership.

Maybe - I thought I'd read something somewhere about them feeling like they'd dropped the ball a bit with Pentium 4 (particularly the move to Netburst), and when Core 2 was launched they explicitly stated they were going to increase their R&D efforts to make sure that they kept moving forwards. Of course, AMD's troubles with both Phenom I and Bulldozer have helped Intel look particularly good in that regard...

The only thing that changed was that the Israel office which was tasked with making the Pentium M did a much better job than they were expected to, so serendipity won over their original failing Pentium 4 roadmap.

Perhaps - I can't help but wonder if they could've been faster to move away from Netburst, and if there was a certain amount of complacency; a little arrogance that their big core was good enough to compete. Pentium M was around for a long time before it was developed into Conroe: it still took work taking a decent mobile chip and turning it into an all-conquering desktop chip.

I've seen at least one article kicking around the suggests AMD could do the same thing, scrapping their big core and scaling up Jaguar, but iirc it has the same basic IPC as the Richland core, which means it would have to scale well past 4GHz (which it *really* doesn't atm) to become a genuine enthusiast CPU core. tbh it's hard to see where AMD's next big advance is going to come from - they have two cores of similar IPC that hit clock speed limitations before they can be competitive in straight line performance. They're really pushing the energy efficiency angle with relatively better performance at lower TDPs - but that won't get them a Core 2 moment, because Core 2 had significantly better absolute performance at lower TDPs. HSA will only get them so far, and it only takes one badly-coded but very popular game to tarnish a CPU family's reputation....

[kalniel]

Maybe - I thought I'd read something somewhere about them feeling like they'd dropped the ball a bit with Pentium 4 (particularly the move to Netburst), and when Core 2 was launched they explicitly stated they were going to increase their R&D efforts to make sure that they kept moving forwards.

I think it was more a change in direction - they were partly focused on headline ghz speeds and not IPC as a performance goal - that was shown up by AMD producing chips with better IPC that outperformed the netburst arch while also using less power. Luckily the mobile group came up with the core arch that concentrated on IPC and they then focused on efficient performance.

[watercooled]

On the subject of brand image disparity, there are a few recent examples we can compare. Incoming ramble...

AMD's Hawaii release wasn't brilliant in terms of PR - the reference cooler allowing the GPU to approach the throttle temperature, and not being particularly quiet about it. Of course, the issue was categorically fixed when the non-reference designs hit the market shortly thereafter, and at identical pricing. Nevertheless, the reputation of Hawaii being a 'hot' GPU has stuck, especially amongst people who haven't bothered to look at reviews of newer designs, and/or people who just like to flame AMD wherever possible for whatever reasons.

To contrast, we can look at Intel's Haswell cooling issues; a few threads on Hexus recently have reminded me of it, but it really doesn't seem to have attracted all that much negative attention, relatively speaking. It seems that the frankly terrible thermal path through the IHS essentially means the stock cooler is, at best, barely adequate to cool the desktop i7 CPUs even at stock settings, and things get worse with anything calling AVX2 instructions, with 90C+ temps and throttling not uncommon. Even after-market coolers seem to struggle keeping temps down to the point where seriously expensive closed-loop coolers seem almost de-facto for stock or mildly overclocked settings.

Now, it didn't seem all that long ago, where stock coolers were actually adequate for stock settings, and you might spend 15-30 quid on an aftermarket air cooler if you wanted a decent overclock and/or lower noise. With Haswell i7, it seems that 15-30 quid on top of the 250 quid CPU is *necessary* if you don't want it to start throttling and get the performance you paid a not-insignificant premium for (lets ignore reaching blistering temps for now, even though the same AMD flamers seemed to think they were a big deal when talking about Hawaii). If you should want to overclock that overclocking edition CPU you also paid a premium for, that's another sizeable chunk of cash you'll need to spend on a closed-loop cooler to keep it under throttling temps.

Then, as if to add insult to injury, they release Devil's Canyon, which is amongst other things specifically meant to resolve this issue. On the contrary, it doesn't address the root cause of the issue (die-IHS gap) and it seems sample variance comes into play meaning people are still getting extremely hot-running, stock-clocked chips. I'd consider the gesture of filling the assembly machine with some slightly different TIM then re-releasing the same chip to much fanfare, almost insulting to the enthusiast crowd it's aimed at.

And yet, where are the crowds with pitchforks we saw with Hawaii? Short of undergoing the risky procedure of delidding your expensive CPU, there's nothing you can do about the poor thermal characteristics of Haswell, and yet the attitude towards this seems to be more 'oh don't worry the high temperatures and throttling is normal' and 'oh dear that's unfortunate'.

I'm not saying one response or another is any more right or wrong, just there's a *massive* difference in attitude towards comparable fails of the two companies.

[kalniel]

I don't think there were any pitchforks from the knowledgeable Hexus folk over Hawaii.

[DanceswithUnix]

Pentium M was around for a long time before it was developed into Conroe: it still took work taking a decent mobile chip and turning it into an all-conquering desktop chip.

I expect there was a short delay whilst the politics at Intel turned the roadmap around, and then after that it really does take that long to create a new CPU.

But as for the P4, yes they did drop the ball. That wasn't for lack of development effort though, I'm sure they worked very hard to produce a device that could do multi threading and was supposed to scale to 10GHz and the design was the best they could do.

As for AMD adopting their low end CPUs for big core work, I think that misses several points. First is that IPC at low clock speeds won't translate directly to what happens if you double the clock speed. The chip just isn't capable of that, so you need to add pipeline stages extend buffers & queues and make caches bigger. That will tend to increase latency and hence drop IPC, so you need to make the chip wider issue to compensate. See it wasn't the Pentium M hardware design that became Core 2, it was the design ideals and mindset of the team behind it. The hardware was after all a well executed re-targeting of the P-III design.

So if you apply the Jaguar core mindset to Bulldozer? Well, I think they are already well down that path. Just look at how laptop centric the current cores already are.

Last thing is, I get the impression that the big and little core teams at AMD do actually talk to each other. So if one has a unit or technique the other can benefit from, they can use it. So if big core can take any bits they need from the cat cores, why do they need the entire core?

[watercooled]

I don't think there were any pitchforks from the knowledgeable Hexus folk over Hawaii.

Just the general impression I got reading around other forums, news sites, etc.

And as I said I'm said I'm not saying whether the response was or wasn't justified - it was a definite fail IMO considering they could have just allowed aftermarket coolers from the start and/or waited longer if that's what was required, and none of that would have happened.