Athlon 64 architect returns to AMD

[tommo123]

Fabs aren't their main issue. First and foremost they need a competitive architecture.

it may not be the main issue but it's important surely? i mean even if they had an arch that was as good as intels they would still get stomped right due to intels fab advantage? (referring not to capacity as such but superior process node). i mean their chip would use more power, and couldn't scale as high. then there's the price aspect.

correct me if i'm wrong but the only reason the k7 onwards was so great was because intel got lazy. if intel had carried on with something more like the P3 (lading to merom and a core type arch) then the athlons would have been nice but not have been so competitive as they were. IMO to get on a level playing field with intel or even close they would need to design a chip that is way beyond what intel have since there's no way they're going to level the fab playing field. can't see it happening after bulldozer.

:wub:

 

[SickBeast]

AMD is going to need some serious luck to compete with Intel any time soon. Things will not bode well for them if they continue to develop the Bulldozer architecture. It will be 3-5 years before they even have a chance to recover from this fiasco, if they even exist at that point in time in their current form. Another couple of years with the way things are going and I could see AMD selling off their CPU division and focussing solely on GPUs and maybe even ARM chips.

 

[sefsefsefsef]

He was with AMD for a year. He didn't have much of an impact on anything in that short of a time-frame, and he certainly didn't architect a CPU in a year (not that one person could do it anyway).

Basically for a year he managed the CPU team.

I think you'd be surprised what a very talented person can do managing other talented people in just one year. Look what Asus did with the Nexus 7 in just a few short months.

Also, CPUs are just a collection of smaller component parts, which are generally developed independently of each other, so a "lead architect's" job is mostly choosing which parts to use, and influencing the research direction for what parts are developed next. Keller could easily have had a huge impact within even one year.

 

[grkM3]

amd is basically doing what intel did and went back to there core roots and left netburst.

after all isnt a ivy bridge a glorified P3?

the intergrated memory controller was from alpha along with the point to point system bus but intel bought them before the next gen ev8 could come out and scrapped it.dec also used a custom design and did not automate anything and could clock alot higher then the competition back then.

a beefed up athlon 64 is what amd needs right now

 

[Rvenger]

Could it be possible that they can make some last minute adjustments on the Piledriver or probably Steamroller architecture? (if it even has any hope at this point as an architecture)

 

[sm625]

I been telling amd what they need to do for years.

1. An integrated SSD controller. Get rid of the sata interface, since it is cost prohibitive to oems. 256GB of MLC storage would only cost oems about $80 more than a hdd if it was placed on the motherboard or on a DIMM with no controller, case, power supply, buffer, etc. For that price, almost every AMD system would be fast and have ample storage. Cheaper units would have 64GB and they would still be cheaper than HDD based notebooks today.

2. Integrated SIMDs inside the cpu floating point unit. No more gpu at all. Just integer cores and GCN fpu cores. They are woefully behind in this. Not only should they already have hardware, they should have had drivers too by now!

3. Most important of all: get rid of the southbridge. Fit whatever you can in one package, and get rid of the rest. You need 3 displays, 4 usb, 2 sata, audio, and LAN, and a few pcie lanes. That can all fit in a sub 1000 pin package.

4. Make memory controller capable of supporting GDDR5. Even if it is only 64 bit. The bandwith will still be greater than 128 bit DDR3-1600.

 

[Phynaz]

I think you'd be surprised what a very talented person can do managing other talented people in just one year. Look what Asus did with the Nexus 7 in just a few short months.

Also, CPUs are just a collection of smaller component parts, which are generally developed independently of each other, so a "lead architect's" job is mostly choosing which parts to use, and influencing the research direction for what parts are developed next. Keller could easily have had a huge impact within even one year.

While the engineering effort that goes into building anything is amazing, I can't believe that you are comparing assembling off-the-shelf parts to CPU design.

 

[sefsefsefsef]

While the engineering effort that goes into building anything is amazing, I can't believe that you are comparing assembling off-the-shelf parts to CPU design.

Do you have experience working in either field? So are you saying that you still think that Keller couldn't have a big impact in 1 year?

 

[Abwx]

Certainly that Seymour Cray did design his Cray1 almost alone ,
with a very reduced engineering crew but that was another time.

Currently , although individual talent is a plus , a CPU design involve
thousands specialized engineers and has become more of a marathon
in montagneous terrain than a straight 200 meters course.

 

[Phynaz]

Do you have experience working in either field? So are you saying that you still think that Keller couldn't have a big impact in 1 year?

Define "big".

 

[sefsefsefsef]

Define "big".

OK, fine. Don't answer my question. If you do want to answer my question, use whatever definition of "big" that you want, which is what I expected you to do in the first place.

 

[Atreidin]

I don't know if this guy is any good or not, but if he is, I think 1 year is more than enough time to do a lot of good things, maybe make a big difference. You might not see the effects immediately but I think it is silly to discount the value of effective leadership.

Way too much emo in this thread.

 

[ShadowVVL]

Could it be possible that they can make some last minute adjustments on the Piledriver or probably Steamroller architecture? (if it even has any hope at this point as an architecture)

I hope this is what they are going to do.I think If he is a good engineer he may be able to make some decent improvements with BD.

I wonder if amd will stick with 2 cores on 1 module with what ever arch comes out 5 years from now.

 

[Olikan]

I hope this is what they are going to do.I think If he is a good engineer he may be able to make some decent improvements with BD.

I wonder if amd will stick with 2 cores on 1 module with what ever arch comes out 5 years from now.

IMO...
CMT design is very nice idea, but AMD done a bad approtch to it...

instead of having TONs of resourses for a very high peak ST performance and then adding another core to share it in MT performance...

this way, you have great performance in low threaded, but lose the relative performance when using more then 5+ threads...(but it makes up for having 8 cores)

AMD did the oposite, they cut down the peak performance of ST, and go more cores aproach...to save area and power...well, seems to be working as intended for trinity...bulldozer is just a rushed chip, like fermi

 

[exar333]

I think you'd be surprised what a very talented person can do managing other talented people in just one year. Look what Asus did with the Nexus 7 in just a few short months.

Also, CPUs are just a collection of smaller component parts, which are generally developed independently of each other, so a "lead architect's" job is mostly choosing which parts to use, and influencing the research direction for what parts are developed next. Keller could easily have had a huge impact within even one year.

Not applicable and no.

 

[pelov]

He was with AMD for a year. He didn't have much of an impact on anything in that short of a time-frame, and he certainly didn't architect a CPU in a year (not that one person could do it anyway).

Basically for a year he managed the CPU team.

It's not quite what you're implying.

The biggest screw-ups on account of AMD's recent CPU designs have been in the management's decisions and their direction. So, while it's true that a completely new CPU uarch would take 4-5 years, the goals and the planning are outlined in that first year and make or break what happens in the following 3-4 years. That first year determines the direction after which the rest is essentially written in stone.

He's not going to make an immediate impact. Steamroller should already be done and he'll likely only be able to tweak just a few minor things in the Excavator cores. That isn't why they brought him in, though, and nor is that why he joined AMD (rejoined?). He's likely there to decide what happens with the bigger decisions that he can have an impact 3-4 years down the line and provide a bit of direction with respect to where Apple, and in turn the market, are headed in the next 5-10 years.

 

[Phynaz]

OK, fine. Don't answer my question. If you do want to answer my question, use whatever definition of "big" that you want, which is what I expected you to do in the first place.

Don't go and get all pissy, I just want to know, when you say a big contribution, what do you mean?

My idea of big could be a long ways away from your idea of big. Because we aren't using the same definition, we could argue about it forever.

 

[sefsefsefsef]

Not applicable and no.

So are CPUs designed on one big canvas, one gate at a time (rhetorical question, I already know the answer)? That's what the opposite of CPUs being constructed out of "a collection of smaller component parts" would be.

 

[BenchPress]

I just want to know, when you say a big contribution, what do you mean?

What's not to understand? Top-level managers make key decisions that can make or break a product, even in as little as one year. So big means big.

 

[CTho9305]

Not applicable and no.

So are CPUs designed on one big canvas, one gate at a time (rhetorical question, I already know the answer)? That's what the opposite of CPUs being constructed out of "a collection of smaller component parts" would be.

They're generally designed with many smaller canvases in parallel. But I think ExarKun333 is taking issue with your implication that the lead architect just picks well-defined pieces off a shelf to slap together... it's much more complex than that (I don't think component selection is a remotely useful analogy).

 

[intangir]

I think this is the first good news for AMD CPUs I've heard in way too long. This guy Keller seems to have had a hand in everything good AMD has produced. We might not see the results for a few years yet, but it shows that AMD still intends to invest long-term in their CPU products.

 

[Idontcare]

The fact that AMD's ex-CEO Dirk Meyer, who was the chief designer of the original K7 Athlon having come from DEC beforehand, was sold on the idea of Bulldozer being the successor to the K10 does not leave me with all that much confidence that just because AMD has rehired their chief designer of the K8 (which was basically a jazzed up K7, albeit jazzed up in all the ways that mattered) that AMD is suddenly going to turn on a dime, in industry timeframe, and herald a new era of k8-like competiveness.

Now it is true that he could be the equivalent of Steve Jobs who reinvented Apple and ushered in their era of dominance once he was brought back in to head the company...Keller could be all that and a bag of chips...but that doesn't mean he'll be allowed to do the things he may feel are necessary to effect the kind of change that is necessary. He's not being brought in at the CEO level, he's still going to be swimming upstream with respect to the existing decision-making structure as well as whatever corporate culture issues that may abound within the engineering teams that resulted in the creation of bulldozer in the first place.

I'm hopeful, but to be honest I thought if anyone stood a chance of knowing what it would take to drive AMD into greatness it was Dirk Meyer given his success in creating the original K7 for AMD and then watching it evolve into the K8 and eventual K10.

But Dirk got shown the door, so if even that guy couldn't get AMD turned around as CEO and as an experienced CPU designer then what chance does an even less experienced CPU designer have, given that he won't have the far-reaching decision making capabilities the likes of what Dirk had?

There's hope, and then there's history, and history shows we should have little hope if we are attempting to be realistic about the relevance of this latest addition to team AMD :\

 

[Olikan]

Now it is true that he could be the equivalent of Steve Jobs who reinvented Apple and ushered in their era of dominance once he was brought back in to head the company...Keller could be all that and a bag of chips...but that doesn't mean he'll be allowed to do the things he may feel are necessary to effect the kind of change that is necessary. He's not being brought in at the CEO level, he's still going to be swimming upstream with respect to the existing decision-making structure as well as whatever corporate culture issues that may abound within the engineering teams that resulted in the creation of bulldozer in the first place.

IIRC, during that time bulldozer didn't had the curent design...i bump the old slides later

 

[sefsefsefsef]

They're generally designed with many smaller canvases in parallel. But I think ExarKun333 is taking issue with your implication that the lead architect just picks well-defined pieces off a shelf to slap together... it's much more complex than that (I don't think component selection is a remotely useful analogy).

That's lame. I neither said, nor implied that. I just re-read what I originally wrote, and I stand by it.

If I were to make a clarification, I would say that mostly the architect's job is more to decide the *ratio* of the components with respect to each other (e.g., how many ALUs, versus how large the instruction window will be, versus the load queue depth, versus how wide the memory controller(s) will be, versus how many cores there will be, etc.), and not picking which ALU to use from a vast library of ALUs (although there is of course some of this also).