FX - Worthy Successor to K10?

Discussion of all forms of processors, from AMD to Intel to VIA.

Moderators: Flying Fox, morphine

Re: FX - Worthy Successor to K10?

Postposted on Fri Oct 14, 2011 9:14 am

Bulldozer has some very bold ideas, but the first implementation is a poor one. I hope they will improve and optimize it much as the R600->RV670->RV770 in the GPU space.
Cache hierarchy is just plain weird and I expect (hope) that it will receive some re-balancing in the future. I hope they increase the L1 cache size, decrease L2 cache size and hopefully decrease L2 cache latencies as well. L2 cache is likely also burning a lot of power currently as my understanding is that it must run at the same speed as the core.
Uncore is also much larger than expected. The 4 modules should be around 850 million transistors and about 125mm-sq die area if their ISSCC papers are to be believed. So more than 1 billion transistors and almost 180mm-sq die area are dedicated to the uncore. No idea why the uncore is so huge.
codedivine
Gerbil Elite
Silver subscriber
 
 
Posts: 705
Joined: Sat Jan 24, 2009 8:13 am

Re: FX - Worthy Successor to K10?

Postposted on Fri Oct 14, 2011 9:21 am

Glorious wrote:Interesting article Jigar.

I can totally see managers thinking like he describes.

AMD wants to save money, so their management looks at ATI which heavily relies on automated layouts (as does Nvidia) for their graphics products and thinks:

It works for them, right? And it's cheaper, right?

So what could go wrong? It's proven method that's cheaper, so we're obviously throwing money away!

Oops.

They may have also believed that they couldn't hit their launch window with a time-consuming hand-tuned design, and felt that they had no choice but to go with automated tools. OTOH they missed the window anyway; and how much of the delay has been due to them trying to fix the problems caused by the automated design tools?

Lack of hand-tuning also hurt them back in the Athlon XP days, when they released the "Thoroughbred" core XPs. The T-Bred was a straight die shrink of the earlier "Palomino" core, and did not perform as well as expected. It wasn't until they re-tuned the design for the smaller process that the "Thoroughbred" core really came into its own.

"There's never time to do it right, but always time to do it over."

Glorious wrote:This kind of thinking is running rampant everywhere from what I can tell. All over the place managers & leaders are looking at numbers and treating them the same as reality, ignoring the fact that the numbers are, and will always be, an abstract represention of reality, not a depiction of reality itself. That mistake is called reification, and it's turning into a big problem.

Even worse is when management tries to bend reality to their will by manipulating the numbers. I have seen situations where projects are literally managed by enabling and disabling time charge codes to control what people are (theoretically) working on! (Result: Engineers continue working on whatever they need to be working on to meet their schedules, and charge time against "overhead" instead. Management then gets royally pissed off because there's too much "overhead". :roll:)
(this space intentionally left blank)
just brew it!
Administrator
Gold subscriber
 
 
Posts: 37830
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: FX - Worthy Successor to K10?

Postposted on Fri Oct 14, 2011 10:00 am

JBI wrote:Even worse is when management tries to bend reality to their will by manipulating the numbers. I have seen situations where projects are literally managed by enabling and disabling time charge codes to control what people are (theoretically) working on! (Result: Engineers continue working on whatever they need to be working on to meet their schedules, and charge time against "overhead" instead. Management then gets royally pissed off because there's too much "overhead".


Yup, that sort of problem is universal at this point. Managers then spend big money buying "better" tools with more granular "control" in order to better circumvent such things.

Which, as you've noticed, is entirely missing the point. :(
Glorious
Darth Gerbil
Gold subscriber
 
 
Posts: 7877
Joined: Tue Aug 27, 2002 6:35 pm

Re: FX - Worthy Successor to K10?

Postposted on Fri Oct 14, 2011 10:12 am

Glorious wrote:Yup, that sort of problem is universal at this point. Managers then spend big money buying "better" tools with more granular "control" in order to better circumvent such things.

Which, as you've noticed, is entirely missing the point. :(

Indeed. At my current job, we're basically down to sub-one week granularity now. And that's for *individuals*, not for groups/projects as a whole. And this level of detail has visibility... oh... probably 3 levels up the management chain? WTF... :o
(this space intentionally left blank)
just brew it!
Administrator
Gold subscriber
 
 
Posts: 37830
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: FX - Worthy Successor to K10?

Postposted on Fri Oct 14, 2011 2:37 pm

Jigar wrote:This looks like a good read.

This is interesting, but difficult to understand. While the chip was being designed, the man all the way at the top was an engineer -- one of the best out there. This is not some corporate suit who can't tell the difference between designing a GPU and designing a CPU; he had to have known that there is a price to pay for automation. I guess the choice could have been between using this automation and getting a larger and slower chip and delaying the launch (yet again...), but they wound up delaying it anyway since their first revision was almost certainly not competitive even with Phenom II, never mind Intel.
Althernai
Gerbil
 
Posts: 84
Joined: Thu Nov 19, 2009 8:02 pm

Re: FX - Worthy Successor to K10?

Postposted on Fri Oct 14, 2011 3:24 pm

Althernai wrote:
Jigar wrote:This looks like a good read.

This is interesting, but difficult to understand. While the chip was being designed, the man all the way at the top was an engineer -- one of the best out there. This is not some corporate suit who can't tell the difference between designing a GPU and designing a CPU; he had to have known that there is a price to pay for automation. I guess the choice could have been between using this automation and getting a larger and slower chip and delaying the launch (yet again...), but they wound up delaying it anyway since their first revision was almost certainly not competitive even with Phenom II, never mind Intel.

It might not have been the engineer that made the decision. It may have been an executive decision.
flip-mode
Gerbil Khan
Silver subscriber
 
 
Posts: 9084
Joined: Thu May 08, 2003 12:42 pm
Location: Cincinnati, OH

Re: FX - Worthy Successor to K10?

Postposted on Sat Oct 15, 2011 6:11 am

Krogoth wrote:Bulldozer is superior to K10.

How sad would it be if that were not the case!?

It easily beats X4 and catches up with X6 despite having two less "real cores" at its disposal at multi-threaded applications.

Yes, but doing so on 8 cores that together have way more transistors.

AMD is already working on desktop versions of the architecture.

I think that is an "oxymoron", there is no way to make this a good desktop chip.
CBHvi7t
Gerbil In Training
 
Posts: 2
Joined: Thu May 19, 2011 10:35 am

Re: FX - Worthy Successor to K10?

Postposted on Sat Oct 15, 2011 8:11 am

CBHvi7t wrote:there is no way to make this a good desktop chip.

That seems overly pessimistic to me. If they can tweak the design/process to get the power consumption down and/or clock speeds to ramp, and drop the price a notch, it'll be competitive in the mid-range. It may never be an *excellent* desktop chip, but IMO it still has the potential to be a *good* one.
(this space intentionally left blank)
just brew it!
Administrator
Gold subscriber
 
 
Posts: 37830
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: FX - Worthy Successor to K10?

Postposted on Sat Oct 15, 2011 11:47 am

just brew it! wrote:That seems overly pessimistic to me. If they can tweak the design/process to get the power consumption down and/or clock speeds to ramp, and drop the price a notch, it'll be competitive in the mid-range. It may never be an *excellent* desktop chip, but IMO it still has the potential to be a *good* one.

According to AMD's own slide, Piledriver's performance per watt is expected to be 10-15% higher than Bulldozer's. If this turns out to be the case, then I'm not sure it can be competitive in anything except price. Also, unless they release it in the next six months or so, it will be competing not with Sandy Bridge, but with Ivy Bridge and its 22nm process and 3D transistors.

The key question here is how much of the problem is due to manufacturing and how much is inherent in the architecture. This is why I commented on the "speed-demon" aspect of Bulldozer in the massive review thread. If the issues are because they didn't handcraft enough transistors or because Global Foundries couldn't perform up to expectations, then the next iteration should remedy them. However, if the architecture itself is flawed, then AMD is in serious trouble. Intel revised NetBurst for half a decade before realizing that no amount of polishing that turd will get it to perform better than an updated P6. AMD doesn't have Intel's deep pockets or market position and it would not survive that kind of mistake.
Althernai
Gerbil
 
Posts: 84
Joined: Thu Nov 19, 2009 8:02 pm

No

Postposted on Sat Oct 15, 2011 12:31 pm

just brew it! wrote:If they can tweak the design/process to get the power consumption down and/or clock speeds to ramp, and drop the price a notch, it'll be competitive in the mid-range. It may never be an *excellent* desktop chip, but IMO it still has the potential to be a *good* one.

Ok, let us assume they make a two module 150(mm)² Desktop version, this they could sell at $125
Had they shrunk Thuban, it would be 150(mm)² too and could be sold at $200
This is not even comparing AMD to intel because they would look like a garage firm that made a deal with an insolvent foundry to produce a mammoth chip.
Bulldozer - Worthy Successor to K10? :roll:
The only way I see, to make this less than a disaster, is to make a three module version of this that has lost half of it's L2-cache
and introduce a lot of tweaks by human engineers to save power raise clock and save area. The talk of overuse of CAD and poor yields at leased signal room for improvement. The architecture might not be the failure it seems to be, after some more man hours are sunk into it.

Edit: I agree with Althernai
CBHvi7t
Gerbil In Training
 
Posts: 2
Joined: Thu May 19, 2011 10:35 am

Previous

Return to Processors

Who is online

Users browsing this forum: No registered users and 3 guests