AMD righting the ship?

[Daveburt]

There's been a lot speculation about the internal changes going on AMD lately (most of it negative)...

There's an interesting article over @ VR-Zone concerning the "Steamroller (K13)" improvements.
Here's a quote:

"Steamroller is not Bulldozer Enhanced. F*** no. The layout might look the same but our LEGO blocks are completely different. When all is said and done we should get 45% improvement and this goes to show how the Bulldozer was f***** design. This is all what Bulldozer was supposed to be."

Source: http://vr-zone.com/articles/amd-pus...erformance-increases/17088.html#ixzz2GYA1FnAK

Hopefully the changes made will payoff and this isn't more hyperbole from management...

Guess we'll find out later this year.

 

[EarthDog]

I will believe it when I see it.

 

[trekky]

ya i agree with earthdog show me proof of a 45% improvement

 

[bmwbaxter]

Only time will tell I guess. It would be nice for sure, better performance will ensure better competition and prices

 

[Bobnova]

45% is one hell of a claim. 45% in what?
45% higher per-"core" performance? That'd be insane.
45% x87 performance? That'd bring it back on par with Intel.
45% floating point? That'd be both insane and on par with Intel, and it's possible if they double the floating point "cores" and give each core its own again.
I predict 20%, on average, over BD.

 

[ssjwizard]

With the number of weak points in the BD core I would not be surprised if they manage it. I believe I saw estimates of 2014 for SR which would put it in the realm of DDR4 land also(maybe). Of course this far out any performance numbers are just smoke mirrors and rumors. Im happy with my PD rig but if SR puts it down the way they say it will, I might be convinced for an early upgrade.

 

[juane414]

I would think 15-20% improvement in most categories would be reasonable. Like Bobnova said, 45% in what? If Steamroller is 45% better than Piledriver in some areas I might be looking at an early upgrade as well. But my 8-core ~4.5GHz FX-8320 is flaming fast for anything that I do now, or will do in the foreseeable future. Quite the improvement over my 2.4GHz Core 2 Duo laptop!

 

[Hicksimus]

This would be nice. But I don't believe in it...AND...so what? Intel made this performance jump 2 years ago. AMD could release this 45% improvement tomorrow and be 2 years late. Plus I seem to remember a ton of Bulldozer hype(including info from AMD to fuel it) and that sure turned out to be a load of.....

 

[Black C5 Z06]

I look forward to it, but I'm not getting my hopes up.

 

[ssjwizard]

So I decided to do a bit of research on the two people named in that article. Jim Keller is responsible for dozens of revolutionary processor designs including the original Athlon 64 and Opteron CPUs, along with Apples A(x) series of chips in all there current devices. John Gustafson is an expert on parallel computing design, and engineered some of the first cluster computing.

Assuming that they're as good as it seems, with a full year of research I dont see why they couldn't improve performance by that much. BD was a good idea they just didnt implement everything correctly. I like the look of this new CPU in theory, and if each of those parts is ALSO faster I can see the total package going up by the claimed 45%. Im not going to count on it though.

 

[Rotary]

Well the fact that AMD knows its the end of the road for them if they cant come up with something soon. Maybe this is it, finally something to pull them out of debt.

 

[Daveburt]

I didn't write the article, just shared it...
And I understand the doubters, I was sucked in by the BD hype too!
BD still wasn't as bad as some folks make it out to be (especially in MT), and Vishera was a nice improvement.

ssjwizard: That slide you posted is also what caught my attention.
I've seen people a lot smarter than me say the front end was a big part of the problem with K12.
So the fact that they gave each core it's own decoder might offer a decent performance boost!
The other thing I don't quite understand is the addition of the MMX module to the FP compute units.
Like you though, I don't doubt the credentials of Jim Keller & John Gustafson!

I've always rooted for the underdog (and AMD), they certainly fit that description right now...
It'd be nice to see them keep Intel honest, like they did with the A64...

 

[ssjwizard]

So the fact that they gave each core it's own decoder might offer a decent performance boost!
The other thing I don't quite understand is the addition of the MMX module to the FP compute units.

If you also see that in the new design each decoder can also send directly to the floating point scheduler, so the previous limitations of the shared FPU should theoretically disappear.

As I make it out the MMX unit is being added to keep legacy support while also removing unnecessary muck from the other compute pipelines, since its rarely employed now.

 

[irrykuxeh]

I'd love to see this, but we'll all have to wait and see :/. I will remain optimistic though

 

[Anonaru]

45% is one hell of a claim. 45% in what?
45% higher per-"core" performance? That'd be insane.
45% x87 performance? That'd bring it back on par with Intel.
45% floating point? That'd be both insane and on par with Intel, and it's possible if they double the floating point "cores" and give each core its own again.
I predict 20%, on average, over BD.

I'm thinking 20% on average over PD, not BD. If Steamroller is done correctly, the larger gains they've been talking about are very feasable, particularly since they left the L3 cache issue and the resource sharing issues from BD in the PD design.

I am really, really hoping that AMD nails this one, we need the competition, and I really enjoy having two unique platforms to tweak and push to the limit

 

[EarthDog]

particularly since they left the L3 cache issue and the resource sharing issues from BD in the PD design.

That will help in FPU performance (someone correct me if Im wrong but that is what shares the cache)...but what about the rest?

 

[Anonaru]

That will help in FPU performance (someone correct me if Im wrong but that is what shares the cache)...but what about the rest?

The resource sharing is a big problem; the long pipeline and other inefficiencies add up quick too.. If they do nothing more than fix existing problems and leave the architecture the same, they can and probably will get a pretty substantial increase in performance. If they actually bust their butts and progress the technology a bit, then it will be a good day for AMD.

If you recall how much more per-core performance you get on BD by disabling a core per module, you can pretty easily measure just how big of a bottleneck these shared resources are on these chips

AMD needs a large jump in IPC. Lets hope

 

[EarthDog]

I dont recall, no. Last I recall, and could be wrong, was that only the FPU shares that cache. If you arent crunching FPU, the integer cores (or w/e) get access to the full cache. BUT this is a bit above my head and I'm simply regurgitating what I have read/remember.

 

[Anonaru]

I dont recall, no. Last I recall, and could be wrong, was that only the FPU shares that cache. If you arent crunching FPU, the integer cores (or w/e) get access to the full cache. BUT this is a bit above my head and I'm simply regurgitating what I have read/remember.

No, you're actually 100% correct about the cache, to the best of my knowledge. The shared scheduler and pipelines are what I'm talking about though, and what effed over BD so badly (And why their performane scales so poorly beyond 4.3GHz). An interesting find some BD users found, particularly with the 8150, was that if you disabled one integer core per module, you got large (and at times extreme) gains in per-core performance. Bottlenecked by design

I believe they placed the per-core performance efficiency of BD at like 75~% of what it should be, so basically a 20-something percent performance hit just because of poor design choices. I'm still a firm believer that PD is a good architecture that needs to have the living turds tweaked out of it

Edit: I completely forgot about the IMC. AMD's IMC, for lack of better words, blows. If they gave this chunk of hokey a revamp, they could pinch a pretty substantial gain here too

 

[EarthDog]

I thought PD's IMC was much more robust (but still not IB levels) than BD?

Thanks for the info... regardless, Im looking forward to seeing a big jump.