Barcelona overview article at RWT

Rys

Graphics @ AMD
Moderator
Veteran
Supporter
Real World Tech posted a very nice architectural overview of AMD's upcoming Barcelona architecture (also previously known as K8L or K10, with the apparently final nomenclature being 'AMD Family 10h processors'), comparing it both to its predecessor and Intel's competition in terms of Conroe and Penryn.

Read the full news item
 
The projected performance seems to be in line with what has been discussed earlier in other threads about Barcelona and its relatives.

Single-threaded performance that is most important for the majority of desktop apps and games will remain an Intel strong point.

Dual-socket server will be a toss up between Intel and AMD, depending on the workload.

Multi-socket server will heavily favor AMD due to the platform design.


For gamers, Intel will most likely be the best choice well into 2008.
 
Will it? Seems more and more new games are dual+ core capable.

Yes. Intel will still be the best bet. Their CPUs have been dual-core for a while now and they perform better than AMD's. Intel also offers quad-core cpus today. At the end of July, Intel's quad-core cpu will drop down to $266. AMD's quad-core cpus are no where to be found.
 
Yes. Intel will still be the best bet. Their CPUs have been dual-core for a while now and they perform better than AMD's. Intel also offers quad-core cpus today. At the end of July, Intel's quad-core cpu will drop down to $266. AMD's quad-core cpus are no where to be found.

Ah, ok. Until Conroe came out, AMD was easily still top dog in games! Granted Intel did leap-frog ahead with Conroe (which was about damn time!). So that gives AMD 3+ years to leap frog Intel... ;)
 
Well we can only hope the IPC of "K7 part 3" is a lot better than Athlon64. I'm sorta concerned because AMD is being so quiet. They not only need to be faster than Conroe, which is like a 20-50% deficit for them right now, but they have to match or best the not-so-far-off Penryn.
 
Ah, ok. Until Conroe came out, AMD was easily still top dog in games! Granted Intel did leap-frog ahead with Conroe (which was about damn time!). So that gives AMD 3+ years to leap frog Intel... ;)

The only catch being the fact that I doubt AMD has enough fat to deal without major pains with being in the situation Intel was in. Just MHO.
 
1. The article's conclusion and performance forecast comparison is based on the leaked roadmaps that featured 2.6 GHz as the top clock, for the 4-cores. And as the article says as well, it all depends on the frequencies. Recent rumors says 3GHz as top, for the 4-cores. But anyhow, even the earlyer roadmaps feature a 2.9 GHz for the 2-cores, so there is a much narrower gap here, between K10 and Intel offerings. And perhaps the 2-cores also will go upwards in clock.

2. The article (as well as the figures) often speak about micro-ops, where there should be macro-op, instead. F.ex.:
"The pack buffer, which is part of the decoding phase, is responsible for sending groups of exactly 3 micro-ops to the re-order buffer (ROB)."
An exerpt from the Software Optimization Guide for AMD 10h Processors:
''Early decoding produces three macro-ops per cycle from either path. The outputs of both decoders are multiplexed together and passed to the next stage in the pipeline, the instruction control unit.''
A macro-op can contain several micro-ops.
 
Will it? Seems more and more new games are dual+ core capable.

It's hard to see where AMD's platform gives it any decisive advantage at 2 cores.

Dual core isn't far enough away from single-core for the design decisions made for 4-way scalability to have an effect.

Intel's design is native dual-core, and its cache is larger at 4 MB than the split 512 MB L2 caches and the L3 that hangs off the chip's crossbar.

AMD's 512KB L2 latency is 14 cycles, and Conroe's 4MB shared L2 is about 14 cycles as well.

The IMC might keep AMD closer to parity in some workloads, but Conroe's done well without one.

By the time AMD releases Kuma, Penryn will be just around the corner, and that chip will be better than Conroe.
If we assume equal performance between Kuma and Conroe, then Penryn wins out.
 
Dual core isn't far enough away from single-core for the design decisions made for 4-way scalability to have an effect.
I've got to agree completely with that, I really can't see the 1333MHz FSB being a bottleneck to feed two cores, either.

In the end, I'd be very surprised if AMD's per-core IPC in integer workloads was higher than Conroe's. The IMC will help, but AMD is also less aggressive in terms of small tricks (such as load/store reordering, which they added in K10 but isn't as advanced as in Conroe) or in terms of prefetching, so I'm not 100% convinced the IMC's lower latency is enough to give them an advantage there.

And obviously, Penryn has the advantage of possibly clocking at 3.4GHz+, and possibly as high as 3.7GHz if Intel gets aggressive. Even at an equal IPC, that would result in a 20%+ advantage for Penryn. Of course, as RWT also pointed out, this proves just how much things will depend on the frequencies achieved. I'm skeptical AMD could get a 280mm² core at 3GHz+ though, even with redundancy...

And dess, good points - I hadnt noticed the confusion between macro-ops and micro-ops there myself, oopsie. Maybe point that out in the comment thread at RWT?
 
Well, if Intel manage to launch Nehalem next year, I would say blah to both Barcelona and Penryn. :cool:

With an integrate the memory controller, point-to-point serial link technology, 8+ cores with 16+ threads running and a Multi-Level shared cache architecture, Penryn merely looks like a stop-gap for Nehalem. This is what Barcelona will have to fight in the long run, and I'm frankly very excited that Intel is making this move.
 
Well, I don't think AMD is excited. :) Their silence is rather scary. Just how can they ride Barcelona for years like they did A64 and AXP? Intel is pumping out new cores like wildfire.
 
AMD's financial model and structure is not tuned for this strategy. That market isn't large enough to sustain the current company unless they more than double their share in it. The last time I calculated the mix data, which was for the Q1 Warning Analysis, server gross profit was between 35-45% of the company's overall gross profits.

If AMD is losing money right now, I'm not really sure how they'd stop losing money by halving their gross profits? Now, of course, their goal isn't to completely lose their other markets, but if you focus too much on servers you cannot expect to be as competitive as Intel in your other markets, and then your ASPs, your margins and/or your volume will be hurt.

I do agree that AMD wants to create niches for itself though. HPC isn't the only one, however, I think, although it certainly is very significant. Energy-efficient low/mid-end laptops (see: Fusion) is another market they'd like to be good in, and so are "commodity consumer desktops" where the GPU and video decoding might matter more than your CPU (but you can bundle a CPU anyway!)

I really, really hope for their own sake that they are not overestimating the size of these markets, however. It's nice to think that you can be good in some markets, but when you have high expenses because of R&D, fabs and many other things... You need to be good in many markets, and in very large volumes. That's a bit like if NVIDIA said they didn't care about anything but the $399+ market. That's great, they'll get $50M of guaranteed revenue a quarter. So, errr, how do you fund R&D with that?

AMD said they will restructure and that they have a strategy to return to profitability. I sure as hell hope for the sake of the company and of the employees that work there that their 'strategy' is not just to become small and nimble enough again to be able to remain profitable with smaller markets. I still have some faith in AMD's management, so errr, please don't **** up guys? :(
 
Just to pile onto what has been mentioned:

Another point is that AMD's going to have troubles when Intel drops prices on its quad core products, worse issues when Penryn comes out, and a major fight on its hands when Nehalem comes around.

Nehalem especially is worrisome for AMD, as Intel might give AMD less than a year before it comes out.
It will adopt an interconnect scheme very similar to AMD's (maybe more advanced?) and if modular enough with rumored graphics capability would be able to compete top to bottom with AMD's platforms from server to Fusion.

Intel's platform strategy would also take place at a better process node and would most likely have a much better ramp than AMD could match.

AMD's counting on Fusion for a lot, and it's really stretching things with Barcelona.

There are rumored projects, such as Bulldozer, though I know nothing about them.

I don't see any solutions to the problem that AMD can't field the number of in-flight designs as Intel, or how AMD is going to compete when Intel can price them out with better process technology.

We're seeing the gradual balkanization of computing, where one size fits all can no longer count on Moore's law to make it workable.
AMD's problem is that it is suffering from this challenge more accutely than Intel, and it has fewer resources to handle it.

Relying on others to take up the slack seems questionable, since everyone (including AMD) waits to see what Intel will do.

The new SSE4 instruction for streaming is an interesting example.
AMD could use it and other instructions to make Fusion more compelling than just an MCM.
However, AMD can't force greater integration with new instructions without Intel's backing an SSE extension.
If Fusion comes out before Nehalem, it runs the risk of Intel's simply ignoring any GPU extensions, making the first round of Fusion a commercial non-starter.

AMD's company-saving initiative is at the mercy of its competitor.
 
The new SSE4 instruction for streaming is an interesting example.
AMD could use it and other instructions to make Fusion more compelling than just an MCM.
However, AMD can't force greater integration with new instructions without Intel's backing an SSE extension.
If Fusion comes out before Nehalem, it runs the risk of Intel's simply ignoring any GPU extensions, making the first round of Fusion a commercial non-starter.

Presumably MS is quite a powerful power-broker in this equation (as they seemed to be i the rocky road from 32- to 64-bit Windows computing). SSE4 is only important if it's used, and MS is the gating factor there isn't it?
 
Presumably MS is quite a powerful power-broker in this equation (as they seemed to be i the rocky road from 32- to 64-bit Windows computing). SSE4 is only important if it's used, and MS is the gating factor there isn't it?

Microsoft had a pretty signficant interest in x86-64, since its 32-bit x86 platforms were at a distinct disadvantage in higher-end markets that Windows is still trying to expand in.

AMD was lucky in that it didn't have a 64-bit architecture that x86 would eat the bottom out of like Intel did. AMD came first, and Microsoft wasn't going to support two x86-64 ISAs.

What overriding business interest does Microsoft have with Fusion?
Fusion doesn't enable any new market segments from an OS point of view.

to sum up:
Intel has pretty free reign if Microsoft doesn't care either way.
 
Last edited by a moderator:
Who knows. Seems to me that MS is interested in more things than just the bitted-ness of the Windows kernel these days.

The have interest in getting their software to as many people (and thusly, platforms) as possible. Picking favorites in the PC hardware space isn't going to further their goal, so I don't think Microsoft will have any vested interest in whatever Intel, AMD, or NVIDIA have up their sleeves. At least not in the already-entrenched PC space.
 
Back
Top