Barcelona VS K8: nice

[Ace Pace]

Anandtech1 and Anandtech 2.

No need for comments, just look at the benchs. AMD got competetive again. Not a victory over Conroe, but it's pretty good at the buck per bang.

[Starglider]

Not a victory over Conroe, but it's pretty good at the buck per bang.

Posted about this yesterday.

Basically it beats the current Core 2 chips on a clock-for-clock and performance-per-watt basis by between 5 and 15%, but current models only clock up to 2.0GHz. AMD hopes to be up to 2.5-2.7 GHz by the end of the year, which is enough to match the best current Core chips. However Intel can easily launch clock-bumped versions to maintain its lead, and Nehalem-based chips (with major gains over Kentsfield) are only months away from launching.

AMD has priced these chips to beat Intel on performance-per-dollar, but given that their die sizes are so much larger and their financial position is so precarious, I am dubious about their ability to sustain any kind of price war. To stem the market share losses AMD needed to convincingly beat Intel's current generation, not match their mid range chips at a slightly lower price.

I fear that unless AMD somehow manages to scale these chips past 3GHz in the near future, or Intel makes another major blunder, we may see AMD forced out of the desktop x86 market and possibly bankrupt within a year.

[Ace Pace]

Doh, i'm fucking retarded, shoot me.

[The Kernel]

*snort*

SpecJBB: 2% -AMD
Valve: 13% - Intel
CineBench: 12% - Intel
PovRay: 4% - AMD
Myri: Tie
Stars: 15% - Intel
Folding: 3% - AMD
PanFac: 13% - Intel
PicColor: 25% - Intel
WME: 6% - Intel

Someone over at Tech Report created this chart of a 2Ghz Core 2 vs a 2Ghz Barcelona from the Tech Report benchmarks. Granted this is a limited set of benchmarks, but it's all we have right now (thanks AMD for doing a shadow launch).

The Green numbers are AMD wins, the Blue are Intel. Not only does Intel win more tests, but at much greater percentages...

Competitive on the server side maybe (at least until Nehalem) but K10/K8L can't even manage to match Clovertown clock-for-clock on desktop applications AND AMD is dealing with a 50% clockspeed deficit then there are some serious problems at work here.

There's a few important points that everyone should keep in mind about this launch:

1) K8L/K10 is not as fast as Clovertown in most situations clock-for-clock unless the workload favors AMD's bus architecture. Some server and HPC workloads in 2p-4p boxes will benefit here, but Intel is still extremely competitive in most situations.

2) For desktop workloads, Barcelona can't match Clovertown clock-for-clock. To make matters worse, AMD would have to boost the clockspeed of these processors by 50% just to be on equal footing with today's Clovertown processors.

3) Even though Barcelona doesn't look good compared to Clovertown, it's going to look even WORSE compared to Penryn. Intel's all set to launch Penryn in November and it will further increase their IPC and performance advantage.

4) Let's face it: the Core architecture has got some serious frequency headroom. Intel could probably release a 4GHz Core 2 Quad today if they wanted to with no problems at all. THe only reason they haven't done this is that they have no reason to; there is currently nothing that competes with Core 2 at the high end.

5) Penryn + frequency scaling will keep Intel ahead of AMD well into next year. What about beyond this? Intel's Nehalem is scheduled to hit towards they end of next year and it will deliver Intel's CSI platform with an integrated memory controller (negating AMD's two chief advantages) as well as serious architectural improvements.

Honestly, I don't see how this is anything but a loser for AMD. Barcelona is obviously a step up from K8, but it's too little too late for the most part.

[Ace Pace]

It's nitpicking time.

1) K8L/K10 is not as fast as Clovertown in most situations clock-for-clock unless the workload favors AMD's bus architecture. Some server and HPC workloads in 2p-4p boxes will benefit here, but Intel is still extremely competitive in most situations.

Comparisons between Barcelona preformance compared to Desktop processors are not very fair.

Barcelona is currently limited to DDR2-667, we were unsuccessful with attempts to run the memory any faster. Like all other MP Opterons, Barcelona requires the use of registered DDR2 memory, which is inherently slower than the unbuffered stuff we use on desktops. Because of these limitations we refrained from running any comparative benchmarks to desktop Athlon 64 X2s, instead we chose to run a single quad-core Opteron in our server platform against a pair of dual-core Opterons to simulate Phenom vs. K8 on the desktop.

Further differances include a differant cache design for Phenom rather then Barcelona(if I recall the data correctly) and other small things.
Before comparing directly, keep those facts in mind.

5) Penryn + frequency scaling will keep Intel ahead of AMD well into next year. What about beyond this? Intel's Nehalem is scheduled to hit towards they end of next year and it will deliver Intel's CSI platform with an integrated memory controller (negating AMD's two chief advantages) as well as serious architectural improvements.

This might have changed, the real world tech article on CSI mentioned that Intel might not roll out CSI on it's desktop/mobile platforms along with it's server boards, and even then, it might stagger the releases so the 4S and 8S boards get CSI before the rest of the servers.

[The Kernel]

Comparisons between Barcelona preformance compared to Desktop processors are not very fair.

It's a comparison to Clovertown which is a server processor. Yes some of the applications are desktop, but I'll address that in a second.

Barcelona is currently limited to DDR2-667, we were unsuccessful with attempts to run the memory any faster. Like all other MP Opterons, Barcelona requires the use of registered DDR2 memory, which is inherently slower than the unbuffered stuff we use on desktops. Because of these limitations we refrained from running any comparative benchmarks to desktop Athlon 64 X2s, instead we chose to run a single quad-core Opteron in our server platform against a pair of dual-core Opterons to simulate Phenom vs. K8 on the desktop.

Further differances include a differant cache design for Phenom rather then Barcelona(if I recall the data correctly) and other small things.
Before comparing directly, keep those facts in mind.

Oh come on Ace, you've been following tech for a long time, you know as well as I do that Barcelona is not going to make any serious differences in the desktop merely with a different cache design and some other small platform changes like HT3 and faster memory. Barcelona is going to be nearly identical in performance to Phenom, you aren't going to see some magical 15% increase on the desktop just from a little platform tweaking.

5) Penryn + frequency scaling will keep Intel ahead of AMD well into next year. What about beyond this? Intel's Nehalem is scheduled to hit towards they end of next year and it will deliver Intel's CSI platform with an integrated memory controller (negating AMD's two chief advantages) as well as serious architectural improvements.

This might have changed, the real world tech article on CSI mentioned that Intel might not roll out CSI on it's desktop/mobile platforms along with it's server boards, and even then, it might stagger the releases so the 4S and 8S boards get CSI before the rest of the servers.

Which is exactly what they should be doing. Intel's antiquated bus architecture, surprisingly enough, is only a bottleneck on 2p-4p boxes and even then only when the workloads are bandwidth intensive enough.

[Starglider]

Someone over at Tech Report created this chart of a 2Ghz Core 2 vs a 2Ghz Barcelona from the Tech Report benchmarks. Granted this is a limited set of benchmarks

Specifically it is a set of benchmarks using mostly Intel compilers that have extremely dubious behaviour (i.e. disabling all vector optimisations if 'GENUINE INTEL' is not detected). AMD chips have always done much better on GCC-compiled benchmarks and that gap has been increasing recently. The different between the Intel-hand-optimised and the normal benchmarks is quite apparent in the Anandtech review.

The situation is bad for AMD but it is not as ridiculously bad as you are suggesting.

and Nehalem-based chips (with major gains over Kentsfield) are only months away from launching.

Gah, that's about the third time I've said 'Nehalem' when I mean 'Yorkfield' now. I confess I have a hard time keeping track of all these Intel codewords.

5) Penryn + frequency scaling will keep Intel ahead of AMD well into next year.

However don't you also mean Yorkfield/Wolfdale? Penryn is the mobile version isn't it? Though early on it did seen to refer to the whole core generation.

[The Kernel]

Specifically it is a set of benchmarks using mostly Intel compilers that have extremely dubious behaviour (i.e. disabling all vector optimisations if 'GENUINE INTEL' is not detected). AMD chips have always done much better on GCC-compiled benchmarks and that gap has been increasing recently. The different between the Intel-hand-optimised and the normal benchmarks is quite apparent in the Anandtech review.

The situation is bad for AMD but it is not as ridiculously bad as you are suggesting.

Which specific benchmark from the above do you take issue with?

As for the Intel compiler issue, the ICC is a popular compiler (maybe not as popular as say VCC, but a lot more people use ICC under windows than the gcc compiler), so it's perfectly valid for real world benchmarking. Furthermore, AMD chips actually do fairly well historically under ICC compiled applications as most of Intel's enhancements are not specific to their system architecture.

EDIT: And it's a mischaracterization to say that Intel has "hand optimized" anything since you are really talking about third parties that just download and use the Intel compiler, not software that comes straight from the Intel labs. Face it: despite AMD's crowing over the past few days, the ICC is a long established compiler that is used all over the software industry.

However don't you also mean Yorkfield/Wolfdale? Penryn is the mobile version isn't it? Though early on it did seen to refer to the whole core generation.

Whatever, I don't have time to keep up with Intel codenames anymore.

[Ace Pace]

I'm not sure what you people are complaining about with Intel codenames, AMD is far worse.

I mean, the next gen of AMD codenames is:
Bulldozer
Bobcat
Falcon
Sandtiger
Shanghai
Predator
Budapest
Spider
Hardcastle
Pinwheel
Cartwheel

Which is translated thus:
So what they're saying is that they're new Hardcastle architecture will contain a Sandtiger which was Shanghai'd from the old Budapest architecture to act as a secondary Predator along with the new Bobcat. And thanks to Hardcastle the old issues associated with Bulldozer have been overcome by adding the new Pinwheel system. The whole thing sits atop of AMD's new Cartwheel board, allowing it to make better use of the new off-die Spider. Cartwheel also allows the old Falcon to be implimented on the same board without the usual problems that arise when you have a Falcon and a Spider feeding into the same die.

We could *never* do this sort of thing in the days of Sandcastle. God bless Hardcastle -Credit to InnocentBystander