Reason for Brisbane's increased L2 latency

[BitByBit]

"We now have an explanation straight from AMD for the higher L2 cache latencies in its 65nm Athlon 64 X2 processors. The company has confirmed that the higher latencies are the result of provisions built into its 65nm processors for larger L2 cache sizes. They are quick to emphasize that this revelation is not tantamount to a product announcement, but the provisions are there should they decided to release such a product.

On the subject of the die size of current 65nm parts, they say the relatively small reduction in die area from 90nm to 65nm is not the result of added L2 cache being placed into silicon and then deactivated in the 4800+ and 5000+ models we tested. Instead, the modest reduction in die size has origins in the esoterica of process technology and AMD's model of continuous, gradual improvement to its manufacturing techniques. We've asked for additional detail on this subject, but that's all we can report at this time."

Techreport article

 

[Arkaign]

Bleh, sounds like just a stupid excuse for a disaster. This is AMD's Prescott launch. A new core that is inferior to the old one. Even if AMD made AM2 A64 X2s with 8mb L2, they would still be 2nd-class products. Bring on K8L, hopefully the reason for the horrible initial 65nm AM2s is that all their R&D is on that track for now.

 

[Hard Ball]

Originally posted by: Arkaign
Bleh, sounds like just a stupid excuse for a disaster. This is AMD's Prescott launch. A new core that is inferior to the old one. Even if AMD made AM2 A64 X2s with 8mb L2, they would still be 2nd-class products. Bring on K8L, hopefully the reason for the horrible initial 65nm AM2s is that all their R&D is on that track for now.

Actually, it was reported the same with the original Yonah release, where the L2 latency was 14ns to the 10ns of Dothan before it, even though the L2 size were similar. As it turned out, the L2 arbitration mechanism was designed to be implemented in both Yonah and Merom (4MB), hence the added latency in the earlier product. It is a quite common occurrance, saving cost and effort on R&D, with only a small amount of impact on performance.

 

[Duvie]

Anandtech guessed a similar response of larger L2 cache in the future though they said all current roadmaps do not show any chips coming with more cache. If anything most of the line has seemed to have gone towards the lower 512kb of cache per core....

Like someone pointed out in another thread which is quite sad...


90nm A64/X2 sckt 939 > 90nm A64/X2 AM2 > 65nm A64/X2 AM2.....

that is now 2 steps back

 

[BitByBit]

The news also begs the question: just when are these larger cached X2s going to be released? K8L will be released some time in Q3 next year, in both dual and quad varieties, thus rendering existing X2s obsolete.
Perhaps the new chips will feature in an amended roadmap some time soon.

 

[aka1nas]

That answer might just be an excuse and they will fix it in the next batch for all we know. Or they may have caught it already in newer batches that are still being manufactured. Or they are just very inept and didn't catch it.

 

[nullpointerus]

Originally posted by: aka1nas
That answer might just be an excuse and they will fix it in the next batch for all we know. Or they may have caught it already in newer batches that are still being manufactured. Or they are just very inept and didn't catch it.

I think the exact latency cycle count of this kind of cache is related to the design of the chip, not how well a particular batch of chips turns out. If this is true, the decision was made months ago to keep some options open until K8L is out.

 

[n7]

AMD is in big trouble (in terms of performance comparably) IMHO...but we already knew that anyway.

I just hope they can pull something out of nothing eventually, as i prefer to see competition, & right now, we have none (aside from the value segment)...

 

[A554SS1N]

It's a shame to say it, but these excuses from AMD just confirm that they're on a loser.....they better wake up fast, because taking away performance is just plain dumb, no matter what is said, when they are on their backfoot in the market. Surely they knew tech sites will review it and bring it up?

 

[tylerw13]

Originally posted by: n7
AMD is in big trouble (in terms of performance comparably) IMHO...but we already knew that anyway.

I just hope they can pull something out of nothing eventually, as i prefer to see competition, & right now, we have none (aside from the value segment)...

i dont think they are in trouble they will eventually get their problems corrected...however i dont understand why they are wasting time and money on products that are less than up to par??? can anyone help explain this to me??? i just am at a loss with what they are doing....it reminds me of intell a while back....freak dont waste money and time sometimes i dont understand how companies think

 

[SlowSpyder]

I thought this was just a die shrink... I don't understand why everyone seems disappointed in the performance, it's 99% the same performance as the 90nm parts. Why would you expect performance to increase?

 

[MDme]

Just like everyone here, this move really puzzles me. BUT let us not forget that there was probably a reason for all this. The new core is slower for sure but if this move meant something for the future (K8L and derivatives) then it will be a welcome move. remember no matter what they do to K8, it's not gonna beat Core 2, so this may be something intended for K8L. We can judge all we want but in the end, AMD probably knows something we don't.

 

[Stoneburner]

AMD is pretty much intel of 2 years ago. I need to find a cheap 680 board so i can upgrade.

 

[DrMrLordX]

Originally posted by: Duvie

90nm A64/X2 sckt 939 > 90nm A64/X2 AM2 > 65nm A64/X2 AM2.....

Actually, AM2 performance is better if you do a small memory overclock to cut down on latency. I don't know if a fully overclocked AM2 rig is really competitive with s939 (especially due to the fact that it's so easy to get 1 meg l2 chips for s939) but a stock s939 rig with DDR400 2-2-2-5 is a bit slower than a stock AM2 rig with DDDR2-800 3-3-3-10.

Brisbane is pretty disappointing, though. Hopefully Rev G2 will be better.

Originally posted by: SlowSpyder
I thought this was just a die shrink... I don't understand why everyone seems disappointed in the performance, it's 99% the same performance as the 90nm parts. Why would you expect performance to increase?

Mostly we're looking for something cheap and good to overclock, that's all. The promise of a successful die shrink made some hope for 3.6 ghz X2s on air (or better). That's not gonna happen with Brisbane Rev G1, but maybe G2 will fare better.

 

[Duvie]

Originally posted by: DrMrLordX

Originally posted by: Duvie

90nm A64/X2 sckt 939 > 90nm A64/X2 AM2 > 65nm A64/X2 AM2.....

Actually, AM2 performance is better if you do a small memory overclock to cut down on latency. I don't know if a fully overclocked AM2 rig is really competitive with s939 (especially due to the fact that it's so easy to get 1 meg l2 chips for s939) but a stock s939 rig with DDR400 2-2-2-5 is a bit slower than a stock AM2 rig with DDDR2-800 3-3-3-10.

Brisbane is pretty disappointing, though. Hopefully Rev G2 will be better.

Well I wasn't comparing overclocked or tweaked settings. What if we set the DDR on the sckt 939 to 500ddr which had been pretty common? Bottomline stock performance is worse and has continued to have gotten worse the last 2 steps for AMD.

Originally posted by: SlowSpyder
I thought this was just a die shrink... I don't understand why everyone seems disappointed in the performance, it's 99% the same performance as the 90nm parts. Why would you expect performance to increase?

Mostly we're looking for something cheap and good to overclock, that's all. The promise of a successful die shrink made some hope for 3.6 ghz X2s on air (or better). That's not gonna happen with Brisbane Rev G1, but maybe G2 will fare better.

 

[A554SS1N]

Originally posted by: SlowSpyder
I thought this was just a die shrink... I don't understand why everyone seems disappointed in the performance, it's 99% the same performance as the 90nm parts. Why would you expect performance to increase?

The question is, why does it decrease!

 

[RichUK]

Increasing the L2 cache is yet another stop gap between now and the release of their next generation core, K8L. This as we know is obvious, and is required if they are to seem competitive in the market. Albeit with a minor(?) increase in real world performance.

 

[BitByBit]

It still doesn't quite add up. They have increased the cache latency to allow for expansion, but insodoing created a situation in which the cache needs to be expanded to compensate for this latency increase. From what we have seen of the impact of going from 512KB to 1MB cache on Athlons, we can conclude that 1MB Brisbanes would have a hard time outperforming 512KB Windsors. Of course it's possible AMD has plans for 2MB per core, but that seems like a rather costly method of impoving performance.

 

[RichUK]

Originally posted by: BitByBit
It still doesn't quite add up. They have increased the cache latency to allow for expansion, but insodoing created a situation in which the cache needs to be expanded to compensate for this latency increase. From what we have seen of the impact of going from 512KB to 1MB cache on Athlons, we can conclude that 1MB Brisbanes would have a hard time outperforming 512KB Windsors. Of course it's possible AMD has plans for 2MB per core, but that seems like a rather costly method of impoving performance.

Yeah i was thinking along the lines of 2MB's per core too, because i agree with what you've said with regards to previous knowledge of 512KB Vs 1MB performance improvements on the Athlon K8.

Who knows? As long as it's a K8 derivative there?s not going to be much if any performance increase without heavily modifying the core logic.

So at the end of the day, AMD?s future focus could be on marketing BS, so to try and make out that increasing the L2 cache makes a respectable difference, and to thus generate a reason to allow AMD to rehash and bring out a new line of processes in wake of K8L.

I hope this is not the case, but it is a possibility to stay in the game until K8L, I think it is indeed quite viable. All AMD have to do is provide an improvement over the previous generation, even if it is a minuscule performance improvement. People (the average consumer) usually get roped into this marketing trash, where as people like ourselves can obviously see past it and make a more informed judgment.

I?m intrigued to say the least.

 

[Munky]

Meh, at least now I dont feel so bad for not having s939 65nm K8 cores.

 

[myocardia]

Originally posted by: DrMrLordX

Originally posted by: Duvie

90nm A64/X2 sckt 939 > 90nm A64/X2 AM2 > 65nm A64/X2 AM2.....

but a stock s939 rig with DDR400 2-2-2-5 is a bit slower than a stock AM2 rig with DDDR2-800 3-3-3-10.

The problem with your argument here is that it in no way involves the processor; it only involves the memory. I can pretty much guarantee you that if it were possible to run your BH-5 @ 1½-1½-1½-5 on an nForce 4 board, it would be at least as fast as any AM2 chip @ 3-3-3-10 timings.

 

[oRdchaos]

All I know is that I paid $95 for an AM2 3000+ and its been running on the Foxconn 590 SLI board at 9x324 (2.91 ghz) perfectly fine. Was going to upgrade to dual core, but this processor seems to game just peachy.

 

[DrMrLordX]

Originally posted by: myocardia

The problem with your argument here is that it in no way involves the processor; it only involves the memory. I can pretty much guarantee you that if it were possible to run your BH-5 @ 1½-1½-1½-5 on an nForce 4 board, it would be at least as fast as any AM2 chip @ 3-3-3-10 timings.

The reason why I made that argument is that memory is the only thing that changed between s939 and AM2 (so did the preponderance of 1megx2 l2 cache chips, though AM2 does have some . . . just not as many).

I will agree that AM2 chips do poorly at DDR2-400 and DDR2-533. DDR2-667 isn't that great on AM2 either.