Intel's new launch just. . .er, launched

I hope both manufacturers can conjure up more performance headroom with their designs.

Right now, it looks like x86 performance is reaching what is hopefuly a temporary plateau at around 4 Ghz or the equivalent in single-threaded performance.

The 130 nm fx-55 seems to be hitting a point where it is thermally much more demanding with a tdp of 104W.

The 90nm might be better, but currently the chips don't clock as high as the established process.

Intel's going to resort to 2mb cache on the P4s, but given their cramped clock restrictions, they probably won't get much higher performance-wise than an equivalent 4.0 Ghz either in the near term.
 
I'd be interested in a P4 when they get 2MB L2 along with x86-64 support, but not before then. Otoh, AMD should work faster on getting DDR2 support for their CPUs, they always seem to lay back too much whenever they get ahead, and then they get overtaken and end up in deep trouble until they catch up again. Happens over and over, one would think they'd eventually learn from their mistakes but apparantly not...
 
AMD's primary shortcoming is being far smaller than Intel, coupled with poor brand recognition/perception amongst consumers.

They have never been at parity with Intel, not even when they had a consistently better product or value. Any hope of them changing this would require a much larger presence in the market, but capacity restraints and supply snafus will probably continue for a while. Intel might have a few problems, but they can far more easily absorb hits.

It's not a mistake in the sense that AMD is doing something it shouldn't when there are better alternatives, because AMD is very short of alternatives.

AMD has stated it is going to wait until DDR2 667 becomes more common, as they say the latency penalties of the lower speed grades make it a performance liability. From what I've seen of benchmarks on DDR2 at speeds like 533, I tend to agree.
 
3dilettante said:
They have never been at parity with Intel

Sure they have, several times even. Not from a marketshare POV of course, but there are other metrics than that you know.

AMD has stated it is going to wait until DDR2 667 becomes more common

...Which is bullshit and a mistake. You can't sit on your heels and wait in this business, you need to keep the initiative. It doesn't really matter if it gives an actual benefit NOW, it has to SOUND as if it gives a benefit (*cough*netburst*ahem*), and you need to be ready once the market shifts and it actually DOES give a benefit. AMD isn't even trying on either of these points, which is very dangerous. They're not in a position to keep a following even if they miss the start of a trend. They need to be ready BEFORE it even starts, like with x86-64 for example.
 
but if switching to DDR2 533 causes AMD to lose part of its enthusiast base, that's going to REALLY hurt its image.
 
The Baron said:
but if switching to DDR2 533 causes AMD to lose part of its enthusiast base, that's going to REALLY hurt its image.
Sorry, I'm a bit stupid today...why would switching to DDR2 533 cause 'em to lose part of their enthusiast base? :|
 
digitalwanderer said:
The Baron said:
but if switching to DDR2 533 causes AMD to lose part of its enthusiast base, that's going to REALLY hurt its image.
Sorry, I'm a bit stupid today...why would switching to DDR2 533 cause 'em to lose part of their enthusiast base? :|
because A64s are very sensitive to latency, and the really high latency of DDR2 versus DDR1 would probably mean a performance penalty in most things since the bandwidth increase isn't enough to compensate for the latency shenanigans.
 
The Baron said:
digitalwanderer said:
The Baron said:
but if switching to DDR2 533 causes AMD to lose part of its enthusiast base, that's going to REALLY hurt its image.
Sorry, I'm a bit stupid today...why would switching to DDR2 533 cause 'em to lose part of their enthusiast base? :|
because A64s are very sensitive to latency, and the really high latency of DDR2 versus DDR1 would probably mean a performance penalty in most things since the bandwidth increase isn't enough to compensate for the latency shenanigans.

actually the athlon 64s perform very well even with looser timings .

ddr 2 shouldn't affect them as much as it will be hidden with the memory controller .

ALso moving foward a dual core athlon 64 will not be happy with ddr 400 .
 
jvd said:
actually the athlon 64s perform very well even with looser timings .

ddr 2 shouldn't affect them as much as it will be hidden with the memory controller .
It shouldn't hinder them much, but they won't gain anything neither. Even going from single-channel to dual channel didn't do much for performance (less than 10% at current speeds). So as long as ddr2 memory is more expensive than ddr1 it makes sense not to change the memory controller.
That said, it might not take too long before ddr2-667 might be common and not that expensive - ddr2-533 is available at quite reasonable prices and obviously memory manufacturers are all ramping up ddr2 production.

ALso moving foward a dual core athlon 64 will not be happy with ddr 400 .
Shouldn't be that bad. There are some dual-opteron boards out today whose memory slots all belong to the same cpu. If you're not using a NUMA-aware OS the performance loss due to that is generally pretty much zero. And even with a NUMA-aware OS it should not lose too much performance - in the same ballpark as the difference between dual-channel and single-channel today.

How much memory bandwidth is "enough" for a specific cpu?
I like to do some unscientific calculation, based back on PII cpus. PII initially had 533MB/s of memory bandwidth available. When the 100Mhz FSB cpus were introduced, benchmarks showed very little improvement due to the increased bandwidth (PII-333 vs. PII-350). So assume 533MB/s was basically enough for a 400Mhz PII. Scaling linearly, this means the 1066MB/s the later PIII (which has the same IPC as the PII) had was only enough for about a 800Mhz PIII. BUT even single-channel DDR-400 offers 3.2GB/s - "enough" for a hypthetical PIII-2400. And the A64 doesn't really do a lot more work per clock than a Athlon XP or PIII, its advantage is primarily the lower latency memory access. So 6.4GB/s is quite a lot compared to the "core cpu performance" of a A64, if you put that in contrast to what the 533MB/s PII-333 had. (Rules seem to be a bit different for the P4 - my best guess would be because it has more aggressive prefetching, larger cache line sizes, and indeed can reach very high throughput when using carefully crafted SSE-2 code.)
But anyway, it wouldn't surprise me to see DDR2-667 dual-core A64 cpus about the same time as the DDR1 dual-core ones. AMD was vague but I'm sure they won't miss the appropriate timeframe. (Another thought: wouldn't it be beneficial to have two single-channel memory controllers instead of one dual-channel one (especially if you'd use DDR2-667) if you have a NUMA-aware OS? Or would that require a different pin-out?)
 
The Baron said:
but if switching to DDR2 533 causes AMD to lose part of its enthusiast base, that's going to REALLY hurt its image.

Why would they need to *switch*? It's no biggie having a memory controller that supports both DDR and DDR2.
 
There's a chance that current A64s do have ddr2 capability, if AMD doesn't want to make another socket transition to something like s939b.

Pincount-wise, there is little concern, though the pinouts are not the same.

However, the instant they put out an official DDR2 model, in marketing terms they have just devalued their entire previous product lineup. Unless DDR2 has reached widespread market acceptance, there is no point in losing margins for a fringe product.

Additionally, unless DDR2 gives a performance advantage, AMD will have to resort to the same thing it's always done, try to undercut Intel on price and value since they would lose out on performance compared to their DDR lineup (which would already be devalued), and their inconsistent lead on Intel's products would drop further.

Motherboard makers will now have to transition their products to a platform that has weaker performance than the things they have filling up their inventory, hurting their margins and their willingness to support AMD.
 
Current high end CPUs appear to be limited more by latency than by bandwidth considering how well socket 754 CPUs fair compared to socket 939. So going to trade latency for bandwidth seems like a bad idea right now.

When dual core chips goes mainstrem that picture may change, but not until then IMO.

Cheers
Gubbi
 
3dilettante said:
AMD's primary shortcoming is being far smaller than Intel, coupled with poor brand recognition/perception amongst consumers.

They have never been at parity with Intel, not even when they had a consistently better product or value. Any hope of them changing this would require a much larger presence in the market, but capacity restraints and supply snafus will probably continue for a while. Intel might have a few problems, but they can far more easily absorb hits.

It's not a mistake in the sense that AMD is doing something it shouldn't when there are better alternatives, because AMD is very short of alternatives.

AMD has stated it is going to wait until DDR2 667 becomes more common, as they say the latency penalties of the lower speed grades make it a performance liability. From what I've seen of benchmarks on DDR2 at speeds like 533, I tend to agree.

But why no faster ddr speeds? PC3700 seems to be very available(and pc4200 has a decent amount), and nearly all existing athlon 64 motherboards will reach at least 233mhz fsb, and most can hit a 1ghz hypertransport bus with no problems.

BTW, amd would have to create entirely new chips for ddr2, they've already got socket 754 and 939 segmenting the market, do they need even more?(someone should make a combo ddr/ddr2 board, I think there used to be combo sdr/ddr boards)
 
JEDEC only has a standard for DDR400. I don't know of any serious interest by the organization to implement a niche standard for higher speeds that would also weaken the already delayed transition to DDR2.

AMD doesn't officially support any configurations that differ from the JEDEC standards.

I don't think a combo ddr/ddr2 board would be viable, assuming AMD's chips either currently or will soon be able to support both memory types. For one thing, the pinouts and slot keying of ddr and ddr2 are different. Not only could they not use the same slots, but some weird and possibly unstable stuff would have to be done to switch around the various bit lines for there to be electrical compatibility on a common socket.
 
another mb of L2 cache wont save the pentium from the beating it takes in every important category. its faster by a few seconds in some encoding...wow thats a selling point!!!!

pentium 4 + and -

+hyper threading
+good at encoding

-a64 is MUCH MUCH faster in 95% of scenarios
-a64 is cheaper
-a64 is cooler
-a64 is more future proof

basically the a64 is better than the pentium in virtually every area. have to wonder why anyone would buy a p4 based system these days.
 
Back
Top