Ted,
There is a good discussion (related to Coppermine stuff posted on this thread) going on JC's page. Here is a copy of JC's comments on Coppermine, Athlon and AMD -
------------------------------
> For the first time, I have been reading the
> recent parts of the AMD thread at Silicon Investors.
> I found that there are several people who work
> for Intel and keep posting there. They are giving
> more and more clear hints about the Coppermine
> performance. According to them, Coppermine will
> be clearly superior to the Athlon.
Actually, the two people at SI who work for Intel are Process Boy and Tenchusatsu. Their positions have been that Coppermine would be *competetive* with
Athlon, not "clearly superior". You're probably grabbing that latter phrase from Elmer, Paul Engel, or Yousef, none of whom work for Intel.
> (Besides, the RAMBUS problems are said to be
> all worked out now.)
Which particular problems? Right now, all that needs to be contended to is the following:
A) PC800 DRDRAM isn't showing an incredible performance increase over PC100 SDRAM so far.
B) DRDRAM is tremendously more expensive than SDRAM
C) DRDRAM is more difficult to yield, and requires a heatsink (on the memory!) to operate
> The Coppermine release will end up to be
> a worst case scenario for AMD, which is not enough
> taken into account by the analysts now. Any fears
> AMD might have concerning Coppermine are more
> than justified, according to them.
Analysts have been taking the worst case scenario for AMD to heart for the past two years. The current belief of analysts takes into account Intel being at a slightly higher clock rate for the rest of the year, and with AMD producing fewer chips than they project. Believe me, AMD is doing *much* better than the worst case scenario.
> Is this (the Coppermine performance) the
> hint JC recently gave when he spoke about some
> "not-immediately-critical but not good"
> news for AMD?
I totally don't remember the context of that statement. I don't suppose you'd happen to remember which message number it was, would you? I can check through the archives.
Here's the beef: Comparing the 180nm PIII with 256K on-die L2 cache on 133MHz bus with PC800 DRDRAM against the 0.25um current core revision of K7 with 512K off-die L2 on 200MHz bus with PC100 SDRAM...
Coppermine would be equal or perhaps slightly faster in Win9x business applications.
WinNT business apps are uncertain, since K7 has a rather sizable lead even in Business Winstone for NT. But it is likely that Coppemine will be at least close
to parity in these benchmarks.
In benchmarks that are largely dependent on memory, like unix kernel compiles, Coppermine (like the K6-3) will likely be faster than K7.
In high end Winstone, K7 will smite Coppermine utterly.
In 3D rendering programs, K7 will destroy Coppermine.
Coppermine will come close to slicing in half the gap between Intel and AMD in specfp, but will come no closer than this.
Coppermine might be faster in specint.
In some 3D games that feature large texture loads that depend on memory performance, on-die cache has shown to be more important that floating point units.
Coppermine could perform better than K7 in Unreal, a game in which K6-3 outperforms the Pentium III.
In the majority of 3D games, K7 will be faster than Coppermine. In the floating point intensive games like Max Payne (btw, when does that come out or is it out now?) or Half Life or Quake 3, K7 will have a sizeable lead over Coppermine.
K7 will be faster than Coppermine in rc5, despite not having its own optimized core. Turnwise, Coppermine will probably be faster than the K7 in the very memory constrained SETI@home program.
Ummm, did I miss anything else?
Okay, conditions under which these results will change:
A) Future K7 models offer slower L2 cache. This is possible. If SRAMs were unavailable over 333MHz in volume, AMD would have to reduce their L2 divisor to 3 or perhaps a milder 2.5 for a 700MHz part. No clue how this would affect performance, though it's good to remember that 3D stuff will be affected less than "business" benchmarks.
B) Future K7 systems use faster memory. I believe that a jump from PC100 to PC133 would positively affect the K7 microarchitecture moreso than it would the P6 microarchitecture.
C) K7 gets better performing core revision. This sort of thing brings to mind CXT horror stories, but you should also recall the <sarcasm> awesome performance of the K7 previewed by Thresh </sarcasm> back when it was revision B1. Remember how everyone complained about K7 fpu sucking, and only a
few people DARED suggest that it was because it's an early core revision? Well, now fpu performance is a full 50% or more higher than we saw then. Wouldn't it be nice if there were some unforeseen bits that could be improved even byond this? Perhaps in their haste to get the part ready, they weren't able to implement some features they were planning on. Or maybe there are small performance bugs there that are being fixed as we speak. Yeah, I know, option C is a
long shot, but this is definitely worth mentioning because changes like these could totally overthrow the balance.
Okay, more notes:
Coppermine will come out at 733MHz, as reported by Intel a week or two ago. If Coppermine were to come out anywhere below this, it would be considered a disaster on Intel's part. AMD will have a 700MHz part out, probably out October 4th (microProcessor Forum, around the same time as AMD's next round of price cuts, and *possibly* also the release date for higher cached versions of the K7). 750MHz is anyone's guess, but giving the immense clockability of the
Athlon (the 550MHz Athlons seem to enjoy running at 700MHz without voltage increments and without breaking a sweat, it's as if AMD's artifically downbinning), I wouldn't be too surprised to see an 0.22um 750MHz Athlon out at the same time as Intel's 733MHz PIII. In fact, I'd be slightly surprised if we didn't see one.
That's the scary thing. With a much larger process, the K7 is competing rather well with the 180nm PIII in terms of clock rate. This is why I'm rather amazingly ... what's the word? Bullish? ... optimistic about the K7 on 180nm.
Ummm...and that's all I have to say about that.
-JC
jc-news.com
------------------------------
Goutama |
