Dale,
I agree here, to a certain extent. I don't think that TBred will clock high enough for the limited amount of L2 cache to matter, but Barton should probably have 512KB L2 cache. The only way that I see for AMD's decision to cripple Barton with only 256KB L2 cache is if Clawhammer really does ship in volume in October.
If you can't sell out your capacity for 6 quarters in the row, with transition to .13u adding even more capacity, any idea of crippling performance of a processor to make it less desirable is a bad one. Barton with 512K L2 would be a performance notebook chip, (bigger L2 would justify higher QS rating). Barton without small L2 will command (I mean lose) price differntial that is much bigger than cost of adding 256K L2 on .13u process.
If AMD can make the transition to Clawhammer as rapidly as they made the transition from K6 to K6-2, repositioning Barton to occupy the same market segment as Intel's 130nm P-III Celeron did in Q4 of last year, while flooding the Asian markets with Morgans out of Fab25 and Appaloosa in a manner similar to what Intel did with low end Celerons in Q4 of last year, the decision to limit Barton to 256KB L2 cache could prove to be an excellent one. But these are big ifs.
Well, I doubt there will be any flooding of any market with Austin chips. The .18u Al process is going to run out of steam. This is what AMD should have done in Q2, Q3:
- Rename Palomino from AXP to Duron
- Sell Tbred with 512K L2 in desktop and mobile
- ramp down .18u Palomino Duron, transition capacity to SOI
- ramp up Barton (with 512K L2) to replace Tbred in notebook segment (possibly some in notebook)
- disable half of Tbred L2 (or salvage the ones with defects in half of L2) and sell it as Duron
- augment 256 Tbred Duron with chips from foundry (if there is demand
- forget Clawhammer, just do Sledgehammer with disabled HT links
The end of 2002 lineup would be:
- low end 256K L2 Tbred Duron from Dresden and foundry (desktop and notebook)
- notebook Barton with 512K from Dresden
- 1MB L2 Sledgehammer with 2 memory channel for desktop, 2 way server
If AMD can get the yields to the same level as on the .18u copper process, the blended average die size would not be any bigger than that of Dresden output today (a big if). Today, it is 120 mm^2 (and some change)
If we assume:
- 256K L2 Tbred 80
- 512K L2 Barton 100
- 1MB L2 Sledgehammer 150
The ASPs would match Intel's $150, and to produce the same revenue as in Q4 2001, all AMD would need to sell is about 4.5 million units. These products would be in sufficient demand that AMD could sell all they could produce, which IMO could potentially up to twice 1.5 or 2x the 4.5 million, plus output from the foundry.
The plan for 2003 would be to:
- tranistion Duron from 256K L2 Tbred to 512K L2 Clawhammer
- selling Sledgehammer DP and MP with cHT qualified
- notebook chips would continue to be 512K Barton and 256K Tbred (for low end)
Once again, I more or less agree. AMD needs to produce a "crippled" variant of the 90nm Sledgehammer for the desktop. Such a processor might have the same number of HTT links and DP capability like the DP Clawhammer, but be sold into the desktop market at the same peak selling price as Prescott. The 90nm Clawhammer would then need to be reduced to just 256KB L2 cache and marketed against P4 Celeron, perhaps being produced using bulk silicon for economy. But, even for this to work, AMD's transition to 90nm would have to take place within four month of Intel's.
This is going into the future, but I think you have it backwards. AMD needs a crippled Sledgehammer (Clawhammer), but later on, potentially, after the demand for high ASP demand has been satisfied. What you do is produce the product that has the highest potential performance ASAP, get the highest ASP possible, and move it downmarket later on. On 90nm front, it will be at premium originally, so the highest ASP parts should move there first, which is 1 MB (or higher for servers) L2 Sledgehammer and notebook Hammer would move there first.
It is not at all in AMD's interest to sell any $60 parts. AMD's desire should be to make (and sell) parts that command the highest premium possible. Today, it means 512K L2, 12 months from now it will be 1MB. I can't believe how AMD has missed the boat, and will end up being forced to sell the CPUs for sub-$60.
Joe |
