Pete,
Why do you think that Clawhammer will not have enough bandwidth to compete against NW? It may have less theoretical bandwidth, but that is not what CPUs actually get from the memory channel. Look at the memory bandwidth found by Sisoft Sandra for 2.4GHz P4 between PC2100 and dual PC800. PC2100 got 2GB/sec and PC800 got 2.4GB/sec, 75% of the theoretical bandwidth. Clawhammer will run on PC2700 and probably get 2.6GB/sec to PC800's 2.4GB/sec. Even dual PC1066's 4.2GB/sec will get only 3.15GB/sec in a real box which could be matched by a single PC3200 channel for Clawhammer.
AMD will have 2.7GB/s theoretical bandwidth. Intel will have 5.4GB/s theoretical bandwidth from memory modules (2xDDR333), 4.3GB/s from chipset to CPU, later, when they switch to 166 MHzx4, it will be 5.4GB/s.
If you look at the direction Intel forced benchmarking to move, it is in direction of benchmarks becoming a glorified stream benchmark. P4 will simply have better stream scores, and therefore, will do very well on the benchmarks, because the stream nature of the benchmark will hide the deficiencies of P4 in ad-hoc type of computing of real world apps.
Now the actual bandwidth is even less than this due to latency. Other than cache tests, which are not typically done any more, there is no standard benchmark to find this very important number out. Yet the efficacy of decrease the CAS setting has shown that even a 1 cycle decrease in latency yields to significant performance boosts. Going from CAS3 to CAS2 is more than equivalent to boosting bandwidth by 33% overall. An onchip memory controller has been shown to be the equivalent to decreasing the latency by 2 cycles.
Ok, suppose you are right, and lower latency nullifies the 100% P4 advantage in bandwidth. But isn't the lower latency one of the features that was supposed to take Hammer oer the top, a big part of improving IPC advantage further? If it is there just to nullify Clawhammers self inflicted defficiency, what is going to bring it over the top? What is going to nullify the probable clock speed disadvantage?
You seem to look at the off the paper based comparisons without taking a look at experimental based comparisons which trump the former.
Ok, how about specfp. Yet another stream like application, one where the strength of Athlon FPU (even compared to SSE) should bring Athlon XP over the top. Yet, it can't because P4 has 50% more bandwidth available. The gap is toing to grow to 100% advantage next month, when P4 will move to 133 MHz bus.
What really gets me about ClawHammer is that AMD will have gone through all the sweat and pain to get the dual channel memory (for Sledgehammer), and AMD is not going to take advantage of it. With AXP, I can fully understand the additional work and logistical problems involved in increasing FSB to 200MHz x 2. But with Hammer, the work is probably already done, resources have been spent, and AMD is just going to throw it away, limit it to low volume processor that is on the slow track.
Competitively, Clawhammer will just match the fastest P4 on Intel influenced benchmarks, forcing the prices of Clawhammer traditional 35 to 50% below the prices of P4. This will go on throughout Q4, Q1 2003, Q2 2003 until Sledgehammer is finally released in June (when cHT is finally fully qualified). In 3 quarters of self-inflicted pain, AMD will leave a $1 billion on the table, 10M processors x $100 difference ASP dual memory channel Hammer vs. single memory channel.
Also what are you calling a Sledgehammer? Is it a dual core or single core in your definition?
Sledgehammer will be single core, 1MB L2, 2 DDR channels, 3 HT links.
Clawhammer will be single core, 512K L2, 1 DDR channel, 1 or 2 HT links.
This is based on the information gathered, based on AMD presentations, demos, info posted on hardware sites.
A single core 1MB L2 would be about 124mm2
I think the information that has been floating around is that it will be in 140 to 150mm2 range. If you are right, and it would be only 124mm2, it is even dumber not to go with that instead of crippled 104mm2 ClawHammer.
As to Durons, you seem to forget that there is different markets mostly abroad. There Durons are selling. Since AMD is selling WW, you need to consider where even $20 is a big amount. This is where those integrated MBs with Durons sell.
Sell poorly. AMD has had a capacity in Austin to make 3x the number of Durons vs. what is sold now. The $20 difference is not an issue either, since rather than going with AMD, they go with Intel CPUs which are $50 more expensive. AMD could have just shifted the Austin CPUs up in the market desirability above the competing Intel solutions, obtain more sales, receive higher ASPs, but has not done so.
Also many Durons are sold into the mobile market. This is where Appaloosa makes sense.
This is a good point, probably the only valid point in defense of Duron.
Joe |
