There's been a lot of speculation that the performance of the Pentium 4 is tied directly to RDRAM. Therefore, what felt like a collective opinion settled on the matter: any non-Rambus DRAM solution for the P4 would ultimately be a failure. In Tom's tests, the PC133 version of the i845 rode a terrible roller coaster up and down the tracks of performance. Latency-sensitive business apps ran 7-12% slower, while bandwidth-senstive apps like Quake III Arena took a 23% nose dive. The platform's performance wasn't able to match the RDRAM-based 850. In fact, the 845 essentially failed in performance, but the reasons for its failure are not so cut and dry.
Take, for instance, this article about memory bottlenecks by Michael Schuette. He writes:
A more conservative estimate is that doubling the CPU clock rate will need a 50% increase of average memory bandwidth. In the case of a 2 GHz CPU, this will shift the point where a memory bottleneck becomes the limiting factor to approximately 1100 MB/sec average bandwidth. According to our calculations, this point will be reached with DDR 333 at 3:3:3 latencies. Keep in mind that this is an arbitrary threshold and applies only for current CPUs. With the P4 and its quad-pumped pipelines, this number will certainly double to a requirement of around 2.2 GB/sec sustained transfer.While Mr. Schuette was making some future predictions when he wrote the above paragraph, his original, conservative estimate predicts the P4's needs accuratelyand explains why the i845 with PC133 fails to best the higher latency RDRAM. Tom's own STREAM test shows the i845 squeezing an unbelievable 650 MB/s out of his 2:2:2 PC133 sticka value so high that it actually rivals the 1400MHz Athlon DDR system. With all that bandwidthenough to feed an Athlonhow did it still fail? Because even for its greater bandwidth, it was unable to feed the P4's appetite. Tom's 1.7GHz P4 Willamette, according to Mr. Schuette's postulation, would require an average of 935 MB/s. 655 MB/s is only 70% of that value, leaving around a 30% gap, which conveniently matches the 23% impact that memory intensive Quake III Arena suffered.
This certainly opens an opportunity for the DDR version of the i845 to come to the rescue. The i845 and P4 are able to successfully utilize 62% of PC133's maximum bandwidth. A similar result under PC2100 would result in 1300 MB/s of bandwidthenough even for the upcoming Northwood 2.2GHz model. While this result would be a good thing, it would also bring Intel back to a familiar situation: pushing chipsets that are better all-around performers than the processors mated to them. The Pentium 4's 20-stage pipeline continues to be Intel's own worst enemy, and I'm not so sure Northwood will change that fact.
We've also been wondering about the front-side bus (FSB) settings for the i845. Would it use an asynchronous 100(400)MHz bus with 133MHz memory, or would it be one of the first to sport a synchronous 133(533)MHz bus with 133MHz memory? Turns out it will use an asynchronous bus running at 100(400)MHz with 133MHz memory. If the Athlon's performance on an asynchronous bus is any indicator, this is a less-than-ideal solution. We might be able to squeeze some extra horsepower out of the i845 by overclocking the bus to 133(533)MHz. Here's hoping the Abit and Asus mobos, to name just two, will be capable of doing so.
From a techie/geek/fanboy perspective, the i845 is intriguing. This chipset's performance helps accentuate which applications are hungry for bandwidth (Quake III), which are more friendly to the P4's design (MPEG encoding)and even those simply limited by the P4's design (office apps).
From the perspective of price versus performance, the i845 probably won't even make a good budget PC. I'm not sure what Intel is thinking, but until the clock speed of the P4 ends up having a nice spread and their prices at least match AMD's, launching this product is crazy.
Intriguing, but crazy.