Editor's note - We've got a newer review up which has the newer Sandra numbers running on the Prestonias. We're still waiting to get the content creation benchmarks in but for those who are interested, the followup review to this article is now here:

The following was the original quick review that I posted on Jan 4th, 2001.

Well they came in today and I just got them installed and ran some benchmarks. It's a darn good thing that I made some quick benchmark screen captures of my 1.7 XEONs before taking them out (thanks for recommending that earlier b/c I was going to forget to do that). In any event, I had some interesting numbers at first but then when I thought about what was happening, I went back and ran the benchmarks again and wow! .. but then later, I goofed because the comparisons I saw were compared to Pentium 4's and not XEONs, so the final numbers are decent and in line with what I expected. I'll actually write up a longer review later but for now, here goes nothing:

First off, here are some pictures

Prestona 1.8 Pictures

1) Here are the 2 - 1.8 XEON Prestonia with 512K L2 Cache

2) Here's a closeup of one 1.8 XEON showing the data on the underside.

3) Task Mgr in Windows XP Pro - flagreen - you are correct!

Foster XEON 1.7 Benchmarks

So here are the original benchmarks for my Foster XEON 1.7's.
1) Here is a CPU Arithmetic Benchmark

2) Here is the CPU Multimedia Benchmark

3)Here is the memory bandwidth benchmarks

Prestonia XEON 1.8 Benchmarks - first pass

Ok, so here are the interesting #'s initially. I ran all the three benchmarks as above on my 1.8's and I had some really strange numbers. Initially I was baffled though I suspected it had something to do with the Hyper Threading aspect of these CPUs. Here are the benchmarks
1) 1.8 Prestonia CPU Arith Bench

2) 1.8 Prestonia CPU Multimedia Bench

3) 1.8 Prestonia Memory Bandwidth Bench

So at first glance, heck these processors performed worse than some 1.6 XEONs. What the heck was going on? Well let's think about these benchmarks now. I thought, "well these benchmarks had to have been multithreaded somehow.. but how much was it threaded?" Well after some analysis, if you look closely (hopefully you can see it) at the numebr of threads used for all three tests (it doesn't show up on the screen capture for the memory benchmarks), Sandra used 4 Test Threads for the benchmarks. I went back to the 1.7 XEON benchmarks and looked at how many test threads were being used there and guess what - if you look above at the 1.7 bench's, you'll see that it shows 2 threads used. So what does this mean you might ask? Remember - one of the downsides of Hyper-Threading is that if you have multiple threads that are basically doing the very same thing - say all threads are doing integer work - then a Hyper-Thread enabled processor could actually perform worse than a non-hyperthread enabled processor. Why? It's because of all the context switching that is occurring within the processor especially since all the threads are basically asking for the same resources. A good threading model would actually have say 1 thread do integer work while another thread does floating pt work thereby reducing the contention for shared processor resources. Make sense ?

Prestonia XEON 1.8 Benchmarks - Second pass

So it appeared that Sandra knew to create 4 threads (for the 4 "logical processors" on the Prestonia CPUs and they of course ended up making the numbers look worse. SO I went back to Sandra and changed the options for all the benchmarks. Basically, I just disabled the multi-threaded test option. By unchecking this, Sandra now just runs 2 test threads for each benchmark thereby allowing one thread to use each physical processor (with one logical processor presumably used but I'm not sure yet). Actually that is one point that I do not know - that is.. does Sandra actually schedule each thread to the physical processor or to the logical processor in this mode? Well in any case, we have a whole new set of numbers and these are mind numbing. Remember again, only 2 test threads are used here which is consistent with the benchmarks run against my 1.7 XEONs before. Nothing else is changed

1) 1.8 Prestonia CPU Arith Bench with 2 Test Threads

2) 1.8 Prestonia CPU Multimedia Bench with 2 Test Threads

So with 2 test threads, the numbers are much more in line with what I expect them to be but they also appear to blow all the other guys out of the water. At first glance, I was saying to myself "WOW, these guys are awesome!!". But then I went back after posting this and noticed the comparisons in Sandra were all wrong. They weren't comparing the values to MP systems which of course was because I turned off the multithreaded tests even though Sandra still used 2 test threads on the Prestonias. So look at the pictures above and throw out the comparisons..

Mix and Match Pictures to compare results - Third Pass

So given that Sandra kinda goofs up the reporting, I had to compile these numbers by hand and show them to you. You can verify the results by looking at the pictures for the 1.7 XEON benchmarks above. I just took them from there and put together by hand

But there's probably more to this and I'm going to have to go back and look through some of the Hyper-Threading slides from Intel to see if I can find out why these scores are what they are. Note, I don't know what Sandra is actually doing behind the scenes here so I am merely guessing as to what the thread pool looks like within Sandra. The floating point benchmarks are pretty good for the multi-threaded enabled tests.. but why? who knows for now.

Ok.. that's it for now. I'll write this up on my site in more detail later tonight when I get back home. I'm heading out for drinks and a small party. Good luck guys and sorry for the big picture sizes. I didn't have time to shrink them.