Of course, one thing that I am not considering in all of this is that the Opterons are essentially being "cut in half" as they are being used in 32-bit mode. I really need to run a 64-bit operating system and rerun the 64-bit version of Sandra benchmarks. I bet I see wildly better numbers. Still, all very interesting and it has led me to learn some interesting things. nitro
Why don't evalute these ECC 2- 3-2-6 modules if the mb supports them? http://www.zipzoomfly.com/jsp/ProductDetail.jsp?ProductCode=80097-19
I personally never liked Tyan's version of anything Opteron. I've always said the difference between AMD and Intel isn't the chips but the chipset and MB support. I've been very selective as to what I use with my Opterons and have used this board for 8 or 9 months now with higher memory timings and the latest stepping Opterons http://www.iwillstore.com/product.asp?pf_id=MADK8N&dept_id=21 . I use this memory with it http://www.ocztechnology.com/products/memory/OCZ_PC-3200_ECC_Registered_Dual_Channel_DDR b/c of the aggressive timings but have also been told this one will work as well http://www.newegg.com/app/ViewProductDesc.asp?description=20-145-473&depa=1 . Based on the board you have I wouldn't put too much worry into sandra benchmarks (b/c it's built for quad right?) as the only true measure of memory with this setup is to fill all sockets, then test . I finally found a review on my board and maybe you can extract some info from them here http://www.burningissues.net/hard/DK8N_review/index.htm . I think your going to have to know this board inside and out before you can conclude any benchmark, took me over a month playing constantly with different bios settings to get 100% of it's capabilities.
Thanks for the tip! When you say evaluate, are you saying that they will allow me to try the memory out and I can return it if does me no good? That would be great! I think I will give them a call Monday. nitro
Yeah, the verdict is still out. I will have a better idea when I get new memory with lower latency if it is the motherboard or the chipset or if in fact there is nothing and I am being premature in my assesment. Thanks for the tips. I am going to try much faster memory to see if it makes a difference. Although I am beginning to suspect Sandra's tests (I never got repeatable results on the memory bandwidth test) I don't follow what you mean when you say it is built for Quad. The memory timings, AFAIK, have nothing to do with the CPUs. I believe Sandra uses threads to do the tests, but assume it then normalizes the results. However, you are right in one way, and that is the one reason I am double confounded. One of (if not the whole reason) for the AMD Hypertransport is to reduce memory contention amongst CPUs. That is why I am really dumbfounded that the Tyan should score less in any memory/cache test than the Supermicro (Xeons) machine. Also, I don't understand why filling all the sockets would affect memory performance? The machine already has all the necessary sockets filled to run NUMA. You are right though if you mean I am testing with only two of the CPUs installed, but again, AFAIK, the mem test should not be impacted by the number of CPUS. The one thing that I am clearly handicapping the Tyan machine is that it is a true 64-bit machine, but I am running a 32-bit OS and 32-bit benchmarks. Thanks for the tip, I will look into it. I agree on the BIOS part. I may not even need to get new memory. Some BIOS's have a way of upping the voltage or even making certain settings "less safe." I am not sure if my memory can handle a higher voltage, but in either case the real question is, is it a deficiency in the memory, or does the problem run deeper, e.g., a crummy chipset. nitro
How can the machine have had any effect on my profit in trading when I am still testing and installing hardware onto it? No machine in the world will by itself add any money to a traders pocket simply because it is monstrously fast. You have to have some insights into the market which you then go and test. The creativity come first - that is the hard part. Spending $5k on a machine and writing software and dicking around to make it as fast as possible is the easy part, at least it is for me. That is what I am doing and will continually do - research - It may not pan out, c'est la vie. But La vida loca is better lived than taking the same old paths because the new path entails risk. nitro
Not having all cpu sockets filled does effect memory and cpu benchmarks, which really is not accurate of true performance because they are synthetic. There isn't anything on the web..... because the web is full of gamers type benchmarks and there isn't one website (maybe this one http://www.aceshardware.com/ , go to the tech forum and ask questions they are a very knowledgable bunch, and this site for real world benchmarks for plain vanilla boxes http://www.xbitlabs.com/ , they aren't like the paid websites who whore reviews for hardware companies) that has the info I know, which is based on an AMD engineer friend of mine. All I have at this time is info, the design for quad Opteron needs all sockets filled to get the memory benchmarks correct. In other words it needs a load. The dual design like I have is designed different and would blow away any dual Xeon memory benchmark. Check out the gif and see what happens under a load with a quad Opteron system.
Ok, I think I finally understand what you are saying. On the Tyan machine, you get the 128-bit wide path without having to have all banks filled. The Tyan has four banks per CPU and you get the full 128 effect if you fill the memory banks in pairs the rigth way. No need for all banks to fill. I talked to Tyan and they also said that there is a critical link in the HyperTransport that does not get utilized unless it is running in 64-bit mode. Another clue that this machine is meant to be run in 64-bit mode. Really, I have a realworld test that I will be able to do soon to compare the performance of the Xeons against the Opteron machine instead of using these inane Sandra tests that appear to be more and more meaningless to me. I will let you know what I see. nitro
I spoke to Tyan about these chips. They even guessed what the model Corsair is was before it came out of my mouth - LOL. They said that on several Opteron based motherboards, not just Tyans, the machine will not even post with these memory sticks in. They hinted that it has to do with the AMD 8100 chipset. Not that there was anything inherently wrong with it, but that overclockers do not necessarily like it. So much for that idea, but thanks anyway. nitro