Slashdot Mirror
VIA K8T800 Chipset Preview - Dual Opteron in Action
Mr.Tweak writes "It has been a long time coming but we are finally reaching the beginning stages of 64-bit mainstream computing. AMD has been the first to bring a 64-bit processor to the market with any true support in the Opteron. VIA is one of the key chipset companies supporting AMD64 and today TweakTown takes a preview look at their new K8T800 chipset with AMD Opteron 242 and 244 processors. 64-bit computing is boarding - don't miss the train!"
I doubt that using a 64bit processor would cause one of your devices "not to work". I don't see why only one of your program's wouldnt work.
The Television Wiki
I think the distinction between PC and RISC is quickly becoming obsolete, and so perhaps one should be including the UltraSparc IIe that's in the SunBlade 100/150, which retail for less than most new PC's.
By all accounts this is not the best of the current 64-bit chips, but I think it was the first to be offered in "PC-priced" systems.
The via board hinders the opteron. Instead of utilizing both on-chip memory controllers, it only uses one. You can see better results if you added a second set of memory banks for the second opteron to use.
There are more dual boards also including Tyans Thuder K8W. Rumor has it that nvidia will be coming out with a nforce chipset that will support dual opterons also.
Benchmarks
246 Benchmark
Overclocked 246
While it's slick that they had a dual proc board and all... none of the tests they used used the dual proc-ness of the system. They even indicate in their results that the second proc just threw overhead into the system.
;)
They've asked for help getting some dual proc benchmarking software. It would be great if someone could help them out. I'm really curious what that box is ACTUALLY capable of. IT's a goodly amount of horsepower with a reasonable amount of L2 cache with 64-bits of data-y goodness. It could make a heck of a "workgroup" size database server.... for a lot less than Sun's workgroup servers.
I think this next gen of procs (and their 64-bitness) is going to put another dent in Sun. First, lowend *NIX servers based on x86 put a huge dent in their pizza box market. Now, consumer grade 64-bit procs will probably start to eat heavily into their midrange market (like the 220R and that realm). In the big iron... well, that's contentious already. No need to to mess with them there
I'm down with that, as it were
The only applications that would currently require this are fairly high-end scientific/business processing.
Until the FPS/MMORG games start requiring it that is...
need a free COBOL editor for Windows?
Repeat after me: 64 bit processors aren't new. There's no new "computing fad" leaving the station. No new architectural wonder.
They aren't even new in desktop machines. I just threw out an Alphastation4 with a 64 bit 21064 from 1996 or something (nearly put my back out lugging the thing down the stairs. They built computers to last in those days). That was a competitor with the Pentium Pro IIRC. Many of the machines where I work were 64 bit ultrasparc before we started to go 32 bit x86.
That said, the new athlon does look pretty damn fast.
I'm using a dual Opteron (model 240, 1.4 GHz) for intensive scientific calculations. Our program, on the Opteron system, works 1.5 to 2 times faster than the 32 bit version that runs on a double Athlon MP 1800+ (1.533 GHz), depending on the algorithm.
The Tyan is an incredible board, agreed. However, the lack of AGP, huge size and price render it more suitable for server usage, and we were talking mainstream (i.e. desktop) usage afterall.
If server boards is your thing C'T (again) has a nice review on pages 158 onward. Interesting enough they compare the Opteron to Intel's Xeon Range. At least some industry sources seem to think that Intel's 64-bit offering is not up to scratch yet.
AI stuff can benefit, since Lisp on 64-bit architectures is especially nice. Lisp code on 64-bit architectures with 64-bit address space is significantly better performance than on 32-bit architectures, as you can implement many language primitive datatypes within the 64-bit quad, and still have a few high-bits for type-tagging, making a sort of virtual tagged memory architecture.
That means a dynamically-typed language with comparable performance to statically-typed.
Historically, one of the significant markets for Alpha processors was people running the Genera virtual lisp machine.
Yes, not all applications are or will be 64 bit. You're practically guaranteed not to ever see 16 bit programs again because a 16 bit processor can address only 64 KB of memory without some hack like segmented memory. 32 bit processor can address 4 GB of memory and for many programs, that is more than is needed for the task even in theory. It simply doesn't make sense to use 64 bit programs for such applications unless those applications do some heavy integer math. The 64 bit accessing mode is for programs that require that much memory that 4GB simply isn't enough. Currently, it's only databases that require that much memory.
The future is different, though. I predict that within 5 years, some computer games require more than 4GB of memory to run effectively. In addition to that, multitasking is becoming more and more important and if we want to keep linear address space then we need more virtual memory. Even today, a single threaded program can exhaust the whole 4GB address space with thread stacks (every thread has 2MB of virtual address space for its stack by default; start 1000 threads and you just lost half of the usable addresses...). If that program needs to access some devices via mmap()ed memory, then the address space comes as a limitation again. Think about a web server that runs with 1000 threads and serves 3 GB worth of content: it cannot use mmap()ed access to those files in a 32 bit processor or at least it must close some files before opening the rest. Latest P4 or Athlon CPUs could probably do the require computing but 32 bit addressing isn't enough for the task.
In short, we need at least 64 bit CPU and OS. It could be even wiser to run mostly 32 bit applications on top of that to preserve some L1 instruction cache (64 bit instructions take more space), however. The whole point of using Opteron instead of some other 64 bit processor is that it doesn't take meaningful performance hit running legacy code (= 32 bit code compiled for older CPUs) and it can run those legacy apps native without some software emulation layer. If all you need is a 64 bit CPU, there're better choices.
We really need only 64 bit addressing space but if you do that right, you get a 64 bit CPU as a result.
_________________________
Spelling and grammar mistakes left as an exercise for the reader.
"... because a 16 bit processor can address only 64 KB of memory ..."
... with a 16bit address space.
Uh... no.
Traditionally, a 16 bit processor means a processor with a 16 bit databus.
This doesn't restrict either register size or address space (ie number of address lines).
My 68K was a 16bit CPU... with 32bit registers and a 24bit address space.
My Z80 was an 8bit CPU
Likewise, it would be techically perfectly feasible to make a 32bit CPU with a 64bit address space - without hacks - or a 16bit CPU with an 8bit address space.
Modern 64bit processors are 64 bits for "everything". Well, in reality, a 48bit address space, and 40 address lines, at least for the opteron.
is being pedantic similar to being a smartass? If so, that's me.
And, actually, as another example, IBM is IBM's trademark, and ibm is not.
Error 666 - Satanic SCO code found in your Linux kernel.
I'm always shocked that Crafty (the very strong GPL chess program) isn't used to benchmark. It's easy to get a nodes/second count, it stresses the processor mainly with a bit on the memory, and it works great on multiprocessor systems ( even fake ones like the HT P4s).
I'm also kinda surprised MESA isn't used to benchmark more. A full OpenGL 1.5 equivalent in software would be even better -- one could be the FP performance of a processor through some serious tests.
For the opteron, though, the obvious test is simulations. If you're building a thousand opteron supercomputer, you're not gonna use it for chess or quake. It only makes sense to show both speed and scaling of the processor in what seems to be its current biggest market.