Athlon X2 or Intel Dual Core for Xilinx ISE tools ?

I haven't gotten around to updating that web page yet however I have done the benchmarks. The Core2 with 4M caches is generally about 30% faster on a clock for clock basis then the Athlon 64 X2 with 1M caches. NCVerilog is very cache sensitive so it benefits more from the larger caches of the Core2. With recordvars on I see a 69% improvement on a clock for clock basis for NCverilog on a Core2 vs the Athlon 64 X2. With recordvars off the gain is the same 30% that I see with the Xilinx tools. One more thing, the Core2 is very overclockable. I'm running my E6700 (2.66MHz) at 3GHz and it's been completely stable under heavy load. I've been running it like that for a couple of months, and I've had overnight NCVerilog regressions running on both cores most every night for all of that time. I wrote a system stress test to make sure that the system was reliable in spite of the overclocking, it's available here

Joshua,

thank you very much for these preciuos information. I'll try to convince my buying manager to get a E6700 with 4GB RAM and linux o.s. ...

Best regards, llandre

Joshua,

thank you very much for these preciuos information. I'll try to convince my buying manager to get a E6700 with 4GB RAM and linux o.s. ...

Best regards, llandre

Based on 10 years of experience I recommend to stay away from AMD based system for any serious computing. AMD based systems are low budget systems based on crappy chipsets and crappy components and because of that they lack a very important feature: stability. I've never seen an AMD based system survive a day in the office without crashing. Most people will tell you their AMD system _at home_ works perfectly. But tell me, is a PC at home used extensively for 10 hours straight? I don't think so.

If you want a computer get an Intel cpu based professional workstation from the business section from Dell or HP. You'll probably notice the price difference between the computer shop around the corner, but believe me, the price difference is worth having a PC that just works fine every day. Large companies buy PCs like these by the thousands for a good reason: a PC which doesn't work/crashes costs a lot of money.

--
Reply to nico@nctdevpuntnl (punt=.)
Bedrijven en winkels vindt U op www.adresboekje.nl

A note on RAM for the Core2s. The 2G DIMMs are now affordable, I bought a pair DDR2 800 2G DIMMs from NewEgg for $500. If you plan on using the largest FPGAs you'll want to be able to put 8G in the system so even if you don't want 8G now you should make sure that you use 2G DIMMs so that you can get to 8G in the future without having to throw out any DIMMs.

The speed of the RAM has no direct effect on performance because the Core2s are limited by their FSBs which runs at 1066MHz which exactly matches the bandwidth of a pair of DDR 533 DIMMs. I found that the difference between running the DIMMs at 533MHz and at 800MHz was only about 1%. However if you are planning on overclocking the CPU then you'll need faster DIMMs. The reason for this is that the RAMs and the CPU share the same source clock. The BIOS on my motherboard (Abit AB9 Pro) only offers 3 RAM speeds, 533, 667 and 800. These are actually 2X source clock,

2.5X source clock and 3X source clock. If you raise the source clock speed from 266 to 300, the E6700 has a 10X multiplier so 300 is 3GHz, the RAM will run at 600MHz, 770MHz or 900MHz. DDR2 667 or better will allow you to remain within spec even if you raise the source clock to 330MHz. If you by from HP or Dell you won't be able to overclock so you can live with DDR2 533 RAM, however if you buy a whitebox system then you'll want DDR2 667 or better.

You are completely wrong about this. I've been running three Athlon 64 systems 24/7, the oldest is three years old, the newest is about 18 months old. My systems are under heavy load doing simulations and place and routes. My compute servers have never crashed, the only problem that I've ever had is with X on my workstation. The Nvidia chipsets for the A64s are every bit as good as the the Intel chipsets, and the Nvidia chipsets integrate more functions such as dual Gigabit MACs with IP offload engines. the Nvidia chipsets are also 100% Linux compatible, you still have to do a little tweaking to get Intel 965 motherboards to work. The Core2s are better performers but there isn't any difference in reliability anymore.

My work system is a AMD 6400 X2 dual processor (Hypersonic Cyclone) configured with ECC RAM. It runs 24/7, usually with at least a simulation, synplify synthesis, or a place and route running on it. I have not had any problems with stability with this machine: it is reliable. I've had some windows issues (memory leaks and file handle limits), but nothing that points to hardware. This is my second AMD system. The previous one was a dual K7, which I also had no issues with other than the incredible amount of heat it produced and the attendant fan noise to keep it from melting down.

That may have had some merit as an argument 10 years ago, but it is totally at odds with most people's experiences since then. AMD has been the manufacturer of choice for serious computing since the Opteron's first came out - again and again, they have given more powerful and scalable than Intel's solutions, and the processors left stability problems behind with the K6 generation. There have been issues with heat - many of AMD's chips in the last five years have run particularly hot, and if you buy a cheap system then it's cooling system might not be good enough. And if you want to talk about motherboard and chipset issues, then Intel has far outweight AMD for problems in recent years - mostly because, until the Core 2, it has been rushing out everything it can in hopes of competing with AMD.

In my own experience, I have picked AMD on almost every occasion in the last fifteen years - first purely for value for money, and later for reliability as well. Were I buying a new machine today, I would probably go for a Dual Core 2, simply because of better value for money at the moment, although for a server I might pick AMD for stability (and for a four-core or more machine, AMD is the only realistic choice).

If your machines crash after a day at the office, you are doing something terribly wrong, and the processor is the least of your worries. Most of the machines I use and administer, at home and at the office, are AMD's, and most of them are never turned off. I have a server here at the office with a 300 MHz AMD K6 that has been running for around 8 years, and has only been off a half-dozen times for power cuts and a replacement power supply (this is probably a world record for NT 4.0).

And in the world of gaming, people run their machines for much longer than 10 hours at a time, and often with more demanding loads than any professional use - they generally choose AMD.

There is a reason why AMD captured a large proportion of the server market, especially for multi-core systems, despite Intel's entrenchment (and illegal and/or unethical behaviour, for which they are currently on trial).

This is a totally different issue. If you want a reliable machine, be prepared to spend money on it and get it from a reliable supplier. No one will argue with that. Don't buy AMD processors because they are cheap - buy the appropriate chip for the job.

There is a major risk for me missing the point here, but I'll give it ago anyway! The original question takled about the ISE and Dual Cores.... why would you need Dual Core for ISE? It does never use more than one of them anyway! OK it'll give your windows/Linux better response for other things while running ISE but it will NOT speed up the actual Synthesis+P&R action. Have any of you guys had ANY luck with using the "second" one within ISE? Maybe someone from Xilinx care to answer. If my memory serves me Xilinx made a statement that the ISE

8.1 would add support for Dual-Cores when running the tools. OK i gogled for that too..

I did NOT notice any improvement between 7.1 and 8.1 using my AMD

3800+ X2 when building a bigger project (a big project for me is a project that fills my S3 StarterKit 200K). It actually got slower! And now in ISE 9.1 (webpack) I did not see any signs of the normal ISE toolchain having any support at all for Dual Cores? I did a more through post here: act=ST;f=1;t=57 that you can read and comment on here in this newsgroup.

Because the fast (2.4GHz and above) Core 2 Duo chips get you 4MB of level-2 cache which, if you're using ISE on only one core and not doing anything very intensive in the background, is all available to the ISE process. Given how much of a boost people saw with 1MB-cache versus 512k-cache AMD chips, and how memory-intensive FPGA compilation is, I would expect 4MB to be distinctly useful.

Tom

The Intel Dual Core 2 chips are about the fastest available cpus at the moment, at least for sensible prices. I've heard several times that fpga tools like larger caches, which these chips have. The fact that you have two cores is almost incidental for ISE at the moment. However, there are always a few things going on in the background while waiting for a place and route - the second core can save you as much as a step up in clock speed even if you are not using the computer for anything else. There is also the future to look to - Quartus now has partial parallelism, and can use two cores for part of the work. It might only save something like 10-20% time at the moment, but we can look forward to competition between X and A in the race to get more parallelism out of their tools.

Also the machine I'm going to buy will be occasionally used by two people working on two different FPGA projects at the same time. Since they will run two different ISE istances, I expect they will be benefit from the dual core cpu. Is this speculation reasonable?

PAR in the Xilinx tools has been multithreaded for years as long as you run it on a *nix system.

Sorry for pulling a really old thread out for a follow-on, but I figured it would be better to keep the info together than spread across a bunch of threads.

I'm sure that you are correct that 4 MB would provide super results, but even 2 MB appears to be more than enough to allow a laptop to be overwhelmingly faster than a desktop that was considered unbeatable when it was brand new:

Pentium 4 at 3.2 GHz with 2 GB of RAM results:

Loading device for application Rf_Device from file '2vp40.nph' in environment C:\Xilinx\Xilinx9.1i. "benchmarkT_top" is an NCD, version 3.1, device xc2vp40, package fg676, speed -6

Device Utilization Summary:

Number of BUFGMUXs 9 out of 16 56% Number of DCMs 4 out of 8 50% Number of External IOBs 319 out of 416 76% Number of LOCed IOBs 319 out of 319 100% Number of RAMB16s 105 out of 192 54% Number of SLICEs 13951 out of 19392 71%

Starting Placer

REAL time consumed by placer: 11 mins 54 secs CPU time consumed by placer: 11 mins 36 secs Writing design to file benchmarkT_top.ncd

Total REAL time to Placer completion: 12 mins 4 secs Total CPU time to Placer completion: 11 mins 44 secs

Starting Router

Phase 1: 128932 unrouted; REAL time: 12 mins 42 secs Phase 4: 20607 unrouted; (1604) REAL time: 14 mins 32 secs Phase 5: 20609 unrouted; (0) REAL time: 14 mins 40 secs Phase 7: 0 unrouted; (0) REAL time: 16 mins 9 secs Phase 8: 0 unrouted; (0) REAL time: 16 mins 54 secs

Total REAL time to Router completion: 17 mins 8 secs Total CPU time to Router completion: 16 mins 39 secs

All signals are completely routed.

Total REAL time to PAR completion: 17 mins 58 secs Total CPU time to PAR completion: 17 mins 4 secs

Peak Memory Usage: 792 MB

------ compared to ---------

T7200 (2 GHz mobile version of the Core 2 duo) with 2 GB of RAM results:

Loading device for application Rf_Device from file '2vp40.nph' in environment C:\Xilinx\Xilinx91i. "benchmarkT_top" is an NCD, version 3.1, device xc2vp40, package fg676, speed -6

Device Utilization Summary:

Number of BUFGMUXs 9 out of 16 56% Number of DCMs 4 out of 8 50% Number of External IOBs 319 out of 416 76% Number of LOCed IOBs 319 out of 319 100%

Number of RAMB16s 105 out of 192 54% Number of SLICEs 13951 out of 19392 71%

Starting Placer

REAL time consumed by placer: 6 mins 12 secs CPU time consumed by placer: 6 mins 9 secs Writing design to file benchmarkT_top.ncd

Total REAL time to Placer completion: 6 mins 20 secs Total CPU time to Placer completion: 6 mins 14 secs

Starting Router

Phase 1: 128932 unrouted; REAL time: 6 mins 37 secs Phase 4: 20607 unrouted; (1604) REAL time: 7 mins 31 secs Phase 5: 20609 unrouted; (0) REAL time: 7 mins 34 secs Phase 7: 0 unrouted; (0) REAL time: 8 mins 24 secs Phase 8: 0 unrouted; (0) REAL time: 8 mins 48 secs

Total REAL time to Router completion: 8 mins 54 secs Total CPU time to Router completion: 8 mins 48 secs

All signals are completely routed.

Total REAL time to PAR completion: 9 mins 25 secs Total CPU time to PAR completion: 9 mins 4 secs

Peak Memory Usage: 784 MB

Your comparison makes interesting reading.

How do you think the AMD dual processor chips compare with Intel's. In many circles there seems to be an prejudice against AMD processors without any good reason.

It is just a prejudice - it's well over a decade since there was any conceivable technical reason for considering AMD processors to be cheap inferior processors. But remember, it is not the processor alone that make a system fast and stable.

The Intel Core 2 chips have more in common with the current AMD devices than they do with the P4 cores. The P4's were very much designed to get a high clock rate, and suffered from (amongst other things) very long pipelines and therefore poor performance when there were many branches. The Core 2, like the AMD cores, aim for lower clock rates but more instructions per cycle. For a lot of code, this turns out to be much more efficient.

There is not a huge difference between Opterons and Core 2 cores at the moment - Core 2 is somewhat ahead on single thread speed, and is a bit better value for money for mid range processors, while the Opteron scales much better with multiple sockets (it does not suffer from the FSB bottleneck). I expect AMD to take the lead again once their next core is out, but it's nice to see Intel with the top spot for speed after so many years - competition is good for everyone.

For FPGA design, popular opinion has it that the cache memory is particularly critical. I don't actually do much FPGA design, so I can't comment on that theory. But if you are looking for raw speed, with little regard for the lifetime of the pc, then it's easy to overclock the processor speed, but you can't add cache at that easily.