With the advent of dualcores and development of workstation-class motherboards for them, there's little reason (as of now) for the enthusiast to consider the Xeon platform any longer. So, I am retiring this thread. The info's still good, but its future's going nowhere.
Note: this thread deals with P4-generation Xeons only
The CPU
What are Xeons?
Xeons are Intel's CPUs for multiprocessor systems. It used to be that regular Pentium CPUs could be run in SMP, but the only way have dual P4s is to have dual Xeons. The only differences between Xeons and P4s of the same core are a different socket and the SMP capability. Occasionally, Xeons will bring new core enhancements to the market before regular P4s, such as HT, EM64T, etc.
Why go with Xeons?
Why bother now with dualcores?
What are the differences between P4 Xeons?
There are two main iterations of the P4 Xeon, the DP and MP. DPs are "normal" 2-way Xeons, which are what are generally available. MPs are for 4- and 8-way systems, and cost thousands of dollars- they're completely out of place for us.
The cores are:
The sockets are:
All Xeons have a downwardly unlocked multiplier- you can lower it freely below stock, but can't raise it above. However, rumor/speculation over at 2CPU indicated that there might be a split second on boot where the high, stock multiplier is in effect until the lower one comes in- you might have to have enough voltage to get over this short hump so that you could get at the stable lower multi when overclocking.
How does the cache affect things?
Cache is very important to the Pentium 4 "Netburst" architecture, and with the extra options in some Xeons, should be considered a bit more. Performance boosts depend on the application. Photoshop seems to get the largest boost, followed by some renderers, like 3DSMax, and games.
L1 cache is very small and contains mostly instructions and important immediate things, it's quite standard and not usually looked at. The Prescott/Nocona has a doubled L1 cache, which does bring a notable performance increase and helps compensate for the longer pipeline.
L2 cache is the standard cache that we look to for performance. It serves as a local bunch of RAM so that the CPU can access program information and other things far quicker for processing.
L3 cache is an extra, periphery level of L2. However, in the P4 architecture the contents of each cache are mirrored in each other, so the L3 is effectively cut by 512kb. More L2 is better than an extra chunk of L3, but remember that the Nocona/Irwindale has other architectural changes; it doesn't get performance differences just from its cache.
Hyperthreading vs two physical CPUs
Hyperthreading is nowhere near equal to having two CPUs. HT boosts specific applications that are written for it by jamming in more threads at the same time. Depending on the application, this may significantly increase performance. Two real CPUs not only boost SMP-optimized applications by an even more significant amount, but they also handle twice the load at the same time. This is what gives the “silky-smooth” feeling to SMP, and why it is so attractive- you can burn a DVD and do something else at the same time, for example. HT has a negligible effect IMHO on being able to handle extra load, which is the main reason why you go with SMP.
Which one should I get to overclock?
The Intel Processor Finder has info on all available steppings and speed grades of all these CPUs. If you're unsure on what S-spec corresponds with what stepping or don't know what you have, check here.
The most popular is a 2.4/533 MHz Prestonia with a recent stepping, such as M0. If you buy recent stock new from a place like Newegg, you're pretty much guaranteed to get a good M0. I don't see a reason to go with 2.66 or higher- the 2.4 gives you the same overclocking ability with a better chance for a high FSB and a better price. There is a 2.4 with 1 MB L3 roaming around that looks attractive, and it doesn’t appear to have any less overclocking ability (I’ve seen results w/ it at 3.6 GHz).
LV chips have started to pop up, too. A very rare, but incredible deal has been the 1.6 GHz LV (low voltage) Xeon on a 400 MHz bus Socket 604. These are C1/D1 steppings at ridiculously low prices, meaning that you might be able to get 2x 3 GHz for really cheap. They aren't as good as M0 2.4s, though. These chips have now sold out, but if you find them used, well, you might be quite lucky. SL6XK is what to look for, these are the D1 steppings- they sell around $250 for a pair, which means $125 per CPU, which is incredibly good for their performance. Now, higher-speed LVs are becoming popular, as further binning to get to LV status may make these chips more overclockable.
The next big overclockers are 2.8/3.0 GHz Noconas and Irwindales, E0 and N0 steppings, respectively. They haven't emerged as much because of FSB limitations- Canterwood boards are generally limited to around 225 MHz and the only overclockable e7525 board so far has been the IWill DN800-SLI. Once FSB restrictions loosen up- as they eventually did for Canterwood- these guys will fly. I've personally gotten a pair of 2.8s to 3.6 GHz on stock cooling with stock voltage (takes a bit more than that for Prestonia), and that's barely scratching the surface.
Overclocking expectations (maxes with full tweaks and cooling):
Prestonia
Nocona/Irwindale
The Xeon Database further down this thread lists the overclocks of some of our members, from which you can get an idea of what to expect from a particular setup.
Parts of this post, especially cache, contributed by zachj
Note: this thread deals with P4-generation Xeons only
The CPU
What are Xeons?
Xeons are Intel's CPUs for multiprocessor systems. It used to be that regular Pentium CPUs could be run in SMP, but the only way have dual P4s is to have dual Xeons. The only differences between Xeons and P4s of the same core are a different socket and the SMP capability. Occasionally, Xeons will bring new core enhancements to the market before regular P4s, such as HT, EM64T, etc.
Why go with Xeons?
- Price- A typical Xeon platform is cheaper than an Opteron setup, and highly overclockable Xeon CPUs can be found slightly cheaper than their Opteron counterparts. Also, Canterwood-based Xeon motherboards, the most popular right now, let you use your existing AGP cards and DDR, so you can still get excellent performance without the DDR2-PCIe upgrade expense. Athlon MPs are cheaper, especially good for folding
, but lag a generation behind. - Performance- For particular applications, especially for media encoding, photo editing, and rendering, a pair of Xeons offer the best performance available. Excellent real-world benchmarks can be found at GamePC. Of course, Opterons own the server and gaming segments, along with a few media apps. Xeons still hold their own in gaming, at least.
- Stability- Pretty much an older issue, maybe not an issue, but it might be an advantage to some people. Intel chipsets are known for stability, whereas AMD's are not. Older Athlon MP chipsets and perhaps early Opteron chipsets had varous hardware compatibility and other issues. Now, with the NForce professional and errata in the first shipments of the e7525, this may be changing.
- Motherboards- One of the biggest advantages Xeons have over Opterons and especially Athlon MPs. There are multiple generations of ATX-sized, AGP/PCIe, overclockable Xeon motherboards both with and without PCI-X, which means that you can get the right amount of capabilities you need at a good price (compared to other SMP motherboards) and without the hassle of the EATX form factor. There's also an ATX-sized SLI board (IWill DN800-SLI), for which there's no Opteron equivalent.
Why bother now with dualcores?
- Price- Kind of the same thing mentioned above, you can get SMP power without having to switch other system components. You can also buy one CPU now, and wait and save up for the next one.
- Performance- You don't get HT with anything but the Extreme Edition, whereas you can buy a pair of low-end Xeons with HT for around half the price. Also, dualcores only have 1 MB cache per core, whereas newer Xeons have double the amount, 2 MB per core. Depending on your application, this can make a big difference.
- Motherboards- Xeon motherboards generally offer more capability for I/O, with PCI-X and more PCIe, which lets you use higher-end storage adapters to take advantage of larger RAID configurations. There are a couple Supermicro LGA775 boards with PCI-X floating around, though.
- Overclockability- Xeons tend to overclock slightly higher than the P4s of the same core- since they're Intel's high-end chip, they're probably binned a bit. Also, two physical CPUs means that the heat output is not put on one heatsink, but two. With less stress on the CPU cooler, it may be easier to achieve lower temps and therefore overclock higher. While the highest achieved FSBs on Xeon motherboards are typically lower than those on regular P4s, the default multipliers are higher, so it evens out.
- Future CPUs- Rumors have Paxville coming out in a DP version for socket 604, which is supposed to be a dualcore chip with 2 MB L2 per core. Also, there’s Sossaman, a Pentium M derivative to become a sort of LV Xeon in the near future.
- Coolness Factor- Admit it, the allure of high-performing, exotic, or otherwise special hardware is part of being an overclocker. It's fun to play around with stuff not normally associated with the desktop
What are the differences between P4 Xeons?
There are two main iterations of the P4 Xeon, the DP and MP. DPs are "normal" 2-way Xeons, which are what are generally available. MPs are for 4- and 8-way systems, and cost thousands of dollars- they're completely out of place for us.
The cores are:
- Foster- found with 256k L2, HT on a 400 MHz bus. Akin to a Willamette- watch out and don't get these.
Fosters are not compatible with most any motherboard newer than the e7505 chipset. - Prestonia- found with 512k L2, HT on a 400 or 533 MHz bus. Akin to a Northwood- these are the most popular and first to be overclocked.
- Gallatin- an MP chip that slipped to DP. 533 MHz Prestonia with 1 or 2 MB L3 cache. Akin to a P4EE.
- Nocona- found with 1 MB L2, doubled L1, improved HT, EM64T, on a 800 MHz bus. Akin to a Prescott.
- Irwindale- Nocona with 2 MB L2. Akin to a Prescott-2M.
The sockets are:
- Socket 603- Fosters and 400 MHz Prestonias. Older socket, these CPUs, except for Fosters, are compatible in newer motherboards.
- Socket 604- Where the action is. 533 MHz Prestonias and 800 MHz Noconas/Irwindales. Just Socket 603 with an extra non-functioning pin to make it special.
All Xeons have a downwardly unlocked multiplier- you can lower it freely below stock, but can't raise it above. However, rumor/speculation over at 2CPU indicated that there might be a split second on boot where the high, stock multiplier is in effect until the lower one comes in- you might have to have enough voltage to get over this short hump so that you could get at the stable lower multi when overclocking.
How does the cache affect things?
Cache is very important to the Pentium 4 "Netburst" architecture, and with the extra options in some Xeons, should be considered a bit more. Performance boosts depend on the application. Photoshop seems to get the largest boost, followed by some renderers, like 3DSMax, and games.
L1 cache is very small and contains mostly instructions and important immediate things, it's quite standard and not usually looked at. The Prescott/Nocona has a doubled L1 cache, which does bring a notable performance increase and helps compensate for the longer pipeline.
L2 cache is the standard cache that we look to for performance. It serves as a local bunch of RAM so that the CPU can access program information and other things far quicker for processing.
L3 cache is an extra, periphery level of L2. However, in the P4 architecture the contents of each cache are mirrored in each other, so the L3 is effectively cut by 512kb. More L2 is better than an extra chunk of L3, but remember that the Nocona/Irwindale has other architectural changes; it doesn't get performance differences just from its cache.
Hyperthreading vs two physical CPUs
Hyperthreading is nowhere near equal to having two CPUs. HT boosts specific applications that are written for it by jamming in more threads at the same time. Depending on the application, this may significantly increase performance. Two real CPUs not only boost SMP-optimized applications by an even more significant amount, but they also handle twice the load at the same time. This is what gives the “silky-smooth” feeling to SMP, and why it is so attractive- you can burn a DVD and do something else at the same time, for example. HT has a negligible effect IMHO on being able to handle extra load, which is the main reason why you go with SMP.
Which one should I get to overclock?
The Intel Processor Finder has info on all available steppings and speed grades of all these CPUs. If you're unsure on what S-spec corresponds with what stepping or don't know what you have, check here.
The most popular is a 2.4/533 MHz Prestonia with a recent stepping, such as M0. If you buy recent stock new from a place like Newegg, you're pretty much guaranteed to get a good M0. I don't see a reason to go with 2.66 or higher- the 2.4 gives you the same overclocking ability with a better chance for a high FSB and a better price. There is a 2.4 with 1 MB L3 roaming around that looks attractive, and it doesn’t appear to have any less overclocking ability (I’ve seen results w/ it at 3.6 GHz).
LV chips have started to pop up, too. A very rare, but incredible deal has been the 1.6 GHz LV (low voltage) Xeon on a 400 MHz bus Socket 604. These are C1/D1 steppings at ridiculously low prices, meaning that you might be able to get 2x 3 GHz for really cheap. They aren't as good as M0 2.4s, though. These chips have now sold out, but if you find them used, well, you might be quite lucky. SL6XK is what to look for, these are the D1 steppings- they sell around $250 for a pair, which means $125 per CPU, which is incredibly good for their performance. Now, higher-speed LVs are becoming popular, as further binning to get to LV status may make these chips more overclockable.
The next big overclockers are 2.8/3.0 GHz Noconas and Irwindales, E0 and N0 steppings, respectively. They haven't emerged as much because of FSB limitations- Canterwood boards are generally limited to around 225 MHz and the only overclockable e7525 board so far has been the IWill DN800-SLI. Once FSB restrictions loosen up- as they eventually did for Canterwood- these guys will fly. I've personally gotten a pair of 2.8s to 3.6 GHz on stock cooling with stock voltage (takes a bit more than that for Prestonia), and that's barely scratching the surface.
Overclocking expectations (maxes with full tweaks and cooling):
Prestonia
- C1- ~3.2 GHz if you're lucky
- D1- 3.2 GHz, maybe a little higher
- M0- roughly up to 3.6+ GHz
Nocona/Irwindale
- D0- ??? early steppings, avoid them
- E0- 3.6 GHz+ Nocona
- N0- 3.6 GHz + Irwindale
The Xeon Database further down this thread lists the overclocks of some of our members, from which you can get an idea of what to expect from a particular setup.
Parts of this post, especially cache, contributed by zachj
Last edited:
