Modern processorsWell the processors above are all a few years old. Only the SL8GU supports 64bit instructions. By itself, 64 bits isn't a big deal. What makes it important is it easily allows for more than 4gb of memory, which even cheap cell phones now have. Another benefit is you can do 64 bit instructions, such as XOR which is useful for software raid. This can double the performance in theory of raid parity computations. Likely useful for software raid-5 and likely quite useful for software raid-6. The most important feature of modern processors is what Intel calls speed-step and AMD calls cool-n-quiet. What these do is dynamically change the frequency that the processor runs as a function of load. At low loads, the frequency is set to some low value. At high loads the frequency is increased. Since power is a function of voltage^2 * frequency, lowering the frequency saves on power. What is better is as the frequency is reduced so is the voltage. This saves big time on power. What is even better is the voltages at each clock speed isn't fixed in the chip, but rather stored in special hardware registers. With a free program called RMClock it is possible to adjust the voltages. Using this program it is possible to save significant amounts of power by reducing the voltage of your processor(s). Unfortunately, there isn't an equivalent program for Linux. |
I recently purchased the 64 core EPYC 7B13 (Milan). This is a used OEM peocessor, perhaps from Google. I paid $1100 on ebay for it. Running my own benchmark, it is 3.3 times faster then my older EPYC 7302. Its TDP is 225w. It is a really nice processor, but it is physically quite large.
Using Cinebench Release 23, I got a score of 58,160 which is quite good. One downside of lots of cores is many programs such as Handbrake don't take advantage of all of the cores. Over time, this issue is getting less sevre, but with 64 cores, it may take awhile for software to catch up
I recently purchased the 16 core EPYC 7302P (Rome). Running my own benchmark, it is 2.8 times faster then my dual Xeon L5640. Its TDP is 155w. It is a really nice processor, but it is physically quite large. The 'P' at the end means it only works in a single socket system. Most of my recent systems have been dual socket, but I can get up to 64 cores in a single socket, so I decided I didn't need a dual socket system.
Using Cinebench Release 20, I got a score of 7778, which is quite good. The EPYC Rome is roughly twice as power efficient as the best Xeon processors. This is nice in general, but mission critical in data centers. The EPYC Rome is no more than half the price of the best Xeon processors. This is a big deal to anyone who cares about money. All Rome processors support up to 4 TB of ram, and have 128 pcie gen 4 lanes. This is a lot more than any Xeon processor. Almost all Rome processors support 8 channels of memory, and have a higher memory bandwidth than Xeon processors. Any of these advantages is nice, all of these combined makes them a killer micro. This is why I bought a new one, when I usually buy older, used processors. I am very happy with the price and power consumption of my EPYC.
Using PowerMax I was able to draw 236 watts AC with my platinum efficiency power supply. At idle the AC power draw is about 80 watts AC. This is a very efficient processor. Also, the motherboard doesn't have a power hungry support chip for the CPU, unlike my Supermicro X8DTL which has an Intel 5500 (Tylersburg) Chipset, which has a TDP of 27.1 watts.
This processor's TDP is 95w, which is less than my old Phenom II 940. The FX-4100 was introduced in Oct 2011. The Phenom II 940 was released in 2009. The 940 has a slightly higher TDP with a lower clock of 3.0ghz. Despite this, it is slightly faster in my benchmark. The FX 4100 works well with the stock cooler. Like all FX processors, the processor has good ECC support, as well as decent performance. Performance is ok, but not great. It is 25w less than the Phenom II 940, and it is a AM3+ socket, but it isn't any faster. Fortunately the FX-8320 with 8 cores is much faster than my Phenom II 940.
I got this with an inexpensive motherboard as a package deal. This processor's TDP is 95w, which is less than my old Phenom II 940. It works well with the stock cooler. Like all FX processors, the processor has good ECC support, as well as decent performance. Performance is quite good, and I am very happy with it.
I got this with an inexpensive Asus motherboard as a package deal. This processor's TDP is 95w, which is the same as my FX-6300, but it has 2 extra cores. This is why it is an 'E' processor, which means efficient. I am using it to replace my old Phenom II 940, which I use for my main fileserver. The FX has twice as many cores, and runs colder than the Phenom. It is also under warranty, which has long since expired for my Phenom, which still works fine. I used a cooler for a 120w AMD processor, rather than the factory cooler. Still pretty compact and cool running. This is a great low cost, low power processor, which is ideal for computers that need ECC memory, such as my file server.
This is an unexceptional processor. The e3-1270v3 was introduced in 2013 and uses a 22nm feature size. Its TDP is rated at 80w. Let's compare it to the L5640 xeon, which was introduced in 2010 and uses a 32nm feature size and a TDP of 60w. The e3 is about 25% faster using my benchmarks and uses 33% more power. Not a big improvement in 3 years. These processors came out when intel dominated the x86 cpu market, and had little incentive to improve their parts. I bought these because the inexpensive server motherboard needed them.
This is a low voltage Xeon, with a feature size of 45 nm. Each Xeon contains two dies, each with two cores on it. This is not a very elegant solution, but it does keep the individual die sizes small, which likely helps with yield. All of the usual features are supported. Intel's TDP is 50w. There is no hyperthreading. For my benchmarks, 8 cores (2 sockets) is about 50% faster than my 4 core AMD Phenom II at 3ghz. The worst part about the L5420 is that is uses FB-DIMM memory, which takes about 10watts per stick, and really could use some kind of serious cooling. All of the memory hangs off of a memory controller, which isn't ideal. AMD built the memory controller into the CPU, and Intel eventually followed suit, after this Xeon.
This is a low voltage Xeon, with a feature size of 45 nm. Intel calls this family the "Gamestown" Unlike the L5420, there is a single die with four cores on it. All of the usual features are supported. Intel's TDP is 60w. There is hyperthreading, but it doesn't work nearly as well as it did on the older LV Xeons I have owned (the 1.6ghz, 2.6ghz and 2.8ghz single core processors). For my benchmarks, 16 cores (2 sockets) is 215% faster than my 4 core AMD Phenom II at 3ghz. Using 8 threads the processors are 16% slower than 16 threads. Using the fritz chess benchmark, 8 cores are 25% slower than 16 cores. Unlike the L5420 generation of processors, memory is directly attached to each socket. Even better, it is DDR3 (either registered or unregistered), not the dreadful, power sucking FB-DIMMs. The newer Xeons are generally more expensive than their AMD counterparts, but they perform faster and use less power. I could not have afforded the L5420 and L5520 when the processors were new, so I bought them used. The L5520 was $530 when first introduced.
Recently, one or both of my L5520s failed. I had purchased then used. I have never had a CPU fail before. I replaced the memory and motherboard. Only when I replaced the CPUs was a able to get the motherboard to POST. So CPUs can Fail. Fortunately, a pair for L5520s costs $7 shopped, which is stupidly cheap. I really should test the old CPUs individually, but it is a bit of a hassle... I was thinking of updating one of my fileservers if I can get a single socket motherboard cheaply. This would be an inexpensive way to reduce the fileservers power consumption.
This is a low voltage Xeon, with a feature size of 32 nm Intel calls the family "Westmere-EP". There is a single die with six cores on it. All of the usual features are supported. New are AES-NI instructions and some virtualization instructions. Intel's TDP is 60w. The clock speed is 2.27ghz, but a new feature is Intel turbo boost technology which allows the cpu to go up to 2.8ghz. The max speed I have observed is 2.66ghz. There is hyperthreading, but it doesn't work nearly as well as it did on the older LV Xeons For my benchmarks, 24 cores (2 sockets) is 84% faster than my L5520 processors (two 4 core, 8 thread processors) at 2.27ghz. Using 12 threads the processors are 16% slower than 24 threads. Memory is directly attached to each socket. Even better, it is DDR3 (either registered or unregistered). The newer Xeons are generally more expensive than their AMD counterparts, but they perform faster and use less power. I bought two of these processors for $50 used, on ebay. The L5640 was $996 when first introduced. I wish they would throttle down to a very slow clock speed when idle in order to further reduce power consumption, the way mobile CPUs do. The use about the same amount of power at idle as the L5520 processors they are replacing. The "Westmere-EP" is designed to fit in the same socket as its predecessor the "Gainestown".
I bought an AMD 2000+. It uses more power than my Pentium III processor. It is the first processor that I bought that actually consumes more power under load than under idle. Under idle conditions, it stays significantly cooler. It has a bare die visible on the top of the package. It also comes with 4 spacers which are placed near the corners of the package so that the heatsink can't easily break the die. AMD later learned their lesson, and used a heatspreader on top of the die to reduce breakage. My AMD 2000+ is still running fine.
I bought a notebook with this processor. I have been able to undervolt it using RMClock, and it runs quite cool. I couldn't lower the idle voltage much, but I was able to lower the high load voltage significantly.
I bought a notebook with this processor. I have been able to significantly undervolt it using RMClock, and it runs quite cool. At idle, the power consumption is down 50% from the factory default!
I built a computer with this processor. I picked it because the processor and motherboard had good ECC support, as well as decent performance. I have been able to undervolt it using K10stat, and it runs quite cool. The idle voltage is 1.0v and I was able to lower it to 0.675v. The full speed voltage is 1.35v and I was able to lower it to 1.xx v. Performance is quite spectacular, and I am very happy with it.
My first computer used a pentium 120 processor. It had passive cooling. Boy, those were the days. I still have the processor...
My second computer had dual pentium pro's. The pentium pro was a big leap beyond the pentium. It did really fast 32 bit operations. It had lots of fast cache. It was really expensive. The motherboard was the Intel Providence motherboard, which is how a server motherboard should be designed. The actual processors were rated at 166mhz, and had a 512mbyte cache. I bought them used, and they were overclocked to 200mhz. I was a bit worried at first, but I never had any problems with the cpu's being overclocked. I suspect that Intel had great yields of their Pentium Pro's and decided to mark some to run at 166 mhz, since they couldn't sell them all at the more expensive 200 mhz cost. This taught me that it is often a good idea not to buy the fastest rated processor, but rather get a cheaper one in the hope that it will overclock. This has proved to be a good idea, see Intel Xeon for more details. The p6-166 with 512kb of cache uses 27.5 - 35 watts according to Intel. The p6-200 with 512kb of cache uses 32.6 - 37.9 watts according to Intel.
I bought 2 Celerons (433 MHz) for an ABIT-BP6 system that I put together. They worked fine. I bought another ABIT-BP6 system, and 2 more Celerons (466 MHz). These have less cache than the corresponding Pentium II processors, and were much cheaper in their day. They also run quite cool, relatively speaking.
The Pentium III was a major step up from the Pentium II, mainly in clock rate. It comes in two basic flavors, slot-1 and socket 370. Slot-1 is much bulkier. I have an IBM system with with a slot-1 PIII which has a large heatsink and a shroud leading to the power supply exhaust. No actual fan on the processor. A good, reasonably low power processor.
My third computer had dual pentium III 933 'Coppermine' processors. It has 256k of L2 cache. The P3 was much faster than the pentium pro 200. It uses about 29 watts, which is pretty similar to the pentium pro 200. One limitation was the Asus CUR-DLS motherboard motherboard has the processors so closely spaced that I could only use 60mm fans on the cpu coolers, which made them a bit noisy.
Of the socket 370 Pentium III's, most have a bare die visible on top of the package. This is known as package FC-PGA. The newer ones have a heat spreader. This is known as FC-PGA2. Coppermine processors (the 933's and 1000's) can be found with both types of packaging. Needless to say, the FC-PGA2's are far less fragile when installing the heatsink.
My dual pentium III 933mhz machine was showing its age, so I decided to get a new computer. I was looking around, and found some some low voltage Xeon 1.6ghz cpus available. The Xeon 1.6 Prestonia is based on the pentium IV, and designed to support dual processing. I am not a big fan of the pentium IV or the Xeon due to it generating lots of heat (58w at 2.0ghz, 65w at 2.4ghz, 74w at 2.8ghz, and 103w at 3.2ghz), but the low voltage Xeon runs at only 1.3V (the bios reports 1.28v) and is rated at 30 watts (at 1.6ghz) , which isn't much higher than the 28 watts of the pentium Pro 200 or the 29 watts of the Pentium III 933. I was able to overclock the LV Xeon to 2.4ghz by increasing the front side bus from 100mhz to 150mhz without changing the cpu voltage. This was done by changing some jumpers on the motherboard and a BIOS setting. The Xeons have worked fine at the higher clock rate, and likely dissipate 40 watts (at 2.4ghz) each at the higher clock frequency. Like my Pentium Pro's, I think the 50% overclock was possible because Intel was making lots of fast processors, and likely labeled these as 1.6ghz even if they could run much faster. And they only cost $119 for two, which was a great deal!
Using the BIOS of the Asus NCCH-DL or the Asus PC-DL I can set the multiplier for the cpu from 12 to 16, where 16 is the default. For my PC-DL I ran the multiplier at 16 and the front side bus at 150mhz. This yielded pretty fast main memory access as well as fast cpu performance. For my NCCH-DL I have set the multiplier to 14 and the front side bus to 133mhz in order to increase main memory speed without making the processor run very hot. It is likely that I could increase the front side bus further if needed. It is possible to run at higher than 2.4ghz by increasing the cpu voltage. This has the possibility of shortening the CPU's life, as well as generating more heat. I have resisted doing this, in the interests of running cool.
Please see Intel Windtunnel HSF below for details about that heatsink/fan for the Xeon. Please see CoolerMaster 3U Xeon Heatpipe HS below for details about that heatsink for the Xeon.
My original LV Xeons had reached their maximum overclock (at standard voltage) of 2.4ghz, so I bought a pair of 2.6ghz LV Xeons on Ebay for $32. They are quite similar to the SL6GV Xeon LV 1.6ghz processors, being Prestonia, 130nm process, and the same voltage of 1.3v. I believe they came from Google, who ordered these special LV Xeons from Intel. It is unclear what the thermal design power is, but I estimate it is roughly 43w. The part number SL7HU doesn't show in any official Intel documentation. They are Prestonia, Family F, model 2, Stepping 9, revision D1, x26 and 100mhz. They work fine in my motherboard, and I have overclocked them to 2.86ghz where they are running fine. There are many faster processors around, but these are fast enough for now :-) I have been able to overclock them to 3.12ghz by increasing the front side bus from 100mhz to 120mhz without changing the cpu voltage. I know at 133mhz (3.46ghz) they won't work at the default cpu voltage.
Using the BIOS of the Asus PC-DL I can set the multiplier for the cpu from 14 to 26, where 26 is the default. For my PC-DL I am running the multiplier at 26 and the front side bus at 110mhz. This yielded pretty fast main memory access as well as fast cpu performance. So I have improved my overall clock speed from 2.4ghz with the 1.6ghz lv xeons to 2.86ghz with the 2.6ghz lv xeons. Not too bad for $32.
The SL8GU is based on 90nm technology, unlike the SL7HU and SL6GV. It has twice the L2 cache, 1mb compared to the slower Xeon LV's. It is based on the Nocona core, unlike the slower Xeon LV's which are based on the Prestonia core. It runs at a lower voltage of 1.1125-1.2v, unlike the slower Xeon LV's. It it unclear what the thermal design power is, but I estimate it is roughly 55w. It features the EM64T (64 bit) instructions. I just bought a pair, at a low price ($39) for my fileserver, as I heard that Solaris likes 64 bit processors for ZFS computations. Hopefully Linux software raid also does 64 bit XOR computations when run as a 64 bit OS.
Using the BIOS of the Asus NCCH-DL I can set the multiplier for the cpu from 14 to 21, where 21 is the default. The voltage according to the BIOS is 1.10 to 1.11v (which makes me very happy). I haven't tried to overclock these processors yet. At 100% cpu usage in my machine, the processors are running at 45C and 51C (upper processor is hotter). This is with the Intel 2U passive coolers with 2000 rpm fans on top. I have a pair of coolermaster 3U Xeon Heatpipes for Nocona I plan on installing later.
I got a new processor with my new ECS motherboard. I picked the E-4500 processor. It runs pretty cool, and comes with a pretty good Intel cooler. If I had a more expensive motherboard, I am sure I could overclock it, but my cheap ECS motherboard doesn't like to be overclocked at all. The processor automatically changes its cpu multiplier and voltage as a function of cpu load. At lower multipliers it runs at lower voltage, and as the multiplier goes up, so does the voltage. If you are running Windows, you can download RMClock, which is a program that can adjust the voltage at each multiplier. There are no similar programs for Linux. Using RMClock, I was able to set the voltage of the maximum multiplier to be the same as the minimum multiplier. I was unable to decrease the voltage at the minimum multiplier, which may be a limitation of the voltage regulator or perhaps the ECS motherboard. As a result, the processor uses significantly less power when loaded, and hence less heat and noise is generated. This is a lot more advanced a processor compared to my low voltage Xeon processors. The wonders of technology...
I bought a notebook with this processor. I have been able to significantly undervolt it using RMClock, and it runs quite cool. I am unable to lower the idle voltage likely due to the motherboard limitation. At 6x the voltage is 0.95v, and at 13x I lowered the voltage from 1.225v to 1.037v with complete stability. Performance isn't nearly as good as I would expect based on the reviews I have read.
If you have comments or suggestions, Email me at turbo-www@weasel.com
Created with gnu emacs and template-toolkit, not some sissy HTML editor.
No Java or javascript needed to view my web pages. They
both have significant security issues.