If geeks love it, we’re on it

Socket 940 vs. 939

Socket 940 vs. 939

Supplied by AMD


AMD have been launching assault after assault intent on battering the competition into 64-bit submission. The introduction of the 939-pin processor has been eagerly anticipated by some and left others with questions. Many have gotten caught up in the drag strip of benchmarks and overlooked the new features this latest introduction brings. Socket 939 does bring support for unbuffered memory but does this mean the death of Socket 940? Cool’n’Quiet has received a cool reception but it’s a gem waiting to be polished. Socket 940 versus Socket 939. Which is the winning combination for you?

Socket 940 FX-53 and Socket 939 Athlon 64 3800+ look identical and for the most part they are. There is a cost of manufacturing advantage for motherboard makers to move to Socket 939 which uses a 4 layer PCB motherboard instead of the more expensive 6 layer motherboards for Socket 940. This was not the driving influence for AMD to pursue Socket 939. AMD needed the next generation socket to carry the processor through another growth curve.

ws_3800

fx_53

Socket 939 is obviously one less pin than Socket 940. It isn’t possible to snip one pin off a Socket 940 processor and have it function in a Socket 939 motherboard. Think of Socket 939 and Socket 940 processors as first cousins that can’t live in the same house. For now Socket 939 will not be able to function in multi-processor setups.

939_pins

 

Side by side

Processor
Athlon 64 FX-53
Athlon 64 3800+
Athlon 64 3700+
Athlon 64 3500+
Athlon 64 3400+
Pin Count
940
939
754
939
754
L1 Cache Size

(64 KB + 64 KB) 128 KB total

L2 Cache Size
1 MB
512 KB
1 MB
512 KB
512 KB
CPU Core Frequency
2.40 GHz
2.40 GHz
2.40 GHz
2.20 GHz
2.40 GHz
CPU to Memory Core Frequency
2.40 GHz
2.40 GHz
2.40 GHz
2.20 GHz
2.40 GHz
Integrated Memory Controller
128 bit registered
128 bit
64 bit
128 bit
64 bit
Dual Channel Memory Controller
Yes
No
Yes
No
HyperTransport Links
1
HyperTransport Spec
2 GHz
1600 MHz
2 GHz
1600 MHz
Effective data bandwidth
Up to 12.8 GB/s
Up to 14.4 GB/s
Up to 14.4 GB/s
Up to 9.6 GB/s
Up to 14.4 GB/s
Up to 9.6 GB/s
Process Technology
130 nm
Approximate Transistor Count
105.9 million
68.5 million
105.9 million
68.5 million
105.9 million
Approximate Die Size
193nm2
144nm2
193nm2
144nm2
193nm2
Nominal Voltage
1.5 V
Max Ambient Case Temperature
70 C
Max Thermal Power
89 W

It can be confusing as to which is “the best”. The 940-pin FX53 and the 939-pin FX-53 and 3800+ and the 754-pin 3700+ and 3400+ all run at 2.4 GHz. The notable difference are:

  • Socket 940 has an effective data rate up to 12.8 GB/s, Socket 939 is up to 14.4 GB/s and Socket 754 is up to 9.6 GB/s.
  • Socket 940 and 939 HyperTransport specification are up to 2 GHz while Socket 754 is up to 1600 MHz.
  • Socket 940 and 939 both have 128-bit wide dual channel integrated memory controllers while Socket 754 is only single channel and 64-bits wide.
  • Only the Socket 940/939 FX-53 and Socket 754 3700+ have 1 MB of L2 cache.
  • Only the Socket 940 processors require Registered RAM.

It’s pretty much six of one half dozen of the other isn’t it? Between the FX-53 Socket 939/940 processors there are only two differences. The first is that Socket 940 requires Registered memory and Socket 939 doesn’t. The second is that the Socket 940 FX-53 has less effective data bandwidth. It comes in at up to 12.8 GB/s versus up to 14.4 GB/s with the Socket 939 FX-53. The only obvious difference between the Socket 939 FX-53 and the 3800+ is 512 KB of L2 cache.

It appears the AMD cooks are varying their recipes by a smidgen of this and a dash of that. The “best” is what suits your needs…and budget.

Why Registered RAM?

Servers and mission critical systems is the explanation given most as to why to use Registered RAM.

Why?

Often this explanation is used as a “because I said so” statement and left at that. It has always been quoted but never really explained. Enthusiasts rejoice at the introduction of Socket 939 because the processor does not require Registered RAM as did its Socket 940 predecessor. Registered RAM is more of a specialty item compared to unbuffered “normal” RAM and as such it is more expensive and the choices fewer. First assumption by anyone would be to put Socket 940 on the road to obsolescence with the introduction of the Socket 939 processor.

Let’s begin with the memory itself. Registered refers to the registers onboard the RAM which hold data for one clock cycle before moving the data to the motherboard. This increases the reliability of the data. The “hold” for one clock cycle explains why Registered RAM is slower than unbuffered memory. “Slower” is a relative term like one dragster is “slower” than the other. Both move pretty fast.

Registered RAM can also be ECC Registered RAM. The “ECC” stands for “Error Checking and Correction”. Others use “Error-Correcting Code.” In both cases it’s the same built-in function of the RAM to automatically correct errors in data as it passes in and out of the RAM. These errors are known as “soft errors” or SER.

The reason for soft errors is incredulous but true. At the chip level a soft error occurs when the radioactive atoms in the chip’s material decay and release alpha particles into the chip. Those particles can “hit” a memory cell in the chip and force a change. As an example that memory cell may, at that time, be holding a binary code value of 1 and the particle collision may change it to a 0. Think of it like a billiards table where the cue ball strikes another ball knocking it out of position.

Chip level errors are extremely rare and there’s no worry of these particles bouncing wildly about ripping tiny holes in everything including the RAM, the motherboard, your legs, etc. These collisions are on the atomic scale and it’s more the physics of electrical interaction than anything else. RAM technology has increased so much making chip level errors extremely rare.

System level soft errors occur because of noise interference. This noise isn’t a stereo that’s too loud but “noise” of unwanted radiation from other components. Think of it like trying to tune in a radio station. The clear signal is what is desired but the static is the unwanted noise. That noise interferes with the ability to hear the song clearly. The song “corrupts” and words may be misheard or not heard at all. In the case of memory the data passing to and from the ram is the song and the noise “corrupts” the data. System level soft errors typically occur as the data is traveling to and from the RAM on the data bus but not within the RAM chips themselves.

How often chip level or system level soft errors occur is most difficult to pin down. These soft errors occur randomly and this randomness if further compounded by the variable of other system components and environment situations. The PSU may not have enough shielding. Someone may have parked a speaker next to the PC case. The cat may have rubbed up against the PC and discharged static electricity. Think on the microscopic level too. Did that trace on the motherboard leak a weak electrical signal and interfere with the trace next to it?

The variables never end and it’s safe to say that the odds are largely against soft errors but they do happen.

Stand alone systems typically use unbuffered memory. That is to say it’s not ECC Registered memory. On average the home PC isn’t being taxed as much compared to a server nor does it handle as much data over an extended period of time. A server is designed to “serve” many users; tens, hundreds or thousands at a time. This load means a tremendous amount of data is being passed about and, at times, during a 24/7 period. The odds of a soft error occurring are much higher. ECC Registered RAM will correct these errors thus everybody gets what they came for instead of glitches, crashes or BSODs.

Remember that any computer that is accessed by one or more users can be thought of as a “server.” This includes that PC that stores MP3s that are played back through the home network or a LAN party game server and so on right up to those corporate servers which are eternally blamed for the email not working.

A workstation is also another PC that may benefit from ECC Registered RAM. 3D programs such as Autocad, 3D Studio Max and Softimage manipulate a tremendous amount of data. If that data corrupts then frequent crashes result in downtime and that means lost time and money. 3D programs and even some 2D programs such as Adobe After Effects can render or “put together” a finished project taking a long period of time. This period may be over an hour or a day but the longer the duration increases the odds of an error. The last thing any graphic artists wants to see is a glitch half way through an 8 hour render. ECC Registered RAM is better suited to avoid those occurrences.

It is true that Registered RAM is slower than unbuffered memory. The reason is because Registered RAM modules have buffers on the address and control lines. These buffers along with the clock driver chip act like signal boosters. In layman’s terms data has a very weak electrical signal. If it had a stronger electrical signal then it could result in leakage and then it’s back to the soft error problem. That data has to reach a lot of places in a PC with a lot of memory such as a server which may have greater than 4 GB of RAM installed. The electrical “oomph” of the data isn’t enough to travel to all the places in the ram module without degridation…aka errors. The buffers give those signals a helping hand. The buffer and clock driver operation requires more time to “do its thing.” The time needed can be measured in nanoseconds and it would be difficult for the average user to tell the difference in the “feel” and “speed” of the PC given two matching setups; the only difference being one used unbuffered memory and the other used ECC Registered. Benchmarks would show a difference but, as will be seen later in this article, that difference is very small.

That’s an extremely simplified view but it is adequate for this level of explanation. This is a reason why motherboards with 2 and 3 DIMMS are common. Less DIMMS do cut down on the overall cost of manufacturing the board but a move to higher memory capability may cross that threshold between requiring Registered memory instead of unbuffered.

That threshold is changing as memory technology and dependability increases. Unbuffered memory modules at the 512 and 1 GB level are common which means that a system with a whopping 2 GB of unbuffered memory is easily attainable. Push to the 3, 4 or 8 GB level and it’s safer to use ECC Registered memory. It used to be less a year or two ago and it will be more in years to come. Technology marches on.

As a final point of conjecture it’s important to know that Registered modules can be buffered and unbuffered. The difference is the inclusion of buffers hence the term. ECC also is an additional feature. For the most part the most common types of memory are ECC Registered Buffered memory, ECC unbuffered and “normal” unbuffered memory (non-parity).

So why use Registered RAM?

Consider using Registered RAM if the goal is to have a PC:

  • That has a large amount of memory. Start considering it at the 3 GB level.
  • That serves another regardless of the amount of users.
  • That supports 2D or 3D data intensive programs especially if those programs can take a long period of time to complete a task.
  • That is used wherever data integrity is of critical importance to the end user.

So now it’s clearer to see that a Socket 939 FX-53 processor would make for a formidable gaming system but a Socket 940 FX-53 would make for a very powerful single processor workstation. Each has their useful niche and, when all things are considered. The choice must be made between saving on purchase of a system that uses unbuffered memory to only spend that savings in downtime and re-done projects. The variable of Opteron has not been introduced at this time.

Cool’n’Quiet

Cool’n’Quiet is now a feature on AMD’s Athlon 64 processors excluding the FX and Opteron lines. Opteron models have been introduced with new low power models but that’s another topic altogether.

Cool’n’Quiet is a function of the processor, when combined with a motherboard that supports it, to automatically adjust its own speed and power consumption based on load. Cool’n’Quiet is an overlooked feature of recent AMD processors and, while still in its early adoption and not without the occasional anomaly, has potential.

Since January of 2000 Intel has had its own version called SpeedStep which was introduced primarily for their notebook processors. SpeedStep kicked in when the notebook was running on battery. The processor adjusted its bus ratio to run the processor at a lower frequency (lesser speed) to conserve batter power thus extending battery usage duration. SpeedStep can be set to:

  • Manual – always run at the highest performance state
  • Constant – always run at the lowest performance state
  • Adaptive – performance state chosen according to CPU demand
  • Degrade – starts at the lowest performance state then uses additional linear performance reduction as battery discharges

Intel P4 desktop processors have a TCC or Thermal Control Circuit “that will attempt to reduce processor temperature by rapidly
reducing power consumption when the on-die temperature sensor indicates that it has
exceeded the maximum operating point.” Basically the P4 runs flat out until it overheats.

AMD’s Cool’n’Quiet, unlike Intel’s similar offering, isn’t just for notebooks and here’s what it does for desktops.

  • Cool’n’Quiet cuts processor speed when the processor is at idle.
  • Cool’n’Quiet cuts processor voltage when the processor is at idle.
  • Cool’n’Quiet instantaneously ramps up processor speed and voltage in accordance to CPU load.

Let’s talk about the benefits. The first is heat output. A processor that runs at lower voltage outputs less heat. Less heat means that a thermistor controlled CPU cooling fan can reduce speed and thus noise level. The second is power usage. A processor that runs at a lower voltage consumes less power and this power consumption saving adds up on the monthly bill for a business; small, medium or large. Basically it’s the reverse of the P4 Thermal Control Circuit; it runs flat out when it needs to otherwise it backs off to reduce power usage and heat output.

 

Cool’n’Quiet Real world numbers

killawattcuangle

Kilowatt usage measured by a P3 International Kill-a-Watt device. (0.2% accuracy). Comparative performance tests featured UT2003 Flyby benchmarks. Aquamark benchmark was executed at 640×480, 800×600, 1024×768, 1280×1024 and 1600×1200 screen resolutions at four levels of graphic settings: AA off & details low, AA off and details high, 16x AA & details low and 16x AA & details high. Sisoft Sandra file system benchmark was run 5 consecutive times on an 80 GB Western Digital hard drive with 10 partitions. (Floppy and external drive tests disabled.) The remaining time after programs completed was inactive and results were recorded 6.5 hours from the beginning of the tests.

My own hydro company charges $0.05770 per kilowatt hour. The test PC without Cool’n’Quiet enabled used 1.05 kWh. Consider that it is powered on for an average of 4 hours per day over a one month (30 day) period. That comes to an extrapolated cost of $7.27 that I pay to operate that PC. (An amazing figure as it’s half the overall usage charge.) To see average kilowatt per hour usage of 4 different power supply configurations without Cool’n’Quiet look to our 4 PSU roundup.

The same battery of tests with Cool’n’Quiet enabled yielded a 0.76 kWh usage. The extrapolated cost over the same time period is $5.26.

costsaving

So what? A difference of $2.01. For an example business situatio multiply that monthly saving per computer times 100 computers over a year’s usage and the savings work out to be $2412.

It adds up.

In the following examples the test system idled for a period of 15 minutes before executing 25 repetitions of Sisoft Sandra CPU Burn-in (Arithmetic/Multimedia benchmarks). After another 5 minute idle period Unreal Tournament 2003 Citadel, Antalus and Asbestos flyby benchmarks were executed.

Cool’n’Quiet reduces processor temperature with the first rise indicating the 25 repetitions of Sisoft Sandra CPU Burn-in and the second spike indicating the flyby benchmarks of Unreal Tournament 2003.

coolnquiet_cputemp_graph

The ramp in voltage between idle and load is very significant with Cool’n’Quiet enabled.

coolnquiet_voltage

Another way to look at the processor temperatures is between idle and load with Cool’n’Quiet enabled and disabled. Most noticeably is the near 10 degree drop in idle temperatures with Cool’n’Quiet enabled. It even affected load temperatures.

coolnquiet_cpu_temps

With a drop in processor temperature comes a drop in fan speed (graphed) and therefore fan speed noise.

coolnquiet_fanspeed

Processor usage is, for all intents and purposes, identical whether Cool’n’Quiet is enabled or disabled. It’s proof that when the processor is under load it is functioning at peak efficiency.

coolnquiet_cpu_usage

Cool’n’Quiet is simple. When it’s not needed, such as during tasks like email or word processing, the processor backs off. When it is needed, such as during multimedia design or gaming, the processor speed and power are right there. Cool’n’Quiet is also dynamic, adjusting in a curve rather than a one level or the other setting. Mind you the curve is rather steep as maximum core speed, voltage and multipliers are usually reached when the processor approaches 50% usage. Cool’n’Quiet maintains the lowest setting when CPU usage is below 10%. Quick jumps are seen between 30-50% CPU usage.

Impact on performance

Nil. Zilch. Nadda. Nope. No siree Bob. Don’t believe me? Unreal Tournament 2003 flyby benchmarks showed little to no change with Cool’n’Quiet enabled or disabled.

cnc_ut2003

 

Aquamark tests showed the same negligible differences

aquamark_offofflow

aquamark_616high

EVP: Enhanced Virus Protection

AMD marketing has their teeth into another route to our pocketbooks with an upcoming feature available with any AMD64 processor. Enhanced Virus Protection (EVP) is a hardware feature that sets aside an area of system ram to contain certain types of virus’ and malicious executables. Make no mistake; this feature is not designed to replace Antivirus software but it is available to enhance protection within the computer.

EVP works by the processor setting aside a portion of system memory. How much memory is unknown but it will be a very small amount. That portion of memory becomes a containment room of sorts. A virus or malicious executable that enters the PC through an open port is automatically sent to that portion of memory and contained. Anything within that memory area cannot be executed. Essentially the bad guys get their fuse snipped on the bomb. The contained virus is then dealt with by the virus protection software.

EVP must work in conjunction with software written for EVP. Windows SP2 will have EVP features for enhanced virus protection and it remains to be seen if other vendors such as Symantec explore the EVP route. If it catches on and can make money then there’ll be no doubt they will. EVP DOES NOT protect a user from every virus nor is it to be thought of as virus protection hardware. It will not protect a user who accidentally opens up that “RIWANDA BANK NOTIFICATION OF 1 MILLION DOLLAR DEPOSIT” email that contains a virus. Nor will it protect you from sticking your mouse where it shouldn’t be.

EVP is said to enhance virus protection and not replace it. If effective then EVP could show saving in both lost manpower time and money.

Benchmarks.

The test systems.

  • AMD Athlon 64 3800+ Processor (32-bit mode)
  • ASUS 8KV motherboard
  • ATI 9800 PRO 256 MB Video Card Catalyst 4.2 drivers (Application preference ticked for Anti-Aliasing and Anisotropic Filtering in both Direct 3D and OpenGL, VSYNC disabled BIOS AGP aperture set to 256)
  • 2 x 512 MB Corsair PC3200LL TwinX DDR RAM in DIMM 1 and 3
  • LG 8x DVD+/-RW.
  • 80 GB Seagate Hard Drive
  • Samsung 950p 19″ Monitors
  • USB Keyboard and Optical Mouse
  • Retail HSF
  • AMK SX1000 modded PC case (window, fans, cables, loom)
  • Enermax 465 Watt FC PSU
  • Windows XP Professional Service Pack 1 updated DX90.b installed.
  • AMD FX-53 Processor
    (32-bit mode)
  • Gigabyte
    8NNXP-940 motherboard
  • ASUS SK8N motherboard
  • ATI 9800 PRO 256 MB Video
    Card Catalyst 4.2 drivers (Application preference ticked for Anti-Aliasing
    and Anisotropic Filtering in both Direct 3D and OpenGL, VSYNC disabled BIOS
    AGP aperture set to 256)
  • 2 x 512 MB Crucial
    PC3200 ECC REG DDR RAM in DIMM 1 and 3
  • Sony 52x CD
  • 80 GB Seagate Hard Drive
  • Samsung 950p 19″ Monitors
  • USB Keyboard and Optical Mouse
  • Retail boxed heatsink
  • AMK SX1000
    modded PC case (window, fans, cables, loom)
  • Enermax 465 Watt FC PSU
  • Windows XP Professional Service Pack 1 updated DX90.b installed.
  • Intel P4 3 GHz processor (HT enabled)
  • ABIT
    IC7-Max3 motherboard
  • ATI 9800 PRO 256 MB Video
    Card Catalyst 4.2 drivers (Application preference ticked for Anti-Aliasing
    and Anisotropic Filtering in both Direct 3D and OpenGL, VSYNC disabled BIOS
    AGP aperture set to 256)
  • 2 x 256 MB Corsair PC3200 DDR RAM in DIMM 1 and 3
  • LG 52x CD/RW
  • 80 GB Seagate SATA Hard Drive
  • Samsung 950p 19″ Monitors
  • USB Keyboard and Optical Mouse
  • Retail HSF packaged with processor
  • AMK SX1000
    modded PC case (window, fans, cables, loom)
  • Enermax 465 Watt FC PSU
  • Windows XP Professional Service Pack 1 updated DX90.b installed.
  • AMD 3200+ 400 FSB
    Processor
  • Gigabyte
    7NNXP motherboard
  • ATI 9800 PRO 256 MB Video
    Card Catalyst 4.2 drivers (Application preference ticked for Anti-Aliasing
    and Anisotropic Filtering in both Direct 3D and OpenGL, VSYNC disabled BIOS
    AGP aperture set to 256)
  • 2 x 256 MB Corsair PC3200 DDR RAM in DIMM 1 and 3
  • Sony 52x CD
  • 60 GB Maxtor ATA133 Hard Drive
  • Samsung 950p 19″ Monitors
  • USB Keyboard and Optical Mouse
  • Globalwin CAK4-76T HSF
  • AMK SX1000
    modded PC case (window, fans, cables, loom)
  • Enermax 465 Watt FC PSU
  • Windows XP Professional Service Pack 1 updated DX90.b installed.

Programs used

All tests were run at default video card settings with VSYNC disabled. Anti-Aliasing
and Anisotropic Filtering was left ticked for application preference. AGP aperture
was set to 256 MB. Windows visual effects
was set for ADJUST FOR BEST PERFORMANCE.

Individual performance will vary with any particular or specific timings or
tweaks enabled by you. A 1024 MB page file was moved to D: partition. Temporary
Internet files moved to J: partition at end of drive. OS installed to C: and
programs installed to E:. All programs were benchmarked with initial monitor
settings at 1024×768@75Hz. Your own mileage may very.

Please note that in the following benchmark graphs the ASUS test motherboard for the Athlon 64 3800+ is incorrectly identified as the SK8V. The correct model is 8KV.

3DMark 2003

3D Mark 2003 was originally designed to measure performance specifically in
shader-heavy titles.

3dmark2003_640

3dmark2003_800

3dmark2003_1024

3dmark2003_1280

3dmark2003_1600

 

Aquamark3

Aquamark3 is a newer benchmark from Massive Development. For the most part
it is a DirectX 8.1 benchmark though it is run with DirectX 90b installed. Four
measurement sets were used. The first has high and low detail with Anti Aliasing
and Anisotropic filtering turned off. The second has high and low detail with
Anti-Aliasing (6x) and Anisotropic filtering (16x) set at max.

aquamark_offofflow

aquamark_offoffhigh

aquamark_616low

aquamark_616high

GL Excess

GL Excess is an OPENGL benchmark that is optimized for DX8.1.

glexcess

 

Quake III high quality

Quake III continues to hang around. This benchmark is one that
most can’t just let go of and it retains grandfather rights in the community.
Many of today’s games are based upon the Quake engine. It wasn’t too long ago that we thought topping 100
FPS was fast. Now we sit at over 300 FPS with the screen set to a high resolution
and detail.

quake

 

Serious Sam

serioussam

 

UT2003 Flyby

ut2003

 

UT2004 Benchmark

ut2004

 

Wolfenstein Enemy Territory: Railgun timedemo

Wolfenstein Enemy Territory uses an improved version of the heavily
modified Quake III engine from Return to Castle Wolfenstein. The Railgun time
demo results were recorded.

wolfenstein

 

X2 Rolling Demo

X2 – The Threat is a teaser with a benchmark option for Egosoft’s
upcoming release. It does not use pixel shaders.

x2

 

Call of Duty Demo

Call of Duty is a new game thus using the latest in optimizations.
FRAPS was used to record the average number of frames per second over a minimum
of 100,000 played frames. Call of Duty also finds its roots in the Quake III
engine.

callofduty

 

SplinterCell (Chinese Embassy Timedemo)

splintercell

 

Specviewperf 7.1

SpecviewPerf measures the 3D rendering performance of systems
running under OPENGL.

specview

 

The following two tests are targeted mainly towards CPU performance and will
show if any “flaws” are in board design affecting the ability of the
CPU to crunch through the data. While in render mode the two test programs virtually
bypass ram and GPU.

Adobe After Effects 5.5

Adobe After Effects is a tool to produce motion
graphics and visual effects for film, video, multimedia and the web. It is primarily
a 2D application using imported graphics or digital footage or self generated
effects. A project was created that was a combination of many video footage
files, resizing and rasterizing effects, text animations and multiple layer
effects. This “average” combination was felt to best demonstrate advantages
and/or disadvantages that a real world user may experience rather than isolating
and benchmarking a particular effect.

There is no official benchmark for After Effects
but tasks can be timed to show specific results. Rendering, or the task of building
and compiling frames, is primarily CPU intensive and After Effects generally bypasses
the video card to rely solely upon the processor for speed. The time taken
to render 900 frames shows how fast the processor is working on the
given task.

The Socket 940 FX-53 does show off its 2D workstation advantages in the After Effects benchmark.

aefx

Softimage XSI can simply bring
any computer to its knees. It’s an incredibly powerful 3D animation program
that has the ability to become so complex that single processor systems have
been known to “think” for days when rendering an animation. A faster and more powerful video
card will translate to a smoother interface where complex scenes are manipulated
in real time. Users can manipulate objects in a
choice of views from wire frame mode to simulated real-time shading mode. A user must render a frame to disk, which
bypasses the GPU, in
order to look at a finished product. A faster processor will result in a faster render.

softimage

Benchmark Conclusions

The Athlon 64 3800+ Socket 939 dances well with the Socket 940 FX-53 and both take a turn at the pole position. Does the larger cache provide a bigger performance benefit or is it the fatter data pipe? It’s difficult to say and it depends on the application. Unfortunately software boxes don’t come with labels that say “Optimized for a large processor cache” or “Warning! Memory bandwidth intensive!”

Conclusions

ws_3800

Socket 940 FX-53 came onto the market and embarrassed the competition. The Socket 939 Athlon 64 3800+ processor takes turns with the FX-53 940-pin at the pole position. Though we did not test the FX-53 for 939-pin, it has been documented that AMD continues to stomp heavily around the benchmark world. Anyone who says Intel wears the crown is sadly mistaken.

Socket 940 owners should not lust for Socket 939 but be cautious if considering Socket 940 for single processor as having an upgrade path. Currently Socket 940 FX-53 owners have one of the top performing processors on the market and all the stability of ECC Registered memory. Moving from lower end Athlon XP processors to Socket 939 is akin to getting out of a race car and into a rocket ship. It isn’t fair to call current processors “slow” anymore. There’s just fast and faster.

But it’s not all about the benchmarks. Cool’n’Quiet takes me back to the WPCRESET tweaks of the KT7A-Raid days. A tweak back then enabled the CPU_HALT command and thus a much cooler processor. Cool’n’Quiet, once firmly established, will be as good an idea as sliced bread. Less heat, lower power consumption and money saved appeals to anyone.

AMD have introduced and are introducing many processor choices in the marketplace from value to performance. The question is why? The majority of consumers purchase a complete system and not just a processor. Consumers also want the best bang for their buck and it’s important to put choices within reach. It’s strategy to place many processor choices in the market that are close together in price and specifications. The next level is “just a few dollar more” or “why not choose this processor? It’s just about the same AND its lower price means you can afford more RAM.”

Sound familiar?

The close price point between Socket 939 Athlon 64 3800+ and Socket 754 Athlon 64 3700+ is most likely the proverbial carrot for consumers to transition to Socket 939 allowing AMD to nudge Socket 754 towards the value spectrum.

AMD is doing processor housekeeping to set itself up for the next few years. Socket 939 has the most potential for a long life span in the mid-level to performance area. It’s the odds on favorite which will make vendors and manufacturers happier because they won’t have to stock as large a variety of parts. It may also provide a bit of upgrade insurance to consumers. Socket 754 will most likely be used to support value based processors and also hang around a bit longer due to its popularity. Socket 940 for the enthusiast could be the first to fall by the wayside. This only puts more emphasis on Opteron. Processors and Registered RAM have their place but will the interest be enough to warrant further development for for single processor Socket 940? Will history show it as the sacrificial olive branch between enthusiast/workstation user and Opteron? Don’t forget that waiting to hear its quiet epitaph is Socket A which has carried AMD through Spitfire, Morgan, Thunderbird, Thoroughbred, Barton, Thorton, Appaloosa, Applebred, and Palomino. I’m sure it will get a proud place of honor in the AMD retirement showcase soon enough.

 

Our thanks to AMD for
their support of this and many other sites.

Highs

  • Impressive performance
  • Cool’n’Quiet technology
  • Upgrade path defined

Lows

  • Top end price

Comments

  1. Thrax
    Thrax
    Spitfire, Morgan, Thunderbird, Thoroughbred, Barton, Applebred, and Palomino

    You forgot Thorton and Appaloosa. ;D
  2. MediaMan
    MediaMan Corrected. Thanks Thrax! :)

    Personally I was quite impressed by Cool'n'Quiet.
  3. Thrax
    Thrax Cool'n'quiet impressed me too, come to think of it.

    I thought the dynamic scaling would impact performance as the CPU adjusted its speed settings slightly behind what the program expected
  4. MediaMan
    MediaMan The dynamic scaling curve is rather steep. Mind you it's hard/almost impossible to get a processor to ramp up in usage in a gradual curve.

    Cool'n'Quiet is a feature that also must be supported by the motherboard. The ASUS 8KV sample I have has it in BIOS as enable/disable. You also have to download software. The software isn't so much a GUI but a "patch" of sorts to the OS.

    The Gigabyte 939 pin board I just got doesn't have the Cool'n'Quiet feature and has remained with CPU thermal throttling. It will prove to be a mistake on their part.

    There are some users all over the web who have posted problems with Cool'n'Quiet but it's because the feature is in its early adoption phase. Motherboard manufacturers haven't got it quite right yet. I don't think it's a question of software conflicts but more likely BIOS buggering up by motherboard makers.

    And as far as an impact on performance? Zilch. Response time was pretty instantaneous between the lesser voltage/clock speed and then needing to be "full on". I noticed absolutely no hesitation or even measured any hesitation. The CPU usage graphs show that.
  5. GHoosdum
    GHoosdum I've got a few questions. First, although ECC Registered RAM is not required by Socket 939, can you still use it and gain its benefits in a Socket 939 board?

    Second, all of your graphs show that the processors with Cool'n'Quiet enabled run cooler even at full load than with it disabled. Does this seem to be because the processor hasn't had enough time at full voltage to catch up in heat? I'm asking because when I upgrade to a Socket 939 system, it will be folding for me 24/7 and I'm wondering whether it's useful for me to turn Cool'n'Quiet on at that time - if load temps will really be lower even with full load full time conditions, then that really IS a feat.

    The most common Cool'n'Quiet glitch I've seen so far is the fact that my A64 notebook can't run at above 800 MHz on battery NO MATTER WHAT LOAD. But I'm not sure if the socket 754 notebooks use Cool'n'Quiet, or still use PowerNow...
  6. citrixmeta
  7. MediaMan
    MediaMan
    GHoosdum wrote:
    I've got a few questions. First, although ECC Registered RAM is not required by Socket 939, can you still use it and gain its benefits in a Socket 939 board?

    Short answer: No. In order to use ECC Registered memory the motherboard must support it.
    GHoosdum wrote:
    Second, all of your graphs show that the processors with Cool'n'Quiet enabled run cooler even at full load than with it disabled. Does this seem to be because the processor hasn't had enough time at full voltage to catch up in heat? I'm asking because when I upgrade to a Socket 939 system, it will be folding for me 24/7 and I'm wondering whether it's useful for me to turn Cool'n'Quiet on at that time - if load temps will really be lower even with full load full time conditions, then that really IS a feat.

    It's best to think over the the duration. Remember that Cool'n'Quiet has a pretty instantaneous reaction time in gearing up/down. Now...in that Sandra CPU burn-in example there is a pause between repititions as the burn-in cycle shuts down....switches to the next task...and repeats. The processor will "gear down" during those "pauses" allowing it to run at a cooler temperature. The ambient room was consistent during those tests. I would supsect that, on average, overall CPU temps will be 1-3 degrees cooler but that depends on the type of test and if there are any "pauses"...allow for some variables before definitely assuming that Cool'n'Quiet will consistently show lower CPU temps on the whole.
    GHoosdum wrote:
    The most common Cool'n'Quiet glitch I've seen so far is the fact that my A64 notebook can't run at above 800 MHz on battery NO MATTER WHAT LOAD. But I'm not sure if the socket 754 notebooks use Cool'n'Quiet, or still use PowerNow...

    "Cool'n'Quiet" is a function in BIOS. It will be an enable/disable setting labelled as Cool'n'Quiet. You would have to do the following to use Cool'n'Quiet.

    1) Enable it in the BIOS
    2) Download the software for Cool'n'Quiet
    3) Set it in the OS properties in the same manner as you would for Powernow.

    Cool'n'Quiet has to have a motherboard that supports it. I cannot confirm if this is as simple as a BIOS upgrade to enable this feature. Remember that notebooks and desktops are treated differently but it would seem that Cool'n'Quiet would be more efficient for notebook battery life since the processor would, on average, run at a lower voltage setting for most of the time thus extending battery life. When you would need full speed then it would kick up to that...such as during gaming.

    But then you are fully aware that you are sucking your battery dry. :)

    1000 MHz, if I'm right, is a new step in A64 socket 939 motherboards. If I remember correctly 800 MHz was the previous HT link maximum and this would be especially true of Socket 754.

    Now where's Thrax to confirm this for me? :)
  8. MediaMan
    MediaMan
    citrixmeta wrote:


    CRAPWEASEL. Damn graphs...input one letter and get distracted and the whole charting system goes fer poop. Yes it is an ASUS K8V. Think of it this way...

    "What mobo is it?"

    "S'K8V."

    ;D
  9. citrixmeta
    citrixmeta hehehe , i guess it ok.

    its still an awesome article
  10. MediaMan
    MediaMan GHoosdum,

    Okay...did some poking about and I've got real answers for you.

    - C'n'Q requires logic (HW) on the CPU chip itself to support, and yes... this has been implemented in all Athlon 64 desktop processors
    - C'n'Q requires corresponding motherboard logic (HW) to support the functionality... that's up to mobo vendors, but I think you'll find most supporting it
    - PowerNow! is the technology for the mobile segment and yes, it requires CPU and mobo logic (HW) support as well.

    I don't think it's possible find any K8 laptops that do not support PowerNow! Check your power settings... it's probably locked in "Max Battery" or "Presentation" mode. Select "minimal power management" in the power settings tab which puts C'n'Q in automatic mode... just like PowerNow!
    Keep in mind that C'n'Q and PowerNow! do not affect the speed of the HyperTransport link operation, just the CPU multiplier used for frequency determination for the processor.
  11. Thrax
    Thrax The new 1GHz hypertransport link is natural evolution for VIA who debuted with an 800MHz 16/16bit down/up HT link. However for nVidia, it's a kneejerk reaction as they debuted with a cripled 600Mhz 16/8bit down/up HT link.

    It's the 200MHz LDT bus with a 5x multiplier. Rumour has it that AMD is about to bless 1GHz HT links.
  12. Straight_Man
    Straight_Man Registered NON-ECC is available in some speeds. THAT can be used, and you can take advantage of the mfring QC behind the Registered RAM. ECC features will not work though, the non-ECC supporting BIOS should ignore the extra module or modules-- in some cases, on non-ECC supporting baords, the sockets themselves do not have spring contacts for the ECC pinouts either. Whether Windows or another O\S will ignore ECC, in this case of using ECC on a board that does not support it, is another story entirely. And therein does lie one BIG problem with using ECC on a non-ECC supporting board: If Windows tries to use ECC features, and board cannot provide underlying logic for that, then you can get a mess of a RAM-use mishmosh as board tries to use the extra ECC module for RAM or igonres it totally and Windows tries to use it for ECC checks. You will get mapping errors that lead to RAM violates and BSODs in O\S out of this, and you will get RAM size mismatches between BIOS and Windows.

    But it is not a fiscally sound thing to use either, right. CAN use registered non-ECC but not worth spending the bucks in most cases unless you soon plan to change abords to one with Registered RAM recognition in BIOS, and ECC or Registered plus ECC is not a good thing to use on a board that does not have BIOS support for it.

    Real old boards and early PC gens ECC was an upgerade option for mots of boxes. BIOSs could use either. A BIOS that can ECC recognize and an OS that can map accordingly are major requirements for ECC, plus tight power cycle timings and voltages on whole board.

    I would slightly mod what MM said as to registered-only RAM, but as to ECC do not use unless your board supports it, whether registered plus ECC or non-registered ECC which is a questionable thing to offer or use these days anyway compared to higher end Registered plus ECC. Registered feature is mostly a quality assurance thing, and used mostly with servers that use ECC for data integrity assurance on hardware. And ECC plus registered RAM needs tighter sync ratios to work right and sync to CPU than other more generic but high quality RAM does-- for pure Extreme OCing, would not use registered plus ECC, period. You will get a system that is very picky about what combos of timings it will accept if you use registered plus ECC-- but much better assurance of integrity of data in RAM at correct timings for RAM and CPU.

    We have ignored BUFFERED RAM here, that is something else having to do with using physical (usually smaller) modules as temp storage right on the stick of RAM. Some ECC can be physically also buffered RAM (and I have seen buffered non-ECC at certain time sin history, in the age of SIMMs as best available, and BIOSs that could detect buffers on sticks). That can give you a RAM to bus latency decrease potential if the stick is QC'd right at mfr. I would not bother buying buffered OEM grade RAM. Expect to pay for QC checking of guaranteed buffered RAM.
  13. Thrax
    Thrax You cannot use ECC or registered memory on socket 939 boards. It will not work.

    Buffered is the same as registered, and it also will not work.
  14. MediaMan
    MediaMan Oh gosh...dare I open this can of worms.

    RAM can have buffers.
    RAM can be unbuffered
    RAM can have registers
    Registers are not buffers.

    Therefore:

    You can have unbuffered ECC Registered memory.
    You can have buffered Registered Memory.
    You can have unbuffered ECC memory.
    You can have unbuffered memory.
    You can have buffered ECC Registered memory.



    I believe non-parity = "non-registered"
    I believe parity = "registered"

    ECC Reg and unbuffered - non-parity are the two most common for our particular world.
  15. Tex
    Tex Not quite. You are slightly confused. And its easy to do. Let me put it this way..... registered memory is = buffered. Its another name for the same thing.

    But...... You can have ecc memory thats either buffered or non buffered. The buffered is "registered ecc". The non bufferred is regular ecc and is not registered.

    All non ecc memory is also non registered and thus non bufferred.

    You actually have three types of sdram/ddr then

    1) non bufferred/non registered.

    2) ecc thats non buffered which is the same thing as ecc thats non registerred. ecc comes in both flavors

    3) ecc that is buffered.... and this is another name for "registerred ecc" if its registerred its also ecc in all cases. You can't have registered memory thats also not ecc.

    parity is = ecc.
    non parity is = non ecc.

    parity is the extra chip that does the error correction and has nothing to do with registerred or non registered.

    Just trying to clear up the confusion. Did that help at all?

    Tex
  16. MediaMan
    MediaMan Tex to the rescue! :)
  17. GHoosdum
    GHoosdum
    MediaMan wrote:
    GHoosdum,

    Okay...did some poking about and I've got real answers for you.

    - C'n'Q requires logic (HW) on the CPU chip itself to support, and yes... this has been implemented in all Athlon 64 desktop processors
    - C'n'Q requires corresponding motherboard logic (HW) to support the functionality... that's up to mobo vendors, but I think you'll find most supporting it
    - PowerNow! is the technology for the mobile segment and yes, it requires CPU and mobo logic (HW) support as well.

    I don't think it's possible find any K8 laptops that do not support PowerNow! Check your power settings... it's probably locked in "Max Battery" or "Presentation" mode. Select "minimal power management" in the power settings tab which puts C'n'Q in automatic mode... just like PowerNow!
    Keep in mind that C'n'Q and PowerNow! do not affect the speed of the HyperTransport link operation, just the CPU multiplier used for frequency determination for the processor.

    Thanks for digging up the info, MM. I'm sorry that I forgot to mention at first, it's not just me having this problem. So far, every A64 notebook I've seen has had the 800MHz on battery issue. Supposedly, manufacturers need to release a BIOS update. I think. There were also ramblings that it's a Microsoft issue, since the speed will step up to full on an A64 nb in Win2K, but not XP.
  18. Unregistered
    Unregistered This article IS on the "top end of informative" piece literature I've come across in some time. Good job and thanks for providing this type of information.

    Guest
  19. Unregistered
    Unregistered Thanks for the excellent article, MM. It really clears up a lot of misconceptions I had about ECC RAM and where the AMD-64 line is headed.

    I'm a gamer who just bought a used system based on the Athlon FX-51 in an Asus SK8V mobo... the system is pretty much loaded (1GB DIMM of Kingston ECC, 160GB SATA + 40GB RAID 0 HDDs, 256MB Radeon 9800), I got it for $1150, it seemed at the time like a great deal but after reading your article, part of me wonders if maybe the guy was dumping it because he knows the upgrade options will be limited for Socket 940.

    Can you or anyone else hazard a guess as to how many more Opterons will come out for this socket, since Athlon FX is moving over to Socket 939? If some Socket 940 Opterons ever get up around 3GHz a year or two from now then I'd have a pretty clear upgrade path.

    -Adam in Philly
  20. MediaMan
    MediaMan Adam,

    Let's not start enscribing the epitaph for Socket 940 as of yet. It's true that Socket 939 will be the top end enthusiast line simple because of the less expensive motherboard manufacturing process (we're talking dollars here...not 10's or 100's of dollars) and, of course, the less expensive unbuffered RAM.

    Let's look at what you got. A heckuva deal on one really fast system. Think of it this way...in order to get a noticeable boost in performance there must be a significant jump in processor speed. Moving from FX-51 to FX-53 is indeed faster but measureable only by benchmarks and not really in the "feel" of the two given identical setups. By the time a processor level comes out that provides a worthwhile processor upgrade. (FX-61? FX-71?) the whole pricing and scope of the platform is going to change.

    I am presuming that Socket 939 will be to AMD what Socket 462 was. Remember how at the beginning of one year we were just approaching the 1 GHz barrier and then zoom...1.4 and 1.6 GHz. Just about that time the PR rating came into effect. There were some significant jumps in processor speeds.

    Now we sit at 3800+ (2.4 GHz in Intelanguagese) at the start of a new Socket. Could we all be in for another year of jumps? I could see 3 GHz perhaps but it's anyone's guess. It is for certain that new processors will come out...they may be faster in GHz but I think they will "mushroom" in other areas such what happened with larger on die cache, HT links and so on.

    You've got a pretty lickety split system at a good price. Be happy. By the time you "need" a new system ...things will have changed again. PCI Express and so on.

    Those that always keep up with the bleeding edge of technology probably are millionaires. I'm selling off body parts to keep me in the game. :) I just have to remember to keep one arm to type with and at least one eye.
  21. Thrax
    Thrax Opteron will stay at 940, says AMD.
  22. MediaMan
    MediaMan You beat me to it.

    Socket 940 will continue on in the form of Opteron 1xx processors. Loosely defined think of FX-xx as the "limited" edition or sports model since FX 940 is based on Opteron 1xx 940.

    Remember that Opteron 1xx processors will always be coming out in new frequency flavors so the upgrade path will be there. You just may not see as many FX-xx launches as those are, at best words, special events.

    Socket 940 isn't dead...it's had it's place in the spotlight and now 939 will be taking it's place in the spotlight for a while.
  23. Unregistered
    Unregistered "Spitfire, Morgan, Thunderbird, Thoroughbred, Barton, Applebred, Palomino, Thorton, and Appaloosa" would have been better if said in as passage of time begining with the earilest, (I think Spitfire), to the latest, Barton. Otherwise I loved the article.
  24. Thrax
    Thrax The earliest was the Thunderbird, and the latest is the Thorton.

    It's actually:

    Thunderbird (Athlon), Spitfire (Thunderbird-derived Duron), Palomino (Athlon XP), Morgan (Palomino-derived Duron), Thoroughbred (.13u Athlon XP), Barton (.13u Athlon XP w/ 512k cache), Applebred (Thoroughbred-derived Durons), Appaloosa (Barton-derived Duron never released), Thorton (Barton with 256k L2).
  25. Unregistered
    Unregistered yo somebody publish this auther - MM. put him in a top-selling comp magazine or something, this article was SO HOT and useful... I wish he would have mentioned all 3 sockets for the 64 more often than just the 939 and the 940 though. The whole adding the Intel CPU to the benchmarks was over-the-edge in putting AMD as king overall, so when they say 3800+ - I guess it really does mean that it performs "like" a 3.8Ghz.
  26. Unregistered
    Unregistered Good article , no bull , just the facts Ma'am !

    You have cleared up my doubts regarding upgrade path between 939-940 sockets. I also agree with getting you published. Your tone and content are what PC people seek... like a good conversation.

    cheers and keep it up.
  27. Shorty
    Shorty That really is still a really good read :)
  28. Unregistered
    Unregistered A very well written and informative article aimed at the right level for a 'non techy'. It would have been nice to see the benchmanrks on the AMD3500+, which is the one i was comparing with the 3800+ and the FX53. But hey, who wants to be picky!! It was just what i needed to know. Thanks from one of the many 'guests' that have read and used the information in this article.
    More Enlightened
  29. Unregistered
    Unregistered This article was so helpful because it explained the real world relationships between these processors & I'm about to spend real world dollars. The local computer shop guys are streets in front of me but I know now that they only think they are tech savy...thank you.

Howdy, Stranger!

You found the friendliest gaming & tech geeks around. Say hello!