Overclock.net › Forums › AMD › AMD CPUs › Hypothetical Question - FX-4300 vs. Phenom II X4 Clock for Clock
New Posts  All Forums:Forum Nav:

Hypothetical Question - FX-4300 vs. Phenom II X4 Clock for Clock - Page 5

post #41 of 131
Quote:
Originally Posted by drmrlordx View Post

One thing I want to know is: why did AMD always keep the Northbridge speeds so low on k10.5, particularly C3 and E0 stepping chips? Deneb/Zosma/Thuban left a boatload of performance on the table that could be unlocked by bumping NB clocks up to 2.4-3 ghz, and the amount of extra power draw or (obvious) heat/risk-to-processor in tapping into that performance seemed pretty low. I know I've probably killed my own K10 chip by stressing the NB too much (that's my guess, anyway), but the gains were pretty big, especially on some synthetic benchmarks where AMD really needs to gain ground (if public perception/press are really that important).

Enthusiast overclockers probably would not have liked such a development, especially since high NB clocks can sometimes limit core clocks somewhat, but bringing out a second-generation Thuban with maybe 100-200 mhz advantage over Thuban on 32nm with a stock 3 ghz NB would have been huge for the benchmarks. NB doesn't seem to make as much of a difference for Bulldozer/Piledriver (does it?), so I can understand why they haven't messed with that much.

The memory controller ratios are different for Phenom II and FX, that's the main difference. Think of the nb as a cog in a gear. For Phenom II it had a designed ratio of 1:3 for RAM:NB frequency. Since Phenom II was spec'd for DDR3-1333 (667MHz), the stock nb was 2000MHz. But since everyone was running DDR3-1600 (800MHz) on Phenom II anyways the nb needed a bump to 2400MHz to take advantage of the RAM speed. The FX has a designed 1:2 ratio for RAM:NB frequency, so high NB frequency isn't needed as much because the memory controller is stronger. In both platforms, increasing NB frequency lowers latency, and also controls the speed of L3 cache. A big problem with the piledriver/bulldozer chips is cache latency, and I'm sure mdocod can explain that wink.gif
Light Gamer #1
(12 items)
 
Light Gamer #2
(10 items)
 
 
CPUMotherboardGraphicsGraphics
i3-6100 ASRock Fatal1ty Z170 Gaming K4/D3 MSI r7850 TFIII OC MSI r7850 TFIII OC 
RAMHard DriveCoolingOS
G,Skill TridentX OCZ Vector 150 TPC-612 Windows 10 
KeyboardPowerCaseMouse
G105 Rosewill Capstone 550M HAF XB M570 
CPUMotherboardGraphicsRAM
Athlon 860K MSI A88X-G45 MSI GTX 460 G.Skill RipjawsZ 
Hard DriveCoolingOSKeyboard
PNY Optima Hyper T4 Windows 10 Devastator 2 
PowerCase
Rosewill Valens DIYPC Ranger-R4-R 
  hide details  
Reply
Light Gamer #1
(12 items)
 
Light Gamer #2
(10 items)
 
 
CPUMotherboardGraphicsGraphics
i3-6100 ASRock Fatal1ty Z170 Gaming K4/D3 MSI r7850 TFIII OC MSI r7850 TFIII OC 
RAMHard DriveCoolingOS
G,Skill TridentX OCZ Vector 150 TPC-612 Windows 10 
KeyboardPowerCaseMouse
G105 Rosewill Capstone 550M HAF XB M570 
CPUMotherboardGraphicsRAM
Athlon 860K MSI A88X-G45 MSI GTX 460 G.Skill RipjawsZ 
Hard DriveCoolingOSKeyboard
PNY Optima Hyper T4 Windows 10 Devastator 2 
PowerCase
Rosewill Valens DIYPC Ranger-R4-R 
  hide details  
Reply
post #42 of 131
Quote:
Originally Posted by Melcar View Post

Overclocking the NB adds a lot of heat and most chips need a rather big bump in voltage. The crappy retail heatsink would have been insufficient. The benefits are mostly on the memory side of things anyway, something that does not always translate directly to real life performance in most situations. Most review sites did gross exaggerations on the benefits of NB overclocking back in the Deneb/Thuban days.

Does it really, though? My personal experience is limited to C2-stepping chips without any L3, but when I OCed the NB and gave it a healthy does of extra voltage (as well as extra voltage on the CPU->NB interconnect), it barely changed thermals. Maybe if I hadn't been using an nh-d14 with CLU, I would have noticed this more?

The behavior of review sites merely serves to confirm the need for AMD to address the NB speed "problem" by jacking up speeds to make benchmarks run faster. I did a whole slew of NB speed testing with a Propus years ago to show what NB speeds would do; in fact, I just found the thread right here.

Quote:
Originally Posted by damric View Post

The memory controller ratios are different for Phenom II and FX, that's the main difference. Think of the nb as a cog in a gear. For Phenom II it had a designed ratio of 1:3 for RAM:NB frequency. Since Phenom II was spec'd for DDR3-1333 (667MHz), the stock nb was 2000MHz. But since everyone was running DDR3-1600 (800MHz) on Phenom II anyways the nb needed a bump to 2400MHz to take advantage of the RAM speed. The FX has a designed 1:2 ratio for RAM:NB frequency, so high NB frequency isn't needed as much because the memory controller is stronger. In both platforms, increasing NB frequency lowers latency, and also controls the speed of L3 cache. A big problem with the piledriver/bulldozer chips is cache latency, and I'm sure mdocod can explain that wink.gif

Actually, from my experiences, the actual memory speed is largely irrelevant; the latency advantages of NB overclocking are responsible for most gains. There are plenty of K10/10.5 chips that have no real need for extra memory bandwidth. What they need is lower latency, and you can always improve the latency between the CPU and NB/IMC by increasing the NB clock speed (well, to a point, but reaching that point is darn near impossible). I recorded gains at NB speeds all the way up to 2850 mhz on Propus, and I've seen Super Pi 32m gains on a lowly Sargas (Sempron 140) from similar NB speeds. There's no way a single-core K10 ever needed extra memory bandwidth, regardless of the clock speed.
post #43 of 131
Quote:
Originally Posted by drmrlordx View Post

Does it really, though? My personal experience is limited to C2-stepping chips without any L3, but when I OCed the NB and gave it a healthy does of extra voltage (as well as extra voltage on the CPU->NB interconnect), it barely changed thermals. Maybe if I hadn't been using an nh-d14 with CLU, I would have noticed this more?

The behavior of review sites merely serves to confirm the need for AMD to address the NB speed "problem" by jacking up speeds to make benchmarks run faster. I did a whole slew of NB speed testing with a Propus years ago to show what NB speeds would do; in fact, I just found the thread right here.
Actually, from my experiences, the actual memory speed is largely irrelevant; the latency advantages of NB overclocking are responsible for most gains. There are plenty of K10/10.5 chips that have no real need for extra memory bandwidth. What they need is lower latency, and you can always improve the latency between the CPU and NB/IMC by increasing the NB clock speed (well, to a point, but reaching that point is darn near impossible). I recorded gains at NB speeds all the way up to 2850 mhz on Propus, and I've seen Super Pi 32m gains on a lowly Sargas (Sempron 140) from similar NB speeds. There's no way a single-core K10 ever needed extra memory bandwidth, regardless of the clock speed.

So are you saying that you benched your K10.5 will high RAM speed like DDR3-2000? Didn't think so biggrin.gif They generally couldn't stabilize much over DDR3-1600, though there were some with lucky to get 1800 or even 2000. That's why for the K10.5 the general goal was 1600CL6 with as high a CPU-NB as possible. I always wanted a set of the DDR3-2000CL7 Flares to test on my Thuban but they were too hard to find. A lot of people ran ECOs though since they had the same Powerchip ICs which were just golden on that platform.
Light Gamer #1
(12 items)
 
Light Gamer #2
(10 items)
 
 
CPUMotherboardGraphicsGraphics
i3-6100 ASRock Fatal1ty Z170 Gaming K4/D3 MSI r7850 TFIII OC MSI r7850 TFIII OC 
RAMHard DriveCoolingOS
G,Skill TridentX OCZ Vector 150 TPC-612 Windows 10 
KeyboardPowerCaseMouse
G105 Rosewill Capstone 550M HAF XB M570 
CPUMotherboardGraphicsRAM
Athlon 860K MSI A88X-G45 MSI GTX 460 G.Skill RipjawsZ 
Hard DriveCoolingOSKeyboard
PNY Optima Hyper T4 Windows 10 Devastator 2 
PowerCase
Rosewill Valens DIYPC Ranger-R4-R 
  hide details  
Reply
Light Gamer #1
(12 items)
 
Light Gamer #2
(10 items)
 
 
CPUMotherboardGraphicsGraphics
i3-6100 ASRock Fatal1ty Z170 Gaming K4/D3 MSI r7850 TFIII OC MSI r7850 TFIII OC 
RAMHard DriveCoolingOS
G,Skill TridentX OCZ Vector 150 TPC-612 Windows 10 
KeyboardPowerCaseMouse
G105 Rosewill Capstone 550M HAF XB M570 
CPUMotherboardGraphicsRAM
Athlon 860K MSI A88X-G45 MSI GTX 460 G.Skill RipjawsZ 
Hard DriveCoolingOSKeyboard
PNY Optima Hyper T4 Windows 10 Devastator 2 
PowerCase
Rosewill Valens DIYPC Ranger-R4-R 
  hide details  
Reply
post #44 of 131
Quote:
Originally Posted by damric View Post

So are you saying that you benched your K10.5 will high RAM speed like DDR3-2000? Didn't think so biggrin.gif They generally couldn't stabilize much over DDR3-1600, though there were some with lucky to get 1800 or even 2000. That's why for the K10.5 the general goal was 1600CL6 with as high a CPU-NB as possible. I always wanted a set of the DDR3-2000CL7 Flares to test on my Thuban but they were too hard to find. A lot of people ran ECOs though since they had the same Powerchip ICs which were just golden on that platform.

Actually, I'm saying I did the opposite: I ran high NB speeds with mediocre memory speeds. The linked thread above featured DDR3-1520 6-7-5-15 1T at NB speeds from 2000-2850 mhz. Point being, I recorded improvements in performance from higher NB speeds when it was obvious that additional bandwidth between the CPU and NB wasn't necessary (it would be in excess of the bandwidth between the NB/IMC and RAM). It was all about latency.
post #45 of 131
Quote:
Originally Posted by drmrlordx View Post

Actually, I'm saying I did the opposite: I ran high NB speeds with mediocre memory speeds. The linked thread above featured DDR3-1520 6-7-5-15 1T at NB speeds from 2000-2850 mhz. Point being, I recorded improvements in performance from higher NB speeds when it was obvious that additional bandwidth between the CPU and NB wasn't necessary (it would be in excess of the bandwidth between the NB/IMC and RAM). It was all about latency.

So you say you proved that high RAM speeds do not help K10.5 without actually testing such? Remarkable. Perhaps you should re-write the scientific method to include non-observations as true when your eyes are closed rolleyes.gif
Light Gamer #1
(12 items)
 
Light Gamer #2
(10 items)
 
 
CPUMotherboardGraphicsGraphics
i3-6100 ASRock Fatal1ty Z170 Gaming K4/D3 MSI r7850 TFIII OC MSI r7850 TFIII OC 
RAMHard DriveCoolingOS
G,Skill TridentX OCZ Vector 150 TPC-612 Windows 10 
KeyboardPowerCaseMouse
G105 Rosewill Capstone 550M HAF XB M570 
CPUMotherboardGraphicsRAM
Athlon 860K MSI A88X-G45 MSI GTX 460 G.Skill RipjawsZ 
Hard DriveCoolingOSKeyboard
PNY Optima Hyper T4 Windows 10 Devastator 2 
PowerCase
Rosewill Valens DIYPC Ranger-R4-R 
  hide details  
Reply
Light Gamer #1
(12 items)
 
Light Gamer #2
(10 items)
 
 
CPUMotherboardGraphicsGraphics
i3-6100 ASRock Fatal1ty Z170 Gaming K4/D3 MSI r7850 TFIII OC MSI r7850 TFIII OC 
RAMHard DriveCoolingOS
G,Skill TridentX OCZ Vector 150 TPC-612 Windows 10 
KeyboardPowerCaseMouse
G105 Rosewill Capstone 550M HAF XB M570 
CPUMotherboardGraphicsRAM
Athlon 860K MSI A88X-G45 MSI GTX 460 G.Skill RipjawsZ 
Hard DriveCoolingOSKeyboard
PNY Optima Hyper T4 Windows 10 Devastator 2 
PowerCase
Rosewill Valens DIYPC Ranger-R4-R 
  hide details  
Reply
post #46 of 131
Quote:
Perhaps you should re-write the scientific method to include non-observations as true when your eyes are closed

No need to, we already have quantum mechanics and Schrodinger's cat, heh. tongue.gif
    
CPUMotherboardGraphicsRAM
FX-6300 @ 4.9ghz M5A99X EVO R2.0 PNY 2gb NVIDIA GeForce GTX 650 PNY DDR3-1600 9-9-9-24 2X4gb 
Hard DriveHard DriveHard DriveHard Drive
Intel 330 SSD Samsung HD103SI Samsung HD204UI X 2 esata externals  Western Digital WD5000AAKX  
Optical DriveCoolingOSMonitor
Optiarc DVD RW AD-7280S Cooler Master Hyper 212 plus Windows 8.1.1 pro with media center Funai 32inch LED HDTV 
KeyboardPowerCaseMouse
Microsoft Comfort Curve 2000 Corsair CX600 Custom twin full tower, 12 - 5 1/4 bays total Logitech M510 
Mouse PadAudioAudio
Combo Mouse Pad, Calculator, USB ports USB stereo speakers Custom 5.1 digital sound system 
  hide details  
Reply
    
CPUMotherboardGraphicsRAM
FX-6300 @ 4.9ghz M5A99X EVO R2.0 PNY 2gb NVIDIA GeForce GTX 650 PNY DDR3-1600 9-9-9-24 2X4gb 
Hard DriveHard DriveHard DriveHard Drive
Intel 330 SSD Samsung HD103SI Samsung HD204UI X 2 esata externals  Western Digital WD5000AAKX  
Optical DriveCoolingOSMonitor
Optiarc DVD RW AD-7280S Cooler Master Hyper 212 plus Windows 8.1.1 pro with media center Funai 32inch LED HDTV 
KeyboardPowerCaseMouse
Microsoft Comfort Curve 2000 Corsair CX600 Custom twin full tower, 12 - 5 1/4 bays total Logitech M510 
Mouse PadAudioAudio
Combo Mouse Pad, Calculator, USB ports USB stereo speakers Custom 5.1 digital sound system 
  hide details  
Reply
post #47 of 131
Quote:
Originally Posted by damric View Post

So you say you proved that high RAM speeds do not help K10.5 without actually testing such? Remarkable. Perhaps you should re-write the scientific method to include non-observations as true when your eyes are closed rolleyes.gif

Look, if I had an AM2+/AM3 rig with DDR2 and an AM3 machine with DDR3 and I could tune the systems to produce similar latency, a substantial difference in memory bandwidth, and identical K10/10.5 chips, I'd do it and show you the results. The best I've got at short notice is this thread which is, sadly, less-than-scientific because there's a 500 mhz NB speed advantage for the DDR3 machine. But, it does show the DDR2 rig coming very close in performance. I've seen similar performance deltas from just NB speed changes alone. I'll bet if that DDR2 machine had a 3 ghz NB speed, he's be right there with the DDR3 guy.

Alternatively, we could look at the link width and speed between CPU->NB and NB->DDR3, calculate how much bandwidth is available in either case, and try to figure out a bandwidth saturation point for the NB if you really want.

If you want to think that it was memory bandwidth from NB clockspeed improvements providing extra performance, go right ahead. I'm not going to stop you. But the idea that there's some cut-off of 2400 mhz or what have you beyond which NB speeds will no longer help is just not accurate. K10.5 benefits from any NB speed boost you can give it (at least in synthetics, and sometimes even in real-world apps), for whatever reason. I'm sticking to my guns and saying that it's latency, but the "why" isn't really the most important thing here. AMD could have put a prettier face on K10.5 with faster NB speeds. Not real sure about Bulldozer/Piledriver/Steamroller.
post #48 of 131
Aww man, I was so exited to bench my FX-6350 @ 4.5 GHz with Black Hole Benchmark 4.2 but it keeps throwing up "unhandeled exception" errors at the start and won't let me play. So unlucky... sad-smiley-002.gif
post #49 of 131
FX-43**>Phenom II x4 > FX-63**>Phenom II x6 > FX-FX-83** for bench test with single and mutli threading.
Phenom II x4>FX-43**>Phenom II x6>FX-63**>FX-83** for most game bench test.
It is weird but the truth.
Edited by dixson01974 - 4/12/14 at 8:46pm
HTPC
(11 items)
 
  
CPUMotherboardRAMHard Drive
AMD A-3870K AsRock A75 Pro (FM1) HyperX DDR3 1866 8 GB SAMSUNG 830 256gb 
Hard DriveOptical DriveCoolingOS
SAMSUNG EcoGreen F4 2TB Plextor Black Blu-ray Burner Phenom II X4 stock Cooler Windows 10 
MonitorPowerCase
LG M2762D-PM Glossy Black 27" 5ms (1920x1080) Antec HG 900 Gerenic ATX Case 
  hide details  
Reply
HTPC
(11 items)
 
  
CPUMotherboardRAMHard Drive
AMD A-3870K AsRock A75 Pro (FM1) HyperX DDR3 1866 8 GB SAMSUNG 830 256gb 
Hard DriveOptical DriveCoolingOS
SAMSUNG EcoGreen F4 2TB Plextor Black Blu-ray Burner Phenom II X4 stock Cooler Windows 10 
MonitorPowerCase
LG M2762D-PM Glossy Black 27" 5ms (1920x1080) Antec HG 900 Gerenic ATX Case 
  hide details  
Reply
post #50 of 131
Quote:
Originally Posted by dixson01974 View Post

FX-43**>Phenom II x4 > FX-63**>Phenom II x6 > FX-FX-83** for bench test with single and mutli threading.
Phenom II x4>FX-43**>Phenom II x6>FX-63**>FX-83** for most game bench test.
It is weird but the truth.

+1 on that
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: AMD CPUs
Overclock.net › Forums › AMD › AMD CPUs › Hypothetical Question - FX-4300 vs. Phenom II X4 Clock for Clock