Overclock.net › Forums › Industry News › Hardware News › [ExtremeTech] Analyzing Bulldozer: Why AMD’s chip is so disappointing
New Posts  All Forums:Forum Nav:

[ExtremeTech] Analyzing Bulldozer: Why AMD’s chip is so disappointing - Page 5

post #41 of 292
Quote:
Originally Posted by Gen View Post
Something straight up broke with BD, hard to say if it is software or the chip itself though. I think I'm leaning towards some wrong with chip...
Read the source, it's the cache...Just like I've been saying since a day after it launched.

Quote:
Originally Posted by kweechy View Post
Yeah something's off. 2600k should be around 6.8 - 7.0 like you said.
Hate to obviously not read the source or even most of the topic, eh?
It's been covered, there's no turbo mode.

Quote:
Originally Posted by Track View Post
Can anyone understand why Bulldozer is WEAKER per-core than Phenom II X6?

That's like, the biggest tech fail in history..

Imagine GTX 480 being 10% weaker than GTX 280.
Not really, there's been worse, case in point, the P4. That thing, despite having a larger clock advantage, was monumentally slow while using around double the power at launch.

And its faster than Thuban, just not when less than 8 cores are enabled and it's not using SSE.

Which, coincidentally, I see BD as a GF100 kinda fail, it's slower than its previous generation core for core, clock for clock (One Fermi shader does less than a GT200b shader in the same clock speed, and each shader is practically a core) and it sucks on the power, how many people remember the Fermi/woodscrews debacle now a fixed version is launched? Same thing can easily happen here.

Quote:
Originally Posted by Track View Post
Yeah, but that was a very long time ago and Intel had no competition, i.e. no reason to better themselves.

AMD had the time, the resources, the incentive..

And we get this.

It's like.. how can their GPU side be so clever while their CPU side is as dumb as needles.
Actually, AMD was beating Intel for quite a bit before the P4 came out...The last time Intel was in the C2D position vs AMD was with the K5.

AMD beat Intel to 130nm and 1Ghz.
I wish those days were back, it was AMD and Intel neck and neck pretty much, if you wanted a chip it was actually difficult to choose instead of "Under $150? Phenom II, otherwise 2500k or 2600k" near universally.

I like how they've repeated what I've been saying since BD launched too, about the cache. They should keep the amount of L2 cache they have but split it into 1Mb chunks that can be accessed independently, rather than 2MB chunks that are slow.

I love it when this type of article is posted, you can easily tell who read it and who didn't.
Edited by Brutuz - 10/25/11 at 4:37am
    
CPUMotherboardGraphicsRAM
Intel Core i5 3570k @ 4.5Ghz ASRock Z77 Pro3 Powercolor Radeon HD7950 3GB @ 1150/1350 4x4GB G.Skill Ares 2000Mhz CL9 
Hard DriveHard DriveHard DriveHard Drive
Samsung 840 250GB Western Digital Black 1TB WD1002FAEX Seagate Barracuda 3TB ST3000DM001 Samsung Spinpoint EcoGreen 2TB 
Optical DriveCoolingCoolingCooling
Pioneer DVR-220LBKS Noctua NH-D14 Scythe Gentle Typhoon 1850rpm Corsair AF140 Quiet Edition 
CoolingOSMonitorMonitor
Arcitc Cooling Acclero Twin Turbo II Arch Linux x86-64, amdgpu BenQ G2220HD BenQ G2020HD 
KeyboardPowerCaseMouse
Ducky Shine III Year of the Snake, Cherry Blue Silverstone Strider Plus 600w CoolerMaster CM690 II Black and White SteelSeries Sensei Professional 
Mouse PadAudioOther
Artisan Hien Mid Japan Black Large ASUS Xonar DX NZXT Sentry Mesh 30w Fan Controller 
  hide details  
Reply
    
CPUMotherboardGraphicsRAM
Intel Core i5 3570k @ 4.5Ghz ASRock Z77 Pro3 Powercolor Radeon HD7950 3GB @ 1150/1350 4x4GB G.Skill Ares 2000Mhz CL9 
Hard DriveHard DriveHard DriveHard Drive
Samsung 840 250GB Western Digital Black 1TB WD1002FAEX Seagate Barracuda 3TB ST3000DM001 Samsung Spinpoint EcoGreen 2TB 
Optical DriveCoolingCoolingCooling
Pioneer DVR-220LBKS Noctua NH-D14 Scythe Gentle Typhoon 1850rpm Corsair AF140 Quiet Edition 
CoolingOSMonitorMonitor
Arcitc Cooling Acclero Twin Turbo II Arch Linux x86-64, amdgpu BenQ G2220HD BenQ G2020HD 
KeyboardPowerCaseMouse
Ducky Shine III Year of the Snake, Cherry Blue Silverstone Strider Plus 600w CoolerMaster CM690 II Black and White SteelSeries Sensei Professional 
Mouse PadAudioOther
Artisan Hien Mid Japan Black Large ASUS Xonar DX NZXT Sentry Mesh 30w Fan Controller 
  hide details  
Reply
post #42 of 292
AMD isn't very good at designing things from the get go, but its definitely very good at fixing ****. I remember when phenom I came out, it would barely hit 2.9, and everyone was like "this sucks just give me a die shrunk 6400+ x4"

But they stuck with it, and fixed the hell out of them and phenom II is actually a pretty damn good chip.

What i think would fix bulldozer:
-extra 128bit fmac per module, eliminating the idling fmac problem when 128bit instructions come in sandwiched between 256bit ones.
-double the L1 data cache
-half the L3 and possibly shrink L2
-widen module scheduler to be able to put out 3+ ops/cycle to each integer core, like SB and deneb have been doing for years. (correct me if im wrong on this)
-add an integer pipe per integer core so thats possible
-design their own god damned interconnects, or re-write the program that does it for them.

Its a rough interpretation, and im no microprocessor engineer, but i think if they made changes along those lines they'd have an incredibly nice chip. Its single thread performance still wouldnt be much to talk about, but *proper* implementation of dx11 is doing a pretty good job of scaling games up for more cores, the bulldozer benchmarks in the bf3 beta i saw weren't too bad?

AMD has the right idea here, they just need to perfect it. If single thread performance wasnt so dismal, this chip would stomp all over the competition.


Maybe bulldozer was just clearing the path for the piledriver to come in and drive a stake into the heart of intel, yeah? Excavator will get rid of the corpse?

Lolololol.
Edited by Rommel - 10/25/11 at 4:52am
The Desert Fox
(14 items)
 
  
CPUMotherboardGraphicsRAM
Phenom x6 1090T ASUS M5A97 Evo AM3+ EVGA GTX570 800/2000 2x4gb Corsair Vengeance 1600 
Hard DriveOptical DriveOSMonitor
2x WD 6400AALS RAID/Mushkin 60gb SSD Sony Optiarc Windows 7 Professional x64 24" Asus CCFL 2ms (LED screens suck, seriously.) 
KeyboardPowerMouseMouse Pad
Logitech K800 Wireless Illuminated Corsair TX650 52amp 12v Microsoft Wireless BlueTrack 6000 Wal-mart $1.99 
Audio
AKG K701 
  hide details  
Reply
The Desert Fox
(14 items)
 
  
CPUMotherboardGraphicsRAM
Phenom x6 1090T ASUS M5A97 Evo AM3+ EVGA GTX570 800/2000 2x4gb Corsair Vengeance 1600 
Hard DriveOptical DriveOSMonitor
2x WD 6400AALS RAID/Mushkin 60gb SSD Sony Optiarc Windows 7 Professional x64 24" Asus CCFL 2ms (LED screens suck, seriously.) 
KeyboardPowerMouseMouse Pad
Logitech K800 Wireless Illuminated Corsair TX650 52amp 12v Microsoft Wireless BlueTrack 6000 Wal-mart $1.99 
Audio
AKG K701 
  hide details  
Reply
post #43 of 292
Quote:
Originally Posted by Brutuz View Post
Hate to obviously not read the source or even most of the topic, eh?
It's been covered, there's no turbo mode.
Care to use logic?
If all threads are used, the turbo mode will apply a boost of 100 MHz on all 4 cores in terms of the 2600k.
Theres NO way a 100 MHz lack would affect performance that much, that it would drop from ~7 to 5,9 in cinebench thats a full 16% drop. In fact, a 100 MHz difference wont change much and most likely improve results by 2-3%, thats nothing.

Something IS wrong with that value and im sure it may just be a typo and is supposed to mean 6,91.

@ article
Very interesting read, but theres a thing I dont understand:

Quote:
A lower latency L2 cache, possibly combined with a larger L1 data cache, would likely result in a significant speed increase, while cutting down on the total amount of L2 would save die space and reduce cost.
If thats the case, why didnt AMD do it in the first place? I mean they've got engineers that have been in the business for years.. if various forumers and hardware reviewers can figure this out, why cant AMD?
Edited by toX0rz - 10/25/11 at 5:25am
Rebel
(13 items)
 
  
CPUMotherboardGraphicsRAM
Intel Core i5-2500k 4,6 GHz @ 1,32 V AsRock Z68 Extreme3 Gen3 PNY GeForce GTX 660 Ti SLI @ 1228 / 3312 MHz 8GB G.Skill Ripjaws DDR3-1600 
Hard DriveOSMonitorPower
Samsung 1TB + 500GB Windows 7 Ultimate 64 Asus VH22H Xilence XQ 850W with a Scythe Kama Flex 
CaseMouse
Aerocool VS-9 Xmas Edition :P Logitech G500 
  hide details  
Reply
Rebel
(13 items)
 
  
CPUMotherboardGraphicsRAM
Intel Core i5-2500k 4,6 GHz @ 1,32 V AsRock Z68 Extreme3 Gen3 PNY GeForce GTX 660 Ti SLI @ 1228 / 3312 MHz 8GB G.Skill Ripjaws DDR3-1600 
Hard DriveOSMonitorPower
Samsung 1TB + 500GB Windows 7 Ultimate 64 Asus VH22H Xilence XQ 850W with a Scythe Kama Flex 
CaseMouse
Aerocool VS-9 Xmas Edition :P Logitech G500 
  hide details  
Reply
post #44 of 292
Quote:
Originally Posted by toX0rz View Post
Care to use logic?
If all threads are used, the turbo mode will apply a boost of 100 MHz on all 4 cores in terms of the 2600k.
Theres NO way a 100 MHz lack would affect performance that much, that it would drop from ~7 to 5,9 in cinebench thats a full 16% drop. In fact, a 100 MHz difference wont change much and most likely improve results by 2-3%, thats nothing.

Something IS wrong with that value and im sure it may just be a typo and is supposed to mean 6,91.
You'd be surprised how much of a difference that clock speed can make, as it alters other internal ones and enables faster core communication, but yeah, it is probably a typo.
    
CPUMotherboardGraphicsRAM
Intel Core i5 3570k @ 4.5Ghz ASRock Z77 Pro3 Powercolor Radeon HD7950 3GB @ 1150/1350 4x4GB G.Skill Ares 2000Mhz CL9 
Hard DriveHard DriveHard DriveHard Drive
Samsung 840 250GB Western Digital Black 1TB WD1002FAEX Seagate Barracuda 3TB ST3000DM001 Samsung Spinpoint EcoGreen 2TB 
Optical DriveCoolingCoolingCooling
Pioneer DVR-220LBKS Noctua NH-D14 Scythe Gentle Typhoon 1850rpm Corsair AF140 Quiet Edition 
CoolingOSMonitorMonitor
Arcitc Cooling Acclero Twin Turbo II Arch Linux x86-64, amdgpu BenQ G2220HD BenQ G2020HD 
KeyboardPowerCaseMouse
Ducky Shine III Year of the Snake, Cherry Blue Silverstone Strider Plus 600w CoolerMaster CM690 II Black and White SteelSeries Sensei Professional 
Mouse PadAudioOther
Artisan Hien Mid Japan Black Large ASUS Xonar DX NZXT Sentry Mesh 30w Fan Controller 
  hide details  
Reply
    
CPUMotherboardGraphicsRAM
Intel Core i5 3570k @ 4.5Ghz ASRock Z77 Pro3 Powercolor Radeon HD7950 3GB @ 1150/1350 4x4GB G.Skill Ares 2000Mhz CL9 
Hard DriveHard DriveHard DriveHard Drive
Samsung 840 250GB Western Digital Black 1TB WD1002FAEX Seagate Barracuda 3TB ST3000DM001 Samsung Spinpoint EcoGreen 2TB 
Optical DriveCoolingCoolingCooling
Pioneer DVR-220LBKS Noctua NH-D14 Scythe Gentle Typhoon 1850rpm Corsair AF140 Quiet Edition 
CoolingOSMonitorMonitor
Arcitc Cooling Acclero Twin Turbo II Arch Linux x86-64, amdgpu BenQ G2220HD BenQ G2020HD 
KeyboardPowerCaseMouse
Ducky Shine III Year of the Snake, Cherry Blue Silverstone Strider Plus 600w CoolerMaster CM690 II Black and White SteelSeries Sensei Professional 
Mouse PadAudioOther
Artisan Hien Mid Japan Black Large ASUS Xonar DX NZXT Sentry Mesh 30w Fan Controller 
  hide details  
Reply
post #45 of 292
Quote:
Originally Posted by toX0rz;15446060 
Very interesting read, but theres a thing I dont understand:


If thats the case, why didnt AMD do it in the first place? I mean they've got engineers that have been in the business for years.. if various forumers and hardware reviewers can figure this out, why cant AMD?

At a guess? They already did, months ago, when they thoroughly tested the first chips at the end of the initial production pipeline.

They were likely disappointed in the results, but some executive probably asked, "This Bulldozer--is it good enough that we could turn a profit on it?"

To which an engineer answered, "Barely. And only in limited numbers."

And the executive let out a long sigh. "Finally, something we can sell," he said, "Guess which salaried employees just earned themselves 60 hour weeks instead of 80 hour weeks?"

"Sir, my wife is threatening to leave me for a guy working at Int--" The executive's finger was at his lips.

"Shh," the boss man said, "It's better if you don't talk."
    
CPUMotherboardGraphicsRAM
FX 8350 GA-990FXA-UD3 Zotac 670 corsair vengeance 
Hard DriveHard DriveOSMonitor
WD6401AALS x2 raid 0 Crucial M4 win 7 pro x64 hp zr2740w 
KeyboardPowerCaseMouse
thermaltake meka g1 corsair tx750 nzxt gamma logitech g600 
Audio
HT Omega Claro+ 
  hide details  
Reply
    
CPUMotherboardGraphicsRAM
FX 8350 GA-990FXA-UD3 Zotac 670 corsair vengeance 
Hard DriveHard DriveOSMonitor
WD6401AALS x2 raid 0 Crucial M4 win 7 pro x64 hp zr2740w 
KeyboardPowerCaseMouse
thermaltake meka g1 corsair tx750 nzxt gamma logitech g600 
Audio
HT Omega Claro+ 
  hide details  
Reply
post #46 of 292
Quote:
Originally Posted by Brutuz View Post
Read the source, it's the cache...Just like I've been saying since a day after it launched.



Hate to obviously not read the source or even most of the topic, eh?
It's been covered, there's no turbo mode.



Not really, there's been worse, case in point, the P4. That thing, despite having a larger clock advantage, was monumentally slow while using around double the power at launch.

And its faster than Thuban, just not when less than 8 cores are enabled and it's not using SSE.

Which, coincidentally, I see BD as a GF100 kinda fail, it's slower than its previous generation core for core, clock for clock (One Fermi shader does less than a GT200b shader in the same clock speed, and each shader is practically a core) and it sucks on the power, how many people remember the Fermi/woodscrews debacle now a fixed version is launched? Same thing can easily happen here.
.
Except you bought a GF100 didn't you, lol? So how much of fail was GF100 hint not much. Considering nvidia stock price doubled over night. Where you born when the P4 launched!

There will be no magic bios fix. AMD does not care. They have been working on this since 2006. Thye knew exactly what they went to market with. Again look at the stock price.
Its a server chip they are passing off as mainstream. That's why John F was shooting mouth off. What does he hock/sell? server processors enough said!
MyCleanPC
(13 items)
 
  
CPUMotherboardGraphicsRAM
i7 4770k ASUS MAXIMUS VI EVGA master blaster Corsair Vengence  
Hard DriveOptical DriveOSMonitor
OCZ SSD raid0 samsung Win 7 Samsung 
PowerCase
Enermax rev 1050 Stacker 832 
  hide details  
Reply
MyCleanPC
(13 items)
 
  
CPUMotherboardGraphicsRAM
i7 4770k ASUS MAXIMUS VI EVGA master blaster Corsair Vengence  
Hard DriveOptical DriveOSMonitor
OCZ SSD raid0 samsung Win 7 Samsung 
PowerCase
Enermax rev 1050 Stacker 832 
  hide details  
Reply
post #47 of 292
I'm the author of the piece in question. 2600K scores are lower because I turned Turbo Mode off. Turbo Mode increases the clock speed of the chip by 400MHz, or by 11%. The point was not to penalize any processor, but to gain a clearer picture of Bulldozer's IPC.

Bulldozer scales more poorly than Thuban because its an SMT design. There's nothing inherently wrong with that. Remember, the original idea was that AMD would capture most of the advantage of a traditional dual-core chip in a much smaller die. If the caches on BD weren't so huge, we'd see this reflected in the chip's die size.

A 2M/4C BD configuration is ~80% as fast as a 4M/4C configuration. A 20% penalty for combining as much core logic as BD did is actually pretty good. The real problem is the IPC loss.

Here's why Bulldozer "doesn't make sense" in two easy points.

1) AMD built a chip designed to save die space -- a very smart move -- then loaded it down with huge amounts of sluggish cache.

2) It has a small, write-through L1 backed by very high-latency L2.

Why does it look the way it does? Clearly something went wrong. Everything we know suggests AMD truly expected to get something very different than what came out of the fab. Since we don't know *what* exactly happened, it's extremely difficult to predict how easy the problem is to fix.
post #48 of 292
Quote:
Why does it look the way it does? Clearly something went wrong. Everything we know suggests AMD truly expected to get something very different than what came out of the fab. Since we don't know *what* exactly happened, it's extremely difficult to predict how easy the problem is to fix.
I would say - not that easy . They delayed BD launch for quite a while and it looks like they could not fix it in that amount of time . I am not 100% sure they will be able to fix all the issues even in piledriver .
post #49 of 292
at this point who cares why bulldozer is a disappointment ,i just want to know when AMD will announce the Operation Scorpius bulldozer winners to find out if i will get a chance to play with one
they said on or around Oct 17th,it's now the 25th.i just hope its not delayed(scollins please dont say they never delayed it)
Edited by radaja - 10/25/11 at 9:40am
    
CPUMotherboardGraphicsRAM
AMD A6-3650 8.32GHz so close Gigabyte A75M-UD2H sapphire HD3450 2x2GB DDR3-1600 
Hard DriveOptical DriveOSMonitor
western digital 320GB ASUS DRW-24B1ST DVD Microsoft Windows 7 Ultimate x64 ASUS LED 23.6" MS246H 
PowerCase
Corsair 400W cardboard box w/lots of holes for fans 
  hide details  
Reply
    
CPUMotherboardGraphicsRAM
AMD A6-3650 8.32GHz so close Gigabyte A75M-UD2H sapphire HD3450 2x2GB DDR3-1600 
Hard DriveOptical DriveOSMonitor
western digital 320GB ASUS DRW-24B1ST DVD Microsoft Windows 7 Ultimate x64 ASUS LED 23.6" MS246H 
PowerCase
Corsair 400W cardboard box w/lots of holes for fans 
  hide details  
Reply
post #50 of 292
Depends on which issues you mean. Yield problems and TDP reductions should be do-able with BD--those are more dependent on GF's overall 32nm process. Fixing the chip's IPC issues...yeah. That could take awhile. I don't expect Piledriver to be a miracle cure.
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Hardware News
Overclock.net › Forums › Industry News › Hardware News › [ExtremeTech] Analyzing Bulldozer: Why AMD’s chip is so disappointing