AMD Processors
Decrease font size
Increase font size
Topic Title: Transistor Count
Topic Summary:
Created On: 04/01/2008 12:39 PM
Status: Read Only
Linear : Threading : Single : Branch
1 2 Next Last unread
Search Topic Search Topic
Topic Tools Topic Tools
View similar topics View similar topics
View topic in raw text format. Print this topic.
 04/01/2008 12:39 PM
User is offline View Users Profile Print this message

Author Icon
Firestrider
Senior Member

Posts: 374
Joined: 03/24/2008

Is transistor count a pretty good indicator of CPU and GPU performance (the more the better)?
 04/01/2008 01:04 PM
User is offline View Users Profile Print this message

Author Icon
Decembermouse
Mad Scientist

Posts: 2149
Joined: 03/31/2004

I'd say there is a correlation.

But it also depends on bus type, architecture, memory, clock speed, etc. Basically don't buy anything because of transistor count, but you can brag about how many it has once you own it.

-------------------------
ASRock M3A785GMH/128M
Phenom II X4 925c2 @ 2.80GHz (95W)
OCZ Platinum PC3-1333 2x2GB 7-7-7-21
WD5000AADS
Raidmax RX-500AF
Check it out.
 04/01/2008 04:43 PM
User is offline View Users Profile Print this message

Author Icon
Firestrider
Senior Member

Posts: 374
Joined: 03/24/2008

Do you know of an example where less transistors actually gave better performance?
 04/02/2008 01:40 AM
User is offline View Users Profile Print this message

Author Icon
Overmind
Assimilator

Posts: 8052
Joined: 01/22/2004

The number is dependent on the level of manufacutring technology (nm).

-------------------------

World's best Red Alert 2: Yuri's Revenge mod and Star Trek: Starfleet Command 3 mod: Overmind.ro
 04/02/2008 07:46 PM
User is offline View Users Profile Print this message

Author Icon
technic58
Minister of Spare Parts

Posts: 13953
Joined: 07/25/2004

Yeah you can't really say more is definitely better...I mean I don't have specific examples...but no more is not always better.

But Overmind...that's confusing. Doesn't a 90nm CPU have the same number of transistors as an otherwise identical 65nm CPU? I was thinking manufacturing process in nm just referred to the size of the actual transistors being used?

-------------------------
My GT430 Is Folding For AMD Team 34106
FOR A GOOD CAUSE---Accepting Donations of New, Used, Working, Nonworking PC Components
Thanks to KLEETUS for his generous donation of components!!
 04/03/2008 01:17 AM
User is offline View Users Profile Print this message

Author Icon
Decembermouse
Mad Scientist

Posts: 2149
Joined: 03/31/2004

Yeah I'm pretty sure downsizing from 130-->90nm or what-have-you just makes each transistor 31% smaller or so.

90-->65nm = 28% smaller
65-->45nm = 31% smaller
45-->32nm = 29% smaller

-------------------------
ASRock M3A785GMH/128M
Phenom II X4 925c2 @ 2.80GHz (95W)
OCZ Platinum PC3-1333 2x2GB 7-7-7-21
WD5000AADS
Raidmax RX-500AF
Check it out.
 04/03/2008 01:23 AM
User is offline View Users Profile Print this message

Author Icon
Overmind
Assimilator

Posts: 8052
Joined: 01/22/2004

Originally posted by: technic58
But Overmind...that's confusing. Doesn't a 90nm CPU have the same number of transistors as an otherwise identical 65nm CPU? I was thinking manufacturing process in nm just referred to the size of the actual transistors being used?

At the same die size, it doesn't. At different die sizes, yes, it has.

The mnf. process not only referres to transistor sizes, but internal spacement, internal link parhways and almost everything else except the external conectors and the direct links to them.

-------------------------

World's best Red Alert 2: Yuri's Revenge mod and Star Trek: Starfleet Command 3 mod: Overmind.ro
 04/03/2008 01:31 AM
User is offline View Users Profile Print this message

Author Icon
Decembermouse
Mad Scientist

Posts: 2149
Joined: 03/31/2004

So basically, only if everything in there is scaled down about equally (all those things you listed, OM) would my very simple math really apply. Of course, spacing might have been improved moving from 65-->45, more than say internal pathways, and it would all equal out in the end anyway.

-------------------------
ASRock M3A785GMH/128M
Phenom II X4 925c2 @ 2.80GHz (95W)
OCZ Platinum PC3-1333 2x2GB 7-7-7-21
WD5000AADS
Raidmax RX-500AF
Check it out.
 04/03/2008 02:35 AM
User is offline View Users Profile Print this message

Author Icon
Xajel
Case Modder

Posts: 1554
Joined: 10/08/2003

going down from 90nm to 65nm or 45nm will not change the transistor count, but it will allow the transistors to work faster, so more performance, but this is not a law as some times going down with the process will needs more transistors, for example in power circuits as smaller transistors are able to handle lower power/current than bigger, so here you will need more transistors to have the same power/current, + some times the faster clocks or smaller process will need some change in trasistors count or change in the design in order to eliminate some problems like current leakage or so on.

there's no relation with transistors count to performance, for example, Cell/BE CPU compared to any current x86 CPU, you will see that Cell/BE is much more powerfull than that x86 even thought it has less transistors count ( 250mil for 65nm ) compared to Core 2 Quad with more than dual of that ( 582mil @ 65nm ), the Cell/BE has 5 times more power in GFLOPS than Core 2 Quad, even if you count that Cell/BE works at 3.2GHz and Core2Q works at 2.7GHz, the power diff is more than the clock diff.

another example G80 nVIDIA GPU, has 681 - 686 mil and peaks at 330GFLOPS, ATi R600 has 720mil and peaks at 450GFLOPS

so lets count transistors per GFLOPS ( in mil transistros per GFLOPS ) - lower is better in case of required transistors per GFLOPS

CELL/BE = 1
Core 2 Q = 11.64
G80 = 2.07
R600 = 1.6

reason ?
with the same transistor count you can make another arch. that is much powerfull, that why Cell/BE can make 1GFLOPS out of 1mil transistors, infact, if IBM added more SPE's to the CELL/BE will have more power per transistors, and you will need less than 1mill transistors to make a GFLOPS !!
 04/03/2008 03:30 AM
User is offline View Users Profile Print this message

Author Icon
Firestrider
Senior Member

Posts: 374
Joined: 03/24/2008

So would you measure CPU/GPU performance in FLOPS? If that's the case would my 4000+@2.8Ghz be equivalent to a E6600? Would a Phenom 9500 be equivalent to a Q6400 (X3210)?
 04/03/2008 04:03 AM
User is offline View Users Profile Print this message

Author Icon
Xajel
Case Modder

Posts: 1554
Joined: 10/08/2003

nope, FLOPS is a specific type of performance measurment, Floating Point to be exact, FLOPS stands for ( FLoating point Operations Per Second ), and coz it's only for Floating Points, performance may vary in other tests... but in todays use, GFLOPS is a well known performance measerment for chips, but as it's only 1 type, CPU's and GPU's will have diff. performance scale depending on the bench you want... as they have diff. arch. too, GPU's will handle graphics much faster than a x86 CPU even if that GPU had half the transistor count compared to the CPU...
 04/03/2008 04:10 AM
User is offline View Users Profile Print this message

Author Icon
Firestrider
Senior Member

Posts: 374
Joined: 03/24/2008

Ok, what would be a better performance indicator for a CPU in gaming benchmarks (MIPS / integer operations)? I know quad core CPUs give better scores in 3DMark06 but most games don't utilize all of the four cores.
 04/03/2008 04:35 AM
User is offline View Users Profile Print this message

Author Icon
Xajel
Case Modder

Posts: 1554
Joined: 10/08/2003

there's no performance test that will give you exact performance comparision as games/games engines diff with the way and needs of resources some will need more AI than physics, and some needs physics more...

if you remembred the days before G80 & R600, there was Pixel Shaders & Vertix Shaders, both nVIDIA & ATi had thire own POV of the ratio between PS units and VS units, this resulted into mixed out results in games, even in scence in the same game it self... the same here happend

maybe 3D Mark is agood indicator of performance but you can take it's result for every thing, todays GPU's like 8xxx/9xxx from NV and HD3xxx from AMD has different performance in games, the first is better in some games, and in other games the second is better, while if you take the 3D Mark results, you will see that AMD hold the crown here
 04/03/2008 07:15 AM
User is offline View Users Profile Print this message

Author Icon
Overmind
Assimilator

Posts: 8052
Joined: 01/22/2004

Originally posted by: Xajel

going down from 90nm to 65nm or 45nm will not change the transistor count,

But you can put more of them on the same surface.

-------------------------

World's best Red Alert 2: Yuri's Revenge mod and Star Trek: Starfleet Command 3 mod: Overmind.ro
 04/03/2008 10:02 AM
User is offline View Users Profile Print this message

Author Icon
Xajel
Case Modder

Posts: 1554
Joined: 10/08/2003

Originally posted by: Overmind

Originally posted by: Xajel



going down from 90nm to 65nm or 45nm will not change the transistor count,



But you can put more of them on the same surface.


yep, but I meant in the same design

if you wanted to go to lower process to increase the performance, then ou have to change the design

but having a different design for more performance doesn't always mean higher transistor count
 04/03/2008 10:25 AM
User is offline View Users Profile Print this message

Author Icon
Overmind
Assimilator

Posts: 8052
Joined: 01/22/2004

The safest way to tech-up is to decrease the fab. process using the same design because you know extremely well what that design can do. As soos as the new decrease works with that design as it suppose to, then you implement a new design on that fab.
The only setback is time.

-------------------------

World's best Red Alert 2: Yuri's Revenge mod and Star Trek: Starfleet Command 3 mod: Overmind.ro
 04/03/2008 12:13 PM
User is offline View Users Profile Print this message

Author Icon
Xajel
Case Modder

Posts: 1554
Joined: 10/08/2003

I agree, and that was the problem with AMD

AMD did not mastered the 65nm process but they were forced to use it in a new design ( K10 ) as it's a quadcore and making it with a well mastered 90nm will mean very high power requirement...

so they had a new design + a new process

the bad luck for AMD, is thier 65nm had a lot of problems, that's why AMD is focusing on 45nm right now rather than optimizing 65nm more and more... the other reason is AMD want to come close to Intel with Process technology, that's why they decided to make the jump every 18 months rather than 24 months
 04/03/2008 12:38 PM
User is offline View Users Profile Print this message

Author Icon
Charlie22911
Voodoo Programmer

Posts: 3278
Joined: 04/09/2004

Originally posted by: Firestrider

Do you know of an example where less transistors actually gave better performance?


intels PIII tulation vs the P4 willimette... northwood vs prescott... tbred vs willimette...

-------------------------
Desktop:
Phenom II x6 1055T @ 4GHz | 4x2GB Patriot DDR3 1600 @ 2000 | 3x AMD Radeon HD6970 Crossfire

Laptop:
Core i7 2960xm @ 4.2Ghz | 4x4GB Kingston DDR3 1866 | 2x GTX 580m SLI OC 725/1450/1500
 04/04/2008 02:25 AM
User is offline View Users Profile Print this message

Author Icon
Overmind
Assimilator

Posts: 8052
Joined: 01/22/2004

...and Pentium 1 vs. Celeron.

-------------------------

World's best Red Alert 2: Yuri's Revenge mod and Star Trek: Starfleet Command 3 mod: Overmind.ro
 04/06/2008 05:53 PM
User is offline View Users Profile Print this message

Author Icon
chiplist
Member

Posts: 149
Joined: 12/29/2006

Moore's Law states that the number of transistors on a chip doubles every two years. As a result software performance in general is growing exponentially too, be it in a slower rate.

Like Xajel says, more transistors allow you to improve your architecture (e.g. increase the size of your caches), or even to come up with a new architecture (add hyper-threading, wider buses, extra instructions, extra speculative execution, etc.). So, yes, more transistors mean more performance, but processor architects have to translate the larger transistor count into performance.

Until recently, programmers got this performance gain for free. Entering the multi-core era, developers will have to parallelize their code to improve the performance of their applications. So now they've become part of this transistor count to performance translation process.

-------------------------
Adrian Offerman
Maintainer of The ChipList
Statistics
112018 users are registered to the AMD Processors forum.
There are currently 0 users logged in.

FuseTalk Hosting Executive Plan v3.2 - © 1999-2014 FuseTalk Inc. All rights reserved.



Contact AMD Terms and Conditions ©2007 Advanced Micro Devices, Inc. Privacy Trademark information