K7/K8 Inventor Back at AMD

As far as all the 2,4,6,8 core vs software talk goes: I'm not going to pretend I know jack about coding for multi-threading, but having spoken to several programmers talking about scientific programming (wind field calculations, computational fluid mechanics, molecular/orbital modeling) multi-threading and parallelization is not as easy as just smashing some crap into a compiler and having it split threads up. The working data set needs to actually benefit from parallel computations. You can split all the threads you want but if they're sitting waiting on data because there is a dependency it doesn't mean shit. Sure it helps if you've got lots of different programs running and crunching in the background to keep them all happy, but if you want to get benefits out of 1 single program they won't all scale, and thats not always the programmer's fault.

Fully agreed.
 
Anything that was not SSE optimized was buried by the AMD chips at the time.
The performance was mixed across a large range of applications during the time of Willamette and Northwood vs K7 and SSE was one area where AMD was performing slightly better per clock.

You can look through any number of Athlon XP vs P3, Athlon XP vs Willamette or Athlon XP vs Northwood P4 reviews to you can verify what I wrote is correct. I glanced through several while making my post.
 
Meh. I disagree. My Chromebook is an i3/Sandy Bridge based Celeron. 1.1Ghz Dualcore and it does me more than well enough to be my daily driver. I only use my desktop for video editing, encoding and animation/rendering when not playing games. I have absolutely no problems using the Chromebook (has XUbuntu 14.04 installed to it instead of ChromeOS, mind you)

System was nigh unusable at 2GB Ram installed. Upped it to 10GB and everything runs great.

I believe we have long since passed the point in which the CPU is capable of being anything that would set back a system. Other factors, such as the GPU, disk throughput and amount of ram are far more important once you have any kind of dualcore CPU, IMHO.

So dual-core isn't multicore? :p
 
A lot of misleading things said about high core count and multi-threaded software from people with only 4-core. 16-core loves Frostbite multi-threaded engine in Battlefield 4 and with a budget R9 270X it nearly maxes out the 200 fps cap at ~190 fps on Test Range. HDD disk I/O seem to benefit too since I usually spawn early with the SSD equipped.

Links to any head to head reviews? I need a new CPU, and primarily gaming on BF, if I can go cheaper with AMD and actually get better performance, that's a win.
 
Links to any head to head reviews? I need a new CPU, and primarily gaming on BF, if I can go cheaper with AMD and actually get better performance, that's a win.

Your 2600k OC'd is going to do better then anything AMD has out right now.
 
I loved those K7's. I still have a couple of PC's with them in sitting in a room. One is still hooked up on an empty desk. I'm tempted to blow off the dust and fire it up, lol!
 
same song five times or is it 8 now with amd.
The engineers take center stage and they fly then when they get off the ground they let them go and coast until they suck then after a long while someone says hey maybe we do need some great engineers after all. wash rinse repeat.
 
Your 2600k OC'd is going to do better then anything AMD has out right now.

Haven't edited my sig...that system is dead. Not sure whether to try and replace what I had, slightly cheaper, go a bit more expensive and get current-gen Intel, or, never thought it was an option, but maybe look at AMD.
 
Looks at picture in article..."What the hell happened to Kevin Sorbo???"
 
Dump faildozer in the trash (or keep it for server use as uber core Opterons), go back to K10, die shrink the hell out of it and clock it up to 5ghz+. How hard is that?

APUs are cool and all for laptops and cheapie systems but it's not for everyone.
 
Oh hey another thought, why can't they have some kind of HAL level shim driver that can virtually combine different cores into a super core? Turn a 8 core FX into a wickedly fast 2 core? If the software companies can't write multicore code properly, turn it around and let the driver handle it?

I am not a CPU engineer, so I have no idea how to do that, but it sounded good while I was typing it.
 
"At least at first". Still waiting, how many years later, and even 4 cores are barely used. 7 years since the Q6600...

No kidding...for what I care about (gaming) more than 4 cores is meaningless. Even in a lot of cases still, more than 2 cores is meaningless.
 
And here I thought that AMD's CPUs going to hell in a hasbasket was the genius decision to keep reducing the amount of lvl 1 cache from one revision to the next.
The socket 939s and 754s had 128 KB of lvl 1 cache (64 KB most frequently used instructions + 64 KB open for anything else) per core (256 KB lvl 1 cache for 2 cores) and our current gen Vishera has 16 KB per core (128 KB total).
Don't get me wrong, I'm an AMD guy, but every time they've cut lvl 1 cache performance per core has gotten worse. Now I'm no engineer, but if every time you cut the lvl 1 cache things get worse do you REALLY want to keep cutting it? Honestly, it was all downhill after socket 939.
 
Oh hey another thought, why can't they have some kind of HAL level shim driver that can virtually combine different cores into a super core? Turn a 8 core FX into a wickedly fast 2 core? If the software companies can't write multicore code properly, turn it around and let the driver handle it?

I am not a CPU engineer, so I have no idea how to do that, but it sounded good while I was typing it.

This would not work in software. However I wish for the desktop processors that they would have combined the power of the two weak cores of each module creating a strong core with 2 threads like Intel has done with HT instead of two independent weak cores relying on increased frequency to work around the reduced processing capability of each core (versus previous designs). Remember AMD said that they only needed 125% of the transistors of a Phenom II core to make a bulldozer module with 2 cores. I do not believe the Phenom II was extremely inefficient so to me the reduction in transistors per core is at least part of the problem. Yes some of the transistors can be cut as obsolete CPU instructions can be emulated instead of executed directly in hardware. And yes there is some sharing of functionality inside the module.
 
Now I'm no engineer, but if every time you cut the lvl 1 cache things get worse do you REALLY want to keep cutting it? Honestly, it was all downhill after socket 939.

Reducing L1 cache means reducing area and power. Indirectly this also means increasing yield and decreasing manufacturing cost. These are cost-benefit decisions that most engineering projects have to make.

It's also worth noting that it is very difficult as a consumer to actually know the impact of that sort of change as it's unusual for them to be done in isolation. The CPU architects write cycle accurate simulators to understand performance implications of changes- they are not done in the dark.
 
Back
Top