AMD FX-8350 Pushed to 8.1 GHz via Extreme Overclocking by Der8auer

[seronx]

I've been saying this since 2011, but that's the AMD equilent of Pentium 4.

I wouldn't say that, just because it has higher clocks doesn't mean it is built in the same way.

A lot of stuff in Bulldozer lost what it was initially based on. It was a beefy big core when it was only a 32-bit processor w/ units that would only pop eventually in Zen.
32-bit core => https://patents.google.com/patent/US6240503B1
L1i => 128 KB 2-way associative cache
L0i => 512 byte fully associative cache
L1d => 128 KB 2-way associative cache
{ 2x 32-bit ALUs
2x 32-bit FPUs
2x 32-bit MMXs
1x LD AGU
1x ST AGU ] x2
Shared Scheduler/Retirement => Independent Renamers

Similar core => https://patents.google.com/patent/US6553482B1
L1i => 256 KB 4-way associative cache
Retirement queue => 128-entries
Shared Retirement/Renamer => Independent Schedulers

Both shared L1d/LSU.

 

[EarthDog]

I wouldn't say that, just because it has higher clocks doesn't mean it is built in the same way.

I don't think that was her point, lol. Think 30,000' view, bud.

 

[Ruru]

I don't know crap about CPU architectures and how they compare, but my point is that with FX, AMD did exactly the same thing that Intel did with Pentium 4, more clocks without caring about the heat output and power draw.

I don't think that was her point, lol. Think 30,000' view, bud.

HER?!?! Whoa!

 

[EarthDog]

I don't know crap about CPU architectures and how they compare, but my point is that with FX, AMD did exactly the same thing that Intel did with Pentium 4, more clocks without caring about the heat output and power draw.


HER?!?! Whoa!

I mean, you have a female avatar and name....

my fault, dude.

 

[Ruru]

I mean, you have a female avatar and name....

my fault, dude.

Just my waifu.

 

[xtreemchaos]

just in case anyone want to know im a male cat ...

 

[EarthDog]

Just my waifu.

I had to look that up, lol...

My last 'anime' was Voltron in the late 1980s. lol

 

[seronx]

I don't know crap about CPU architectures and how they compare, but my point is that with FX, AMD did exactly the same thing that Intel did with Pentium 4, more clocks without caring about the heat output and power draw.

Actually, they did.

They ported a 2x2 32-bit ALU, 2x2 32-bit MMX, 2x2 32-bit FPU, 2x1 LD AGU, 2x1 ST AGU; 1 data cache. To a 2x2 64-bit ALU, 1x2 128-bit MMX, 1x2 128-bit FMAC, 2x2 LD/ST AGU; 2 data cache design. Sure somethings vanished or were never implemented. It was still better than K7/K8/10h.

It also achieved high frequency at lower power and lower leakage than Greyhound+/Husky in Deneb/Llano.

Also, AMD already had a Pentium 4 called K9 => Implementation of K9 Trace Cache blocks (not tapeout out) // Chief Architect of K9 designed to be a 5 GHz successor to the 3 GHz Opteron family.
It was replaced by Agena, the big oops of the 2000s.

Jaguar >1.85 GHz
Bulldozer >3.5 GHz
Is a lot different than K9 which was >5 GHz.

Top-dog positon = K8(>3 GHz) to K9(>5 GHz) <== Netburst pathing ;; Replaced with Stars(no longer Kryptonite#) cores (Greyhound, Greyhound+, Husky)
Mid-dog positon = K8L(>2 GHz) to K10<<Bulldozer>>(>3.5 GHz) <== Pentium M pathing ;; Pushed to eventually replace the above.
Bottom-dog positon => Geode/K7(< 1.8 GHz) to Bobcat (<1.7 GHz) <== Atom pathing ;; Continued through evolving into Zen.

 

[Cranky5150]

I believe i still have a 8350 and a 970 board laying around in my closet LOL. I kinda still have fond memories of that chip TBH....

 

[JB_Gamer]

Ha-ha, hi-hi, Der8auer did that ("the CPU remained attached to the cooler, which resulted in some bent pins (screams in horror)"), then l'm in good company. Yes I remember that feeling..., and I also learned the lesson - twist before lift it.

 

[Cranky5150]

I think we have all learned that lesson TBH....LOL

Ha-ha, hi-hi, Der8auer did that ("the CPU remained attached to the cooler, which resulted in some bent pins (screams in horror)"), then l'm in good company. Yes I remember that feeling..., and I also learned the lesson - twist before lift it.

 

[lexluthermiester]

Anything above 8GHZ is impressive! Extreme voltage though. Even at those low temps, the circuit pathways will degrade quickly.

 

[R-T-B]

I want a 14nm fx cpu
I wonder how much Voltage can 14 nm take

Less.

Anything above 8GHZ is impressive! Extreme voltage though. Even at those low temps, the circuit pathways will degrade quickly.

As will the ln2... No one does this for more than the bench.

 

[Fluffmeister]

I wondered why all my lights went dim recently, something was sucking the power grid dry....

 

[Jism]

Cool, new CPU's are boring, as they are maxed out already. AM3 / Vishera is such a overclockers dream platform, anyone not tried should dive into it.

I'm sure there was more headroom then just the CPU clocks; as the CPU/NB is indirect responsible for L3 cache speeds. Had one for year, 8350 running at 4.8Ghz 24/7 with a 300Mhz FSB. If derbauder did'nt focus on just MP oc'ing alone, he would have cranked up that score even higher. My best CB 15 was around 761 points, which for a 4.8Ghz CPU is'nt bad.

 

[Bones]

Did almost the same myself yesterday and I believe I was using less voltage too BUT he had that I could see all 8 cores going, I only had 1 core going but still - Good day for benching FX.
AMD FX @ 8121.79 MHz - CPU-Z VALIDATOR

 

[Jism]

If anybody could lend me LN2 perhaps i could set a WR lol. I know the timings and settings blindly in my head, getting a working 340Mhz FSB, 2600Mhz DDR3 and a CPU/NB speed of over 2800Mhz or so.

 

[Bones]

Anything above 8GHZ is impressive! Extreme voltage though. Even at those low temps, the circuit pathways will degrade quickly.

This is true, LN2 is rough on a chip due to thermal expansion/contraction of the silicon at such low temps plus the voltage required.
My chip is damaged anyway with a bad IMC making it run in single channel so in my case it doesn't matter alot, I just popped it in and turned it loose.

Mine could probrably (I stress probrably here) do the same or close with all 8 going but I'd have to volt it to potentially lethal levels and then hope it can make it. Der8auer is a great bencher, you can't take that away from him at all.

I'll be getting more LN2 tomorrow, not for this but other things I want to do BUT if I have any left over I'll give it a shot.

 

[ARF]

Slow news day?

lol, 8Ghz on a potato from 2011...Who cares...!

This was done shortly after release and isn't even the fastest! Why is this news?

FTFY

Nostalgia about lost times when people saw GHz and went aw wow, but the reality is that CPUs can't go any higher than 5 GHz and probably will never do.

Intel predicts 10GHz chips by 2011

Intel predicts 10GHz chips by 2011 - Geek.com

intel is predicting that its microprocessors will hit 10ghz by the year 2011. in addition, it is currently working on a system bus that is 10 times faster than its upcoming 400 mhz …

www.geek.com

 

[Jism]

Oh they (proberly) can, but at a huge expensive of heat / power, and not suited for at home use with a stock cooler or motherboard.

 

[seronx]

Nostalgia about lost times when people saw GHz and went aw wow, but the reality is that CPUs can't go any higher than 5 GHz and probably will never do.

Actually, Intel would had no issues with Tejas, after Cedar Mill.

If you line things up, yep that is 50 GHz after 2010, not 10 GHz. With 2013+ being 100 GHz.

https://www.researchgate.net/publication/224648567_An_88GHz_198mW_16x64b_1R1W_variationtolerant_register_file_in_65nm_CMOS

Steven Hsu, Amit Agarwal, Mark Anders, Sanu Mathew, Ram Krishnamurthy, Shekhar Borkar
Intel, Hillsboro, OR

https://www.researchgate.net/publication/224648455_A_9GHz_65nm_Intel_Pentium_4_Processor_Integer_Execution_Core

https://www.researchgate.net/publication/2983363_A_9-GHz_65-nm_IntelR_Pentium_4_Processor_Integer_Execution_Unit

The design of a fourth generation Intel Pentium 4 64-bit integer execution core fabricated in a 65-nm CMOS technology, operating at frequency of 9 GHz, consuming 10.36 W at 1.3 V, 70 degrees C is described.


It also states the die size...
The 65-nm Intel Pentium 4 processor is optimized to enable both single core and dual core products. Each die has a 2-MB L2 cache and 188 million transistors. The dual core products are supported by using two die in the same package. The die size for the dual core implementation is 162 mm.



Technically, can tell it is Cedar Mill.

Just so, you all get it.
16 KB L0d, 12K uops trace cache (12K uops ~= >80 KB L0i, any cache after Decode is L0i)
with no buffer to a L2 cache the size of 2 MB.

Then, there was Power6... and other chips doing the >4 GHz push at least.

https://www.researchgate.net/publication/2983721_Design_and_Implementation_of_the_POWER6_Microprocessor

POWER6 wasn't considered good, but it hit 5+ GHz in the same time frame...

IBM thumbs nose at heat concerns, kicks Power6 to 6GHz

Taunts rivals with pure speed

www.theregister.co.uk

https://www-03.ibm.com/press/us/en/...ILE2&fileName=POWER 6 Fact Sheet - 052507.pdf

Also...
"4-GHz Tejas CPU Integration Engineer: Responsible for physical integration of a front end Sequencing partition using Cadence PDP. Also involved in Design Automation."
Tejas wasn't going to clock as high as people thought it was going to...

Compared to AMD's K9 it was going to be 1 GHz slower.

 

[ARF]

Oh they (proberly) can, but at a huge expensive of heat / power, and not suited for at home use with a stock cooler or motherboard.

There is one thing called degradation, too. Have you seen a dead CPU because of too much overclock?

Compared to AMD's K9 it was going to be 1 GHz slower.

With equal end performance which means high frequency doesn't translate necessarily to higher performance.

 

[basco]

wow time flies by :

Game City 2011 - wir sind wieder dabei!

Auch dieses Jahr werden wir dazu eingeladen, Österreichs größte Spielemesse von 22.-24. Oktober mit High-End-Hardware und flüssigen Stickstoff zu versüßen. Gemeinsam mit DiTech haben wir einige Projekte für euch vorbereitet und dürfen im Rahmen einer täglichen Overclocking-Show auf der...

www.overclockers.at

pictures tell more then words

 

[seronx]

With equal end performance which means high frequency doesn't translate necessarily to higher performance.

Pentium 4 => no buffer to one buffer 128-byte instruction buffer, one instruction(two uops(IA32) to four uops(IA32e(64-bit))) decoded per cycle max peak. Which the schedulers for the front-end of the execution core dispatch six uops.
Tejas actually targeting 4 GHz, no where near the >5 GHz or above.

AMD K9 => 5 GHz @ ~30 stages
'It was described by Alsup as: "K9 fetched 8 instructions every other cycle and made 2 branch predictions associated with 3 next fetch addresses every other cycle. K9 issued 4 instructions per cycle and took 2 cycles to issue a fetch width."'
"According to Alsup, it was designed to be close to 95% of original K8 IPC but reach 5GHz frequency in a 35 nm process. At the time of cancellation most of the logic was running in SPICE at 5GHz and majority of the layout was done."
"Our analysis of the device has confirmed, that the 65nm AMD Athlon 64 X2 dual-core processor, produced at AMD Fab 36, uses minimum gate lengths of 35nm..." <== related to above, not an Alsup quote by wikichips/groups.

Both cases Intel and AMD stopped at 65nm rather than going to the next node, Intel's 45nm introduced HKMG and AMD's 45nm just in general being better than 65-nm.

Pentium 4 small front-end, small retirement, with big w/ double clocking execution core.
K9 big front-end, big retire, not double clocked execution core.

Bulldozer continued big front-end, big retire, but two smaller execution cores. In the original patent for CMT, there was behavior relative to Bulldozer that described if logical core1, core3, core5, core7 were parked. Then core0, core2, core4, core6 would have been able extend across and use the physical second core. Thus, giving the capability of eight small logical cores or four big logical cores. With that Bulldozer dropped a lot of advancements that it should have gotten. The behavior of CMT shouldn't be any different from SMT. If a quad-core i7 can have two cores with two active&two parked threads(two big logical cores), and two cores with four active threads(four little logical cores). Then, an ideal FX can also have two modules with two active&two parked threads(two big logical cores), two modules with four active threads(four little logical cores). However, the performance should be different, with latter cores running smaller threads in the i7 would be competing for execution slots. While, the FX modules providing each thread a separate execution resource, thus different slots.

 

[de.das.dude]

Proud fx owner of an 8320 i can still run 4.5GHz stable on all cores... doesnt make much sense though. Feet gets warm :lol: