AMD CPU speculation... and expert conjecture

[gamerk316]

First benchmark that I gave you: FX-8350 @4GHz is up to 27% faster than FX-6350 @3.9GHz. The difference in clocks is only 3%. The rest is from "moar cores".

You assume for every 1% increase in clock you get 1% increase in performance. Based on the load figures, its clear you wouldn't expect a FX-6300 to do significantly worse then a FX-8350 at the same clocks.

Again from first benchmark: i7 IB 12 threads @ 4.3GHz is up to 12% faster than i7 SB 8 threads @ 4.6GHz. With lower clocks the 12 thread i7 win because "moar cores".

Second benchmark: i7 SB 12 threads @3.5GHz is up to 11% faster than i7 HW 8 threads @ 3.5GHz.

Now now, the 4960x has a lot more HW added on then just the extra two cores that could be affecting the results. Hence why I take those results with a grain of salt. Something like this is likely more representative:

http://www.techspot.com/review/827-watch-dogs-benchmarks/page5.html

Spoiler

Expected results. On the AMD side, you see virtually no change from the 6350 to the 8350, which is explainable by the 100MHz clock speed difference, For intel, its impossible to say due to the 82FPS GPU bottleneck.

Techspot is GPU bottlenecked. Toms result looks as those that I gave you: FX-8 is 35--36% faster than FX-6; the 12-threads i7 is 37--38% faster than i5. Again this shows scaling above 6-threads.

For intel, yes. Not for AMD:

The 8350 shows essentially linear scaling with clocks, and the fact it can reach the 82 FPS GPU bottleneck with an OC means the 79 FPS result at stock likely isn't being suppressed by the GPU.

 

[gamerk316]

Interesting move by Intel:

http://techreport.com/news/26656/xeon-hybrid-marries-x86-cores-with-fpga-sidekick

So they essentially slapped a FGPA onto a Xeon, and connected it via QPI. Not pretty, but should work, providing software programmed for the FGPA.

And yes Juan, this approach does satisfy HSA, as the FGPA exists in the same memory context as the CPU.

 

[8350rocks]

@juanrga:

Please check the PCPerspective write up of the conversation with AMD's Richard Huddy regarding new information about MANTLE/Linux and open sourcing the API. That was posted this morning at 9 AM.

 

[blackkstar]

It's a shame the people who don't understand what software betas are, APIs, or what API implementations are won't watch the Huddy interview or read a transcript.

 

[Cazalan]

Interesting move by Intel:

http://techreport.com/news/26656/xeon-hybrid-marries-x86-cores-with-fpga-sidekick

So they essentially slapped a FGPA onto a Xeon, and connected it via QPI. Not pretty, but should work, providing software programmed for the FGPA.

And yes Juan, this approach does satisfy HSA, as the FGPA exists in the same memory context as the CPU.

They've actually been doing x86+FPGA for a while, just with the smaller Atom parts.

The timing is interesting as Microsoft was talking about doing something very similar (except on PCIe) and also with Altera parts. Microsoft went into more detail so it's worth a read. The big winner here is Altera who has been the underdog of Xilinx for a while.

http://www.theregister.co.uk/2014/06/16/microsoft_catapult_fpgas/?page=1

I would object to the "not pretty" part. I'd say 95% of my projects for the last decade have involved one or more CPUs coupled to one or more FPGAs. Having fewer parts to lay down and route on a PCB is usually considered prettier. Unless you're a masochist and like routing 28 layer PCBs.

 

[etayorius]

AMD Gameworks like, but Open:

http://www.chw.net/2014/06/amd-openfx-la-alternativa-abierta-a-gameworks/

 

[juanrga]

You assume for every 1% increase in clock you get 1% increase in performance. Based on the load figures, its clear you wouldn't expect a FX-6300 to do significantly worse then a FX-8350 at the same clocks.

I am not assuming linearity. In any case, the measured 27% increase in performance cannot be explained by the 3% increase in clock. The explanation is "moar cores".

Interesting move by Intel:

http://techreport.com/news/26656/xeon-hybrid-marries-x8...

So they essentially slapped a FGPA onto a Xeon, and connected it via QPI. Not pretty, but should work, providing software programmed for the FGPA.

And yes Juan, this approach does satisfy HSA, as the FGPA exists in the same memory context as the CPU.

This Intel Xeon-FPGA hybrid is not HSA compliant. A single memory pool is only one of the requirements to be HSA complaint.

I mentioned two of the defects of this Xeon-FPGA hybrid when I gave the news about it:

http://www.tomshardware.co.uk/forum/352312-28-steamroller-speculation-expert-conjecture/page-277#13537454

@juanrga:

Please check the PCPerspective write up of the conversation with AMD's Richard Huddy regarding new information about MANTLE/Linux and open sourcing the API. That was posted this morning at 9 AM.

Disapointing! In the previous PCW interview Huddy promised AMD would devote resources to porting MANTLE to linux. In the PCPer interview he has changed the words to:

Huddy mentioned that AMD was considering, but not promising, bringing the Mantle API to Linux. If the opportunity exists, says Huddy, to give the gamer a better experience on that platform with the help of Mantle, and developers ask for the support for AMD, then AMD will at the very least "listen to that."

 

[8350rocks]

@juanrga: You need to read between the lines...it is what he is not saying that matters most there...

EDIT: AMD do not have the resources to facilitate the port entirely themselves...though they will make the API entirely open source, allowing anyone to openly integrate MANTLE into their OS. (like Valve or Canonical for example)

 

[anxiousinfusion]

Also, here is a tidbit...Carrizo is the mobile codename. The DT codename is different, and as of yet unreleased...

This will only apply to mobile variants — drop a Carrizo APU into the FM2+ socket, and the onboard FCH will disable itself while the motherboard FCH takes over.

From: Extremetech. If Carrizo is only the mobile variant, why would there be discussion of FMx compatibility? Maybe those sites have old news but I'm skeptical of your claims.

Funny enough, AMD's Papermaster has just given a talk about products/plans for year 2020

http://www.marketwatch.com/story/amd-accelerates-energy-efficiency-of-apus-details-plans-to-deliver-25x-efficiency-gains-by-2020-2014-06-19

Take 95W Kaveri GFLOP/s

Spoiler

multiply it by the number given by Papermaster and you get an APU with 21.4 TFLOP/s. But have in mind that this is single precision and I gave double precision numbers when I was discussing supercomputers:

21.4 TFLOP/s (SP) ~ 10.7 TFLOP/s (DP)

Your math doesn't take into consideration that 25x greater efficiency can be achieved if performance stays nearly the same and power consumption is cut greatly. I guess you can say up to ~10 Teraflops double precision but I doubt AMD would spend all of these efficiency gains entirely on performance.

 

[esrever]

FX9590 performance in under 10W should be easily doable with 2 process shrink. 16 puma+ cores at 2.5ghz will match the performance of the FX9590 while using 1/25th the power in 2020.

 

[juanrga]

@juanrga: You need to read between the lines...it is what he is not saying that matters most there...

EDIT: AMD do not have the resources to facilitate the port entirely themselves...though they will make the API entirely open source, allowing anyone to openly integrate MANTLE into their OS. (like Valve or Canonical for example)

Hum! I have been discussing if AMD ports or not MANTLE to linux. I have been discussing that in a former interview Huddy promised AMD would devote resources to port MANTLE to linux and in last interview he retracts.

Leaving others to do the work has little merit.

Unlike Intel/Nvidia, AMD doesn't seem interested in SteamOS really. I can smell why.

 

[juanrga]

Funny enough, AMD's Papermaster has just given a talk about products/plans for year 2020

http://www.marketwatch.com/story/amd-accelerates-energy-efficiency-of-apus-details-plans-to-deliver-25x-efficiency-gains-by-2020-2014-06-19

Take 95W Kaveri GFLOP/s

Spoiler

multiply it by the number given by Papermaster and you get an APU with 21.4 TFLOP/s. But have in mind that this is single precision and I gave double precision numbers when I was discussing supercomputers:

21.4 TFLOP/s (SP) ~ 10.7 TFLOP/s (DP)

Your math doesn't take into consideration that 25x greater efficiency can be achieved if performance stays nearly the same and power consumption is cut greatly. I guess you can say up to ~10 Teraflops double precision but I doubt AMD would spend all of these efficiency gains entirely on performance.

You are entirely right. You can divide the power consumption by 25x and then obtain near 800 GFLOPS in a phone-class APU. However, in the part of my post that you quote I was only considering the supercomputer-class APUs. In the other part of my post I wrote:

Believe it or not, but those ultra-high performance APUs will be used on everything from low power phones to 20 MW supercomputers.

 

[choofieme]

Seriously, all expects that AMD might gone wild

 

[de5_Roy]

http://techreport.com/news/26668/gaming-evolved-app-adds-video-recording-challenges-shadowplay
the long-neglected vce gets used, finally. thank you amd for acknowledging the hardware block existed since trinity!
edit: we'll be seeing those elusive video editing/transcoding apps any day now!

fuldis: ofc vce gets used in other stuff like steady video, h.264 encoding on a few apps.

AMD teases new FX desktop chip
http://techreport.com/news/26672/amd-teases-new-fx-desktop-chip

 

[juanrga]

It is a 12-core 6.31GHz Excavator-based FX-10001 CPU on IBM FDSOI 22nm for AM4 socket

/sarcasm

 

[Cazalan]

Maybe that petition worked after all.

 

[etayorius]

I know it`s a Water Cooled Piledriver, some roadmap a while ago was very clear about AMD extending Piledriver through 2015... so it will probably me more Vishera crap clocked higher.

 

[szatkus]

I know it`s a Water Cooled Piledriver, some roadmap a while ago was very clear about AMD extending Piledriver through 2015... so it will probably me more Vishera crap clocked higher.

Agree, I doubt that it's something new. There would be some leaks earlier.

 

[-Fran-]

Since the current FX9xxx doesn't come with a cooling solution, could it be that is just a "special FX9xxx"?

Cheers!

 

[szatkus]

BTW, these creatures in the background look like... trolls

 

[esrever]

I don't know why AMD even bothers, clocking piledriver even higher isn't going to compete with devil's canyon. I don't even know how they would even clock it even higher.

 

[jdwii]

It is a 12-core 6.31GHz Excavator-based FX-10001 CPU on IBM FDSOI 22nm for AM4 socket

/sarcasm

I wish but no Amd would rather not make much money.
Edit i also love how you act like you don't want some high-end CPU from Amd?

 

[jdwii]

http://techreport.com/news/26668/gaming-evolved-app-adds-video-recording-challenges-shadowplay
the long-neglected vce gets used, finally. thank you amd for acknowledging the hardware block existed since trinity!
edit: we'll be seeing those elusive video editing/transcoding apps any day now!

fuldis: ofc vce gets used in other stuff like steady video, h.264 encoding on a few apps.

AMD teases new FX desktop chip
http://techreport.com/news/26672/amd-teases-new-fx-desktop-chip

Also darn as soon as i got my 8350 but i bet its still piledriver and its probably a 220watt CPU which i yawn to

 

[juanrga]

Maybe that petition worked after all.

LOL

It is a 12-core 6.31GHz Excavator-based FX-10001 CPU on IBM FDSOI 22nm for AM4 socket

/sarcasm

I wish but no Amd would rather not make much money.
Edit i also love how you act like you don't want some high-end CPU from Amd?

I want them competitive and well-prepared for the future. You want them out of business like the rest of the blue/green people here.

 

[colinp]

AMD teases new FX desktop chip
http://techreport.com/news/26672/amd-teases-new-fx-desktop-chip

I did wonder whether there would be something new coming when I saw that several manufacturers had displayed some new 9-series boards at one of the trade shows recently (I forget which). I mean, why release new boards for a dying platform?

Of course, this could just as easily be an FM2 chip, as the FX brand doesn't just mean "unlocked chip with L3 cach" any more.