AMD Carrizo: CPU Performance Plateau Inbound?

[patrickjp93]

http://wccftech.com/amd-carrizo-apu-isscc-2015-presentation-leaked-5-ipc-gain-x86-steamroller-die-consists-31-billion-transistors/

 

Videocardz has published the opening benchmark and performance numbers for AMD's Carrizo APUs, and it doesn't look good for team red. AMD was touting up to 20% IPC increases over Kaveri due to HDL design of the CPU components in addition to the HDL GPU components, something most people, including tech enthusiasts, believed. Well, it turns out Carrizo is only producing 5% IPC increases in most games, both in iGPU mode and dGPU mode. Also, to my personal dismay, the iGPU is still GCN 1.1, so there are no dual graphics capabilities beyond what Kaveri allowed. So, where did the extra 0.7 billion transistors go to? Well, H.265 transcoding and HSA compliance for the most part, as well as AVX2, BMI2, MOVBE and RDRAND, leaving only 3 instructions on Haswell which AMD has not implemented.

 

I'm actually really sorry that AMD didn't deliver on much besides efficiency with this chip, but that said it is extremely efficient! Carrizo L can idle at just 1.5W! This means many great things for this little chip when it comes to tablets, but otherwise Carrizo and Carrizo L are lackluster, and that's being kind. Also, it should be noted clock rates really have not taken a hit, maybe 200 MHz in minimums, though maximums seem to have increased almost 20%. We'll see how it balances out in real world performance if the top clocks are sustainable for long periods or not.

 

Don't shoot the messenger!

 

Personally I wonder just how Intel is meant to squeeze 20% more over Haswell from Skylake... Eventually software devs are just going to have to get better at coding and use better compilers. 

 

How the Hell could AMD be lauding HDL for the CPU and not have it deliver any sort of performance gain? It looks like all the work was done on the cache side.

[Ren]

Don't shoot the messenger!

But you posted WCCFTech as source and then said the data was from Videocardz..

[TheSLSAMG]

But you posted WCCFTech as source and then said the data was from Videocardz..

Read the article. Videocardz posted the data, the article was written on the WCCF site.

[patrickjp93]

But you posted WCCFTech as source and then said the data was from Videocardz..

The WCCF article is better written and provides a good summary.

[niofalpha]

*Plays Violin*

[flibberdipper]

Makes me sad. But I half expected it.

[Ren]

Read the article. Videocardz posted the data, the article was written on the Videocardz site.

That's never going to happen if the link says "WCCFTech".

[TheSLSAMG]

That's never going to happen if the link says "WCCFTech".

Okay, now you're just being petty.

[patrickjp93]

That's never going to happen if the link says "WCCFTech".

 

Okay, now you're just being petty.

^ I would also add irrational, or if he's just pointing that out, it really goes to show how irrational the LTT community is when it comes to news. I don't care who prints it as long as the methodology to get it is sound.

[azurezeed]

we need to wait for Zen

[ComradeHX]

That's never going to happen if the link says "WCCFTech".

WCCFTech have been mostly free of bullshit for a while.

Are you from the past?

[patrickjp93]

WCCFTech have been mostly free of bullshit for a while.

Are you from the past?

I believe he's making a run at the LTT community.

[Ertman]

If true, not surprising, if not, then we wait and see.

[SpaghettiCarbonara]

That's never going to happen if the link says "WCCFTech".

 

Isn't it better to focus on the writers themselves? Believe it or not, WCCF isn't just one writer (as far as I know, most aren't). Find which writers spew the bullshit and ignore them, not the entire website.

[Bensemus]

Everyone keeps hating of WCCFT since I joined the forum and I have yet to see them give us some grade A bull which is what I expected given the hate. Websites can change and it seems like WCCFT is doing that.

[Hatsune Miku 「 」]

I could almost bet that the Samsung octa core big.little arm processor would be faster.

[ComradeHX]

I expected nothing.

[EmmaLong]

Not a chance this is a plateau. First off Intel promised 5% IPC gains with Ivy Bridge and they failed to deliver even that.

Excavator, according to the official presentation slides at least, is 29% smaller than steamroller, achieves 20% higher clock speeds at the same voltage (10% of that is actually thanks to the HDL @patrickjp93 ) and a 5% IPC gain on top of that.
That's actually a 26% (1.2 x 1.05) CPU performance gain over Kaveri at the same power.This is nothing short of fantastic. And that's without even including the GPU improvements in the discussion.

 

I could almost bet that the Samsung octa core big.little arm processor would be faster.

Not even close. That's not even an actual 8 core chip. It has 4 high performance cores and 4 low power state, tiny cores.

[patrickjp93]

I could almost bet that the Samsung octa core big.little arm processor would be faster.

Let's not push it... ARM is not going to catch x86 until the turn of the decade, if it ever does. ARM has slammed into a brick wall on frequency, and its branch predictor needs serous work. There are plenty of things ARM is lacking in which will keep AMD's x86 chips safe in the mobile world for a while yet.

[Hatsune Miku 「 」]

Not even close. That's not even an actual 8 core chip. It has 4 high performance cores and 4 low power state, tiny cores.

Why else did I say "big.little"

[patrickjp93]

Not a chance this is a plateau. First off Intel promised 5% IPC gains with Ivy Bridge and they failed to deliver even that.

Excavator, according to the official presentation slides at least, is 29% smaller than steamroller, achieves 20% higher clock speeds at the same voltage (10% of that is actually thanks to the HDL @patrickjp93 ) and a 5% IPC gain on top of that.

That's actually a 26% (1.2 x 1.05) CPU performance gain over Kaveri at the same power.This is nothing short of fantastic. And that's without even including the GPU improvements in the discussion.

 

Not even close. That's not even an actual 8 core chip. It has 4 high performance cores and 4 low power state, tiny cores.

Sorry, it's a bit late for me. I mean it seems both companies are plateauing, even if at different places. It's a bit depressing, though as I've said many times, the problems mainly lie with the software developers.

 

Also, be aware the overclocking headroom on Carrizo has been cut clean in half. You do pay a price for using HDL, and that is in clock rate ceiling.

[Opcode]

Performance plateau?

 

i3-2130 65w @ 3.4 GHz = 6.96 GFLOPS

Carizzo 15w @ 3.4 GHz = 6.54 GFLOPS

 

The jump in performance per watt for Carrizo is insane (50%).

[ComradeHX]

Performance plateau?

 

i3-2130 65w @ 3.4 GHz = 6.96 GFLOPS

Carizzo 35w @ 3.4 GHz = 6.54 GFLOPS

 

The jump in performance per watt for Carrizo is insane (50%).

i3-2130 -> 32nm

Carizzo -> 28nm

And where did you get your "gflops"?

[patrickjp93]

Performance plateau?

 

i3-2130 65w @ 3.4 GHz = 6.96 GFLOPS

Carizzo 35w @ 3.4 GHz = 6.54 GFLOPS

 

The jump in performance per watt for Carrizo is insane (50%).

perf/watt. Also, come on toss in Ivy Bridge if you're gonna do that. Shame on you comparing 28nm HDL against 32nm bulk when we both know 22nm FF is a much better match for density.

[Faa]

Not a chance this is a plateau. First off Intel promised 5% IPC gains with Ivy Bridge and they failed to deliver even that.

 

Vishera was supposed to add 10-15% IPC, Kaveri was supposed to add 10%,

None of them succeeded hitting Phenom's IPC yet. Also (103-97)/97 = 6% gain so yes they managed the 5% IPC they promised.