Kaveri performance

[AtenRa]

But surely you want CPI to go down, not up? (Unless you get a massive clock speed boost.)

Heh, you are correct i was thinking in IPC terms, CPI will go down.
It seems to me we will get more confusion like that, ill stick to IPC.

 

[galego]

I'll take 15% faster CPU and 30% faster GPU. For me the GPU is more important but it must be horribly bandwidth limited still on DDR3.

A 15-20% for the CPU and a 25-30% for the GPU seems feasible.

AMD's current marketing says that Steamroller is supposed to have 30% higher IPC than Piledriver. Obviously, this can't be taken at face value, but Steamroller is a major revision (not just a minor tweak like Bulldozer->Piledriver) and there is enough low-hanging fruit in the new architecture that 15%-20% increased IPC in most applications seems plausible.

AMD claims 30% over Bulldozer, which implies 20% over Piledriver.

 

[galego]

No, generally the 8100 is about 10-20% faster in fully MT situations.

I have benchmarks that say the contrary.

 

[Enigmoid]

I have benchmarks that say the contrary.

Sorry, I misread your post. I thought the comparison was 6300 and 8150 (I cannot find 8100 on newegg).

Key word is fully Multithreaded.

And yes there are exceptions but generally the 8150 is about 10-20% faster than the 6300.

 

[galego]

Sorry, I misread your post. I thought the comparison was 6300 and 8150 (I cannot find 8100 on newegg).

Key word is fully Multithreaded.

And yes there are exceptions but generally the 8150 is about 10-20% faster than the 6300.

Agree on 8150 > 6300, but I was comparing the 8100 vs the 6300.

The 6300 > 8100.

 

[PPB]

I think this thread needs to be locked too, as it probably was a flamebait too made by a troll 🙄

BTW, where is Seronx and his claims of 2m4c SR = 4m8c piledriver?


It was created by galego, who is now banned, but many here are being constructive. Not you however. Please don't follow in his footsteps.
Markfw900
Anandtech Moderator.

 

[NTMBK]

BTW, where is Seronx and his claims of 2m4c SR = 4m8c piledriver?

Shhh, if you say his name too many times then he will appear 😛

 

[formulav8]

Kavari will now have direct memory addressing for the gpu correct (like Intel did awhile back on its igps)? It won't have to do a check back or write back or something which needs a lot of bandwidth? I can't remember the name for it. But anyways, that should definitely provide a strong boost in gpu performance and not rely so much on memory banwidth as before iirc.

 

[PPB]

Hopefully that's the case and in the future AMD would be able to tap in more shaders into the iGP without coming into serious bottlenecks from RAM BW.

 

[NostaSeronx]

16 64-bit ALUs / 16 64-bit AGUs / 16 64-bit fpALUs / 8 128-bit ALUs / 8 threads = 16 64-bit ALUs / 16 64-bit AGUs / 16 64-bit fpALUs / 4 256-bit ALUs / 4 threads

Amiright? Then you have TSX and hUMA, oh my.

 

[LegSWAT]

Hopefully that's the case and in the future AMD would be able to tap in more shaders into the iGP without coming into serious bottlenecks from RAM BW.

DDR4 should do

 

[Ajay]

Everything is going to be heavily dependent on the actual shipping clock speeds. If GFL's 28nm isn't mature enough, even a large increase in IPC isn't going to matter much. At this point I'm just waiting to see what actually ships - to many moving parts to even guess at performance, IMHO.

 

[Blitzvogel]

A 15-20% for the CPU and a 25-30% for the GPU seems feasible.



AMD claims 30% over Bulldozer, which implies 20% over Piledriver.

Top end Kaveri needs to be hitting 1 TFLOP of GFLOP performance, and competitive with laptop i5s if AMD wants to be any kind of relative in the mainstream market. A 30% boost in CPU performance would get them this and doing it for cheaper than what Intel is offering by a good amount. As for not starving the graphics array, AMD might be implementing a more robust cache system, perhaps a giant amount of L3 (more than 16 MB?) or more on-GPU cache memory. That's the only solution beyond GDDR5 or DDR4 I can think of.

 

[strata8]

I suspect we'll see 600 GFLOPS on the top end mobile part (eq. 7730) and 900 on the desktop (eq. 7750). 8 CUs (512 shader cores) @ 550 MHz for mobile and 850 MHz for desktop doesn't sound too far fetched.

 

[monstercameron]

I suspect we'll see 600 TFLOPS on the top end mobile part (eq. 7730) and 900 on the desktop (eq. 7750). 8 CUs (512 shader cores) @ 550 MHz for mobile and 850 MHz for desktop doesn't sound too far fetched.

they said around 900 Gflops on their flagship apu...

 

[Erenhardt]

Top end Kaveri needs to be hitting 1 TFLOP of GFLOP performance,

1 TFLOP of GFLOP what? What kind of riddle is this?!
...
I know the answer:
TFLOP - Total flop
GFLOP - Global flop

So Kaveri needs to be Total flop of Global flopping! But I don't see how that may help AMD sell those chips.
Ohh.. and that would mean bigger failure than bulldozer
😉

No trolling or thread crapping. This is your third offense. Next one is a mandatory vacation
Markfw900
Anandtech Moderator.

 

[inf64]

In one of the slides from last year Kaveri was listed as 1049 Gflops APU... That's CPU+GPU combined flops.

 

[raghu78]

yeah kaveri was estimated to have total compute performance of 1050 GFLOPS (CPU + GPU) . see client product overview pdf (slide 10 and 30)

http://ir.amd.com/phoenix.zhtml?c=74093&p=irol-2012analystday

but given AMD's problems and delays with Kaveri, the number could be slightly below 1 TFLOPS (1000 gflops). Kaveri might be affected by Globalfoundries 28nm High performance process which seems to be having serious issues. the only 28nm process ramping at GF currently is 28nm SLP suitable for ARM chips.

 

[inf64]

Might be could be should be...

 

[Blitzvogel]

Any word on the cache system?

 

[inf64]

Any word on the cache system?

It has been detailed at last year's HC and AT wrote about it in their SR article 😉.

 

[NostaSeronx]

Any word on the cache system?

2T eDRAM or 6T SRAM, choose. 6T SRAM is the safest bet though. Other than the size and pipe width increases it should be exactly the same as Bulldozer.

Kaveri might be affected by Globalfoundries 28nm High performance process which seems to be having serious issues. the only 28nm process ramping at GF currently is 28nm SLP suitable for ARM chips.

TSMC is fabbing Kaveri, just so you guys know.

 

[NTMBK]

TSMC is fabbing Kaveri, just so you guys know.

That's not what AMD said.

 

[Bill Brasky]

edit: nevermind.

 

[NostaSeronx]

That's not what AMD said.

The GlobalFoundries and AMD thing has become increasingly hostile. While the TSMC and AMD thing is becoming increasingly friendlier.

Within four years you will see:
AMD focusing chips on GlobalFoundries to AMD focusing chips on TSMC.
Nvidia focusing chips on TSMC to Nvidia focusing chips on GlobalFoundries.

Project Denver(Parker) is on GlobalFoundries.
Kaveri/Carrizo is on TSMC.