AMD Kaveri OC On Planet Neptune

[PPB]

Guys don't turn this into compiler discussion. Let's discuss Kaveri and the results...

You can't discuss a bench result if you dont first lay the ground about what work that bench is doing and how it does it.

I can roll the charts showing how far from reality CB can be from real renderers. This affects both Intel and AMD equally.

I know these benches are the easiest thing to do when you want to leak information regarding a soon-to-be-released product. But as easy they are to run, they wont be representative of real world scenarios. I personally dont use my desktop to run SuperPi, CB or 3dmark 24/7 (heck, not even for stability stress testing porpuses I do).

 

[inf64]

Well this is all we got. It's hard to discuss other "benchmarks" when you have no data on them, right?

 

[krumme]

Who cares. What matters is can it run bf4 and how well - and the likes - and what is the battery life in notebooks. That and cost is what matters for the market penetration.

Gf shitty process might because of the wsa be a big gain for marketshare. Loads and loads of kaveri and huge amd gpu is going to hit the market. Lol. Its like bt for the tablet market.

 

[PPB]

Yep, thats why I would just wait till more information is leaked.


We had some SPEC benches too, but somehow we arent discussing them at all.

 

[Sweepr]

88poinst at 4Ghz ST?

A10-5800k - 297/86/3.46

Hardly any ST improvement for SR. That IGP part better be fantastic, or its a chip going nowhere.

So basically zero performance improvement for CB single-thread vs PD @ same clock? Dat predictions. After all the talk about IPC improvement and reduced CMT penalty (30% higher ops/cycle throughput) I was expecting 2M/4C Kaveri to whipe the floor with 2M/4C Richland and even come close to ~FX6300 MT performance. Lets hope this app doesnt reflect overall performance.

With 512 GCN cores @ 720MHz I wouldnt expect fantastic things from the IGP portion either, and thats the top model, the other one has 384 GCN cores @ 720MHz.

 

[ShintaiDK]

Well I suppose at least GPU compute will get a decent uplift with the new GPU,and hopefully Crossfire will actually work properly as there are 384 and 512 shader parts.

Forget hybrid crossfire

The GPU part is the last chance for kaveri to show any real benefit. The obvious question is, how limited will it be bandwidth wise. 2133Mhz got a 20-25$ premium and 2400Mhz is not even an option within specs.

 

[inf64]

So basically zero performance improvement for CB single-thread vs PD @ same clock? Dat predictions. After all the talk about IPC improvement and reduced CMT penalty (30% higher ops/cycle throughput) I was expecting 2M/4C Kaveri to whipe the floor with 2M/4C Richland and even come close to ~FX6300 MT performance. Lets hope this app doesnt reflect overall performance.

With 512 GCN cores @ 720MHz I wouldnt expect fantastic things from the IGP portion either, and thats the top model, the other one has 384 GCN cores @ 720MHz.

Don't tell me you have missed this? Not paying attention on the class again :$

@Shintai

You can forget 2400Mhz ram... 2133 is the "sweet" spot if you can call it that (in the light of mem. BW issue).

 

[ShintaiDK]

So basically zero performance improvement for CB single-thread vs PD @ same clock? Dat predictions. After all the talk about IPC improvement and reduced CMT penalty (30% higher ops/cycle throughput) I was expecting 2M/4C Kaveri to whipe the floor with 2M/4C Richland and even come close to ~FX6300 MT performance. Lets hope this app doesnt reflect overall performance.

With 512 GCN cores @ 720MHz I wouldnt expect fantastic things from the IGP portion either, and thats the top model, the other one has 384 GCN cores @ 720MHz.

Its really starting to look grim. CPU clock down, the IGP is also down from the first expected 900Mhz to 720Mhz. And the 384SP parts will have a hard time against the higher clocked Richlands.

 

[inf64]

@ PPB
What SPEC benchmarks? Can you post a link?

 

[Homeles]

You can't discuss a bench result if you dont first lay the ground about what work that bench is doing and how it does it.

I know these benches are the easiest thing to do when you want to leak information regarding a soon-to-be-released product. But as easy they are to run, they wont be representative of real world scenarios. I personally dont use my desktop to run SuperPi, CB or 3dmark 24/7 (heck, not even for stability stress testing porpuses I do).

Real render software use ICC as well, FYI.

I can roll the charts showing how far from reality CB can be from real renderers. This affects both Intel and AMD equally.

So... if it affects them equally, does that not make it a viable benchmark for cross-generation and cross-vendor comparison?

 

[inf64]

I think C15 is perfectly valid for cross-gen comparisons. It might not be the *best* workload to showcase the cross-gen improvements but this doesn't invalidate the benchmark itself. Sure there might be better and more representative workloads out there, especially the (linux) ones compiled with bdver3 flags.

 

[PPB]

Real render software use ICC as well, FYI.

As you like to say, the burden of proof is on you, prove it by showing us real world renderers that implement ICC.

So... if it affects them equally, does that not make it a viable benchmark for cross-generation and cross-vendor comparison?

Nope, as my statement was directed as they both are benefitted from using real world renderers. AMD gets into a better light as like I said for the 90th time, real world renderers have a lot of INT code in it. The BD architecture and it's succesors built on it are particulary weak on FP workloads, even compared to K10, a debatable design decision taken by the big core uArch team. I would rather just start trimming FP resources once the HSA aspect of the design has taken off, not in the very first itineration of the BD design.

But then again, AMD expected to have HSA running in the very first APU.

 

[Vesku]

The compiler stuff is a red herring, imo, at least for Cinebench R15. My 1090T falls right around where I'd expect compared to 4C/8T Nehalem and Ivy Bridge. Posted my results in the R15 thread: http://forums.anandtech.com/showpost.php?p=35896196&postcount=96

I don't think R15 will turn out to be a particular bad bench for Kaveri but as inf64 said there are bound to be much "better" benches for highlighting Steamroller over Piledriver. I'm sure AMD won't neglect to let us know which benchmarks those are. ;p

 

[Homeles]

As you like to say, the burden of proof is on you, prove it by showing us real world renderers that implement ICC.

Actually, since you're the one who made the original claim (that Cinebench's use of ICC somehow invalidates its usefulness as a benchmark), it's on you.

Also, consider that you're the only one in this thread complaining about Cinebench.

 

[JDG1980]

Is AMD still committed to moving all production to Global Foundries? If so management must be losing sleep over being stuck on GF 28nm through most if not all of 2015.

Global Foundries obviously can't deliver. Why on earth would AMD want to move any more production over there than is required to avoid penalties under the WSA?

 

[AtenRa]

Stock 6800K gets 326points.

Now tell me how much you got in singlethreaded. And lets see how much it misses from the fabled ST performance benefits that was told to be. The 30% IPC is obviously long gone.

We measure IPC clock to clock, not product to product

What is the score of PD at 3.7GHz gets in R15 single core ???

 

[Homeles]

We measure IPC clock to clock, not product to product

What is the score of PD at 3.7GHz gets in C15 single core ???

It's really nasty habit of people on his forum to invalidate a single piece of someone's argument, while failing to address the main point.

You're avoiding the point he's making. There's a regression in performance here, when there really shouldn't be. Steamroller is supposed to improve on Piledriver, and here we're seeing the opposite.

Now I'd be inclined to say that it's a combination of two things: Cinebench not being representative of the average per-clock improvement that Steamroller brings, and then the obvious regression in transistor performance.

The first is understandable and forgivable. The second is not, unless that regression is accompanied by a corresponding decrease in cost for consumers. 28nm bulk should be cheaper than 32nm PD-SOI. If those cost savings are passed on to the consumers, and Steamroller's performance doesn't improve substantially over Piledriver, it's the consumers that lose here.

It's way to early to be making that call, though. Cinebench is a single data point, and we should all be saying "hmm, that's interesting," and moving on. There's no point in drawing conclusions from this.

 

[ShintaiDK]

We measure IPC clock to clock, not product to product

What is the score of PD at 3.7GHz gets in R15 single core ???

ST IPC is close to stagnant with a few % tops. You missed that the Kaveri ran at 4Ghz for singlecore. But again, all the numbers have been posted. There seems to be a few percentage improvement for MT. But its not even remotely close to the usual stupid hyperbole hopes made up with the 30% claim froma slide people refused to understand the context of. Kaveris only hope now is the IGP part.

Doesnt look good for all the people that "could upgrade to Kaveri".

 

[AtenRa]

It's really nasty habit of people on his forum to invalidate a single piece of someone's argument, while failing to address the main point.

You're avoiding the point he's making. There's a regression in performance here, when there really shouldn't be. Steamroller is supposed to improve on Piledriver, and here we're seeing the opposite.

Now I'd be inclined to say that it's a combination of two things: Cinebench not being representative of the average per-clock improvement that Steamroller brings, and then the obvious regression in transistor performance.

The first is understandable and forgivable. The second is not, unless that regression is accompanied by a corresponding decrease in cost for consumers. 28nm bulk should be cheaper than 32nm PD-SOI. If those cost savings are passed on to the consumers, and Steamroller's performance doesn't improve substantially over Piledriver, it's the consumers that lose here.

It's way to early to be making that call, though. Cinebench is a single data point, and we should all be saying "hmm, that's interesting," and moving on. There's no point in drawing conclusions from this.

A10-7850K gets almost the same performance(within margin of error) with 200MHz lower clocks vs A10-5800K and 95W TDP vs 100W TDP and you call it regression

Not to mention its only a single benchmark.

 

[Homeles]

A10-7850K gets almost the same performance(within margin of error) with 200MHz lower clocks vs A10-5800K and 95W TDP vs 100W TDP and you call it regression

Not to mention its only a single benchmark.

It's really nasty habit of people on his forum to invalidate a single piece of someone's argument, while failing to address the main point.

Bolded for you, since you seem to have trouble reading.

 

[AtenRa]

Doesnt look good for all the people that "could upgrade to Kaveri".

I wouldn't expect anything less from you, you always use a single benchmark/game to promote your agenda.

 

[Homeles]

I wouldn't expect anything less from you, you always use a single benchmark/game to promote your agenda.

Pot, kettle, black.

 

[ShintaiDK]

Its clear to see what went wrong. CPU was to be 4Ghz baseclock, down to 3.7Ghz. GPU was to be 900Mhz, down to 720Mhz.

And with slides:

The question is, is the problem (yet again) coming from AMDs simulations. Or is it the 28nm process node.

 

[AtenRa]

It's really nasty habit of people on his forum to invalidate a single piece of someone's argument, while failing to address the main point.

Bolded for you, since you seem to have trouble reading.

Do you even know how to measure IPC ??? same clock for both CPUs. How Kaveri at 4GHz is in the same clock as A10-5800K clocked to 4.2GHz ???

He was speaking about IPC, not products. Learn the difference.

 

[PPB]

Actually, since you're the one who made the original claim (that Cinebench's use of ICC somehow invalidates its usefulness as a benchmark), it's on you.

Also, consider that you're the only one in this thread complaining about Cinebench.

You made that explicit claim, my claim is that CB doesnt represent real world renderers, which is true:

CB15:

LINK!

Vishera barely matches IB 4C8T with a 1000mhz advantage

Real world renderer:

LINK (AT somehow wont load PNG images)
Vishera at stock (4000mhz) renders the scene faster than IB 4C8T at stock (3500mhz).

Remember that when you come here with your proof, use respectable sources, cheers!