APU13- AMD Kaveri details- 856Gflops, 3.7Ghz CPU,720Mhz GPU

[sniffin]

The reason AMD sells its APUs for so cheaply is precisely because of their performance/power characteristics. Do you honestly think AMD is a charity that "likes" selling 200mm^2+ silicon dies for 37-40% gross margin? NO!

You're comparing two products at completely different price segments. May as well compare Iris Pro with Hawaii, it's about as relevant as your comparison is considering it is closer in price to Hawaii than it is to Kaveri. Making the assertion that Iris Pro's successor somehow makes Intel a graphics leader is a ridiculous assertion and you know it.

I remember you making a thread about fanboyism on this forum a while ago. When you write statements like below, it's really hard to take what you're saying seriously. How about taking a more measured tone rather than a passive aggressive one designed to stir up a flame war.

AMD's much vaunted graphics lead has now largely evaporated and with the 2014 Intel lineup, AMD will be a clear #2 in graphics.

Not unexpected, but it's nice to see evidence that what many of us have been predicting for quite some time finally playing out.

 

[SPBHM]

https://www.youtube.com/watch?v=HjAM2zYNqko

kaveri vs i7-4770k+gt630

kaveri gpu is ~2x gt630 so that is well within the 720p gaming proficiency zone! and 1080p for older titles.

GT 630 is (most of the time) a renamed GT 430 (Fermi with 96 cores, only 4 ROPs) and 128bit DDR3, but there is a lot of variation in terms of memory speed, and even some 64bit version, so it's hard to know what exactly was running on that video.

so, comparing it to a GT 630 (and not a 640) is not a good sign...
but fortunately there is something quite similar to the Kaveri specs, 512GCN cores at 800MHz and 1600MHz memory perform like this
http://www.xbitlabs.com/articles/cpu/display/amd-richland-intel-haswell-gpu_5.html#sect0

but I'm a lot more curious (any reliable info?) about the CPU performance improvements, if it's significant it would be a shame if AMD didn't release a version with l3 cache and 4 modules soon.

 

[LogOver]

wat? Haswell @ 5+ GHz gets about 130 GFLOPS throughput..IF you have the RAM @ 2000+ (quad-core that is)

the Hyperthreading adds 10-15

You're making mistake comparing FLOPS from some specific application to theoretical FLOPS coming from AMD calculations. First of all the FLOPS in Intel Burn Test are DP (Double Precision) FLOPS (64-bit FP operations). Second - Intel Burn Test currently does not support FMA operations. Comparing apples to apples (theoretical to theoretical) one Haswell core can do 112 SP GFLOPS at 3.5 GHz -> 8 (32 bit units in 256 bit vector) * 2 (fma) * 2 (ops per cycle) * 3.5 (GHz) = 112 SP GFLOPS.

 

[ShintaiDK]

Its not really looking good for AMD. They seem yet again affraid of making proper IGP progress because it eats into their dGPU segment every time. And with the anemic CPU, it just looks like another CPU with very low sales like the previous revisions.

 

[AtenRa]

Its not really looking good for AMD. They seem yet again affraid of making proper IGP progress because it eats into their dGPU segment every time. And with the anemic CPU, it just looks like another CPU with very low sales like the previous revisions.

Do you have any performance data you will like to share with the rest of us ??

 

[Erenhardt]

so, comparing it to a GT 630 (and not a 640) is not a good sign...
but fortunately there is something quite similar to the Kaveri specs, 512GCN cores at 800MHz and 1600MHz memory perform like this
http://www.xbitlabs.com/articles/cpu/display/amd-richland-intel-haswell-gpu_5.html#sect0

Why? Isn't this like comparing 6770(5770) to 7770?
6770(5770) - 1360 GFLOPS : 15FPS
7770 - 1280 GFLOPS : 27FPS.

GFLOPS != GFLOPS

Given changes to memory system, comparing theoretical values is worthless effort.

BTW: I called 512SP long ago

 

[Abwx]

Lies and misinformation characterize almost every one of your posts, SiliconWars, at least with respect to sticking to the facts re: Intel.

Is that a "technical argument" .??..

Where is the relevance with this thread.?.

And that s once the thread started , but curiously
two page later :

Now we're resorting to personal attacks. I guess if you can't have a civilized discussion, just call the opposition "insane".

 

[Abwx]

I wouldn't call these die sizes fairly large. They're pretty conservative for what they are. The current prices are totally realistic for what you get. These are not meant to be super-enthusiast-tier chips. AMD has gone on record saying they could add more modules or CU's if they wanted, but don't since there's not a huge demand for it. Outside of gamers and benchmarkers and other enthusiasts, most people don't need or desire these things.

Size without mention of process cost is useless , it s not nlike GF is particularly expensive , actualy it must be dirt cheap since they got a 2 1/2 years cycle for 32nm and their 28nm will use exactly the same machinery, the only additionnal cost is some light RD but the
two nodes are essentialy the same process.

 

[CHADBOGA]

Now my memory may be faulty, but I am pretty sure a few months ago I expressed extreme skepticism that Kaveri's IGP would be faster than a ATI HD5770, and there were a few regular posters here who claimed that I was wrong and that Kaveri would easily beat a HD5770. :whiste:

 

[R0H1T]

Now my memory may be faulty, but I am pretty sure a few months ago I expressed extreme skepticism that Kaveri's IGP would be faster than a ATI HD5770, and there were a few regular posters here who claimed that I was wrong and that Kaveri would easily beat a HD5770. :whiste:

I'm sure you have definitive proof that the Kaveri IGP is slower than Radeon 5770 right

 

[monstercameron]

so no discussion about hsa?
what do you guys think about gpu-accell. java? if this gets a perf boost on amd parts what do you think will happen for their server business?

also can one compile/translate/convert javascript into opencl/hsail to run it on the gpu or is that kind of workload not suited for a gpu?

 

[Arkadrel]

I'm sure you have definitive proof that the Kaveri IGP is slower than Radeon 5770 right

a 5770 has 76 GB/s memory bandwidth.

Im looking forwards to Kaveri performance too, but I doubt it ll beat a 5770.

Even with DDR3-2133, thats only ~34,2 GB/s memory bandwidth.

Then theres the 800 x VLIW5 should be more than 512 x GCN, even with GCN better effeciency, that should be a small egde to the 5770.

 

[CHADBOGA]

I'm sure you have definitive proof that the Kaveri IGP is slower than Radeon 5770 right

It's called making a prediction.

Not so dissimilar to inf64 predicting that Bulldozer would be the fastest desktop processor money could buy, accept that I will be on the money, unlike him.

 

[R0H1T]

a 5770 has 76 GB/s memory bandwidth.

Im looking forwards to Kaveri performance too, but I doubt it ll beat a 5770.

Even with DDR3-2133, thats only ~34,2 GB/s memory bandwidth.

Then theres the 800 x VLIW5 should be more than 512 x GCN, even with GCN better effeciency, that should be a small egde to the 5770.

I'm not so sure about that, see my post on page three towards the end, because in real life usage DX 11.2 should aid the Kaveri IGP quite a lot & if power is not a constraint I'd expect an overclocked APU to come damn close to HD5770 GDDR5, however you're right when it comes to bandwidth constrained situations it will struggle big time though I do expect HSA/Huma to solve some of those problems, talking about real world tasks & gaming here.

 

[Abwx]

John Taylor followed up Su's speech with a demo of Kaveri running Battlefield 4. The game was set to 1080p with medium detail, and the frame rate hovered between 27 and 33 FPS during the single-player campaign's intro sequence. AMD showed the same demo on a Core i7-4770K machine with a GeForce GT 630, which sputtered along at 11-12 FPS.

http://techreport.com/news/25636/amd-kaveri-hitting-desktops-on-january-14-with-mantle-support

 

[LogOver]

so no discussion about hsa?
what do you guys think about gpu-accell. java? if this gets a perf boost on amd parts what do you think will happen for their server business?

also can one compile/translate/convert javascript into opencl/hsail to run it on the gpu or is that kind of workload not suited for a gpu?

HSA is a good intention... But execution is another story. Lets hope it wont go nowhere like another AMD initiative.
http://www.amd.com/us/press-releases/Pages/amd-announces-new-levels-of-realism-2009sept30.aspx

 

[EndlessWargasm]

what do you guys think about gpu-accell. java?

also can one compile/translate/convert javascript into opencl/hsail to run it on the gpu or is that kind of workload not suited for a gpu?

Javascript =/= Java

They only thing those two languages have in common is coffee (and a bit of C syntax).

 

[Asterox]

On Topic,
This time a single Module two threads Kaveri with 256 SPs will be a nice entry APU. CPU performance will be considerably faster than Trinity due to higher IPC and lower impact of CMT.

I will really like to see A8-10 Kaveri vs Core i3 Haswell.

A lot of meaningless trash/troll posts in this topic, so back to the main topic now you mention Kaveri APU with 1 Module or 2 Cores.Speaking of that in this comparison we can see the obvious difference, in the L1 Cache configuration between Dual Core Kaveri APU vs. Dual Core Trinity APU.

 

[NTMBK]

I wonder if those reported clocks are base clocks, or turbo clocks? Obviously AMD has put a lot of work into turbo clocks for GCN1.1- just take a look at how much R9-290 performance is affected by fan speed. Hopefully Kaveri will have an aggressive GPU turbo.

 

[JohnDC]

Kaveri looks promising.If the price is similar to their previous apu's this should make a nice budget PC.I can't wait for review sites to get these and see how they perform in benchmarks and how they overclock.

It Will be interesting to see the future of this architecture.Like what happens when developers start using the SDK that AMD developed for it and if AMD will make these with more CPU modules and even more graphics cores.

 

[monstercameron]

Javascript =/= Java

They only thing those two languages have in common is coffee (and a bit of C syntax).

cute. I know what i posted

 

[mikk]

BTW: I called 512SP long ago

This was known since many months, AMD himself confirmed it in a slide long time ago. They also claimed to have 1050 Gflops, now they just have 850 Gflops. The GPU alone 737 Gflops.

 

[USER8000]

Lies and misinformation characterize almost every one of your posts, SiliconWars, at least with respect to sticking to the facts re: Intel.

Broadwell goes into HVM during Q1 2014 for launch likely 1 year from Haswell, implying June launch.

Hardly 2015, although keep spinning, buddy.

Now we're resorting to personal attacks.

??

 

[AtenRa]

HSA is a good intention... But execution is another story. Lets hope it wont go nowhere like another AMD initiative.
http://www.amd.com/us/press-releases/Pages/amd-announces-new-levels-of-realism-2009sept30.aspx

This one will help a little

 

[EndlessWargasm]

cute. I know what i posted

Fair enough.
The way you wrote that seemed like two connected thoughts.