Kaveri performance

[galego]

Which would be the performance of this APU compared to existent chips?

Taking the IPC gain claimed by AMD and the replacement of 8-core piledriver opteron by 4-core steamroller Berlin APUs I think that

Kaveri = FX-8100

for existent software and

Kaveri >>> FX-9590

for HSA software, because the FX has not iGPU.

 

[Enigmoid]

Honestly, not a chance. (No chance kaveri is going to be 70%+ faster CPU wise than richland).

(HSA requires a GPU so you can't compare to FX).

 

[NTMBK]

Depends on what workload. In plenty of games, the 6800k will already be significantly faster than the 8100- greatly higher clock speed, plus the Piledriver IPC improvements, will make it a lot better.

Multithreaded performance? Nah, not seeing it. If it can reach the 6300 in MT performance I will be happy.

 

[parvadomus]

I can see up to 30%-40% IPC increase (in MT at best for ceritan workloads) if leaked specs are right. So it should perform like a low clocked i5.

 

[galego]

I can see up to 30%-40% IPC increase (in MT at best for ceritan workloads) if leaked specs are right. So it should perform like a low clocked i5.

30-40% over what? Bulldozer?

Honestly, not a chance. (No chance kaveri is going to be 70%+ faster CPU wise than richland).

(HSA requires a GPU so you can't compare to FX).

I don't think that 8100 is 70% faster than Richland. I have benchmarks where a 8150 is similar to Trinity.

Are you certain that HSA requires a GPU? I believed that HSA enabled software can work on CPU, except that will run slower.

Depends on what workload. In plenty of games, the 6800k will already be significantly faster than the 8100- greatly higher clock speed, plus the Piledriver IPC improvements, will make it a lot better.

Multithreaded performance? Nah, not seeing it. If it can reach the 6300 in MT performance I will be happy.

Do you mean

Kaveri > 8100 for ST
Kaveri < 8100 for MT

Right?

Does not the 6300 is faster than the 8100 for MT?

 

[SiliconWars]

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.

There is one thing that nobody has picked up on yet though, and that is the lower end chips are starting to get very interesting. With 512 shaders on the top chip, 384 shaders might be the low end, or definitely the midrange APU. That's 384 GCN, which should be better than the current 6800K performance for less than $70. We should now be starting to see the very low end ($55 chips) being 1080p capable (low settings, 30 fps), and this should also start to make the i3's 4600 graphics appear useless and overpriced, which hopefully will force Intel into releasing GT3 chips on their i3's.

 

[frozentundra123456]

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.

There is one thing that nobody has picked up on yet though, and that is the lower end chips are starting to get very interesting. With 512 shaders on the top chip, 384 shaders might be the low end, or definitely the midrange APU. That's 384 GCN, which should be better than the current 6800K performance for less than $70. We should now be starting to see the very low end ($55 chips) being 1080p capable (low settings, 30 fps), and this should also start to make the i3's 4600 graphics appear useless and overpriced, which hopefully will force Intel into releasing GT3 chips on their i3's.

Perhaps, but you are still going to be bandwidth limited. Even the highest apu will have the same shaders as a 7750 and will be slower than a DDR 5 7750 due to bandwidth limitations. Not to mention that the 7750 is quite old now and hopefully faster discrete cards in the same class will be out soon. I agree that it would be nice if Intel released gt3 on the i3, but I still see a cheaper CPU with a discrete card as a better solution on the desktop for gaming.

 

[Enigmoid]

30-40% over what? Bulldozer?

I don't think that 8100 is 70% faster than Richland. I have benchmarks where a 8150 is similar to Trinity.

I'm talking about fully multithreaded situations. 8150 isn't even close to trinity.

Are you certain that HSA requires a GPU? I believed that HSA enabled software can work on CPU, except that will run slower.

Do you mean

Kaveri > 8100 for ST
Kaveri < 8100 for MT

Right?

Does not the 6300 is faster than the 8100 for MT?

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

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.

There is one thing that nobody has picked up on yet though, and that is the lower end chips are starting to get very interesting. With 512 shaders on the top chip, 384 shaders might be the low end, or definitely the midrange APU. That's 384 GCN, which should be better than the current 6800K performance for less than $70. We should now be starting to see the very low end ($55 chips) being 1080p capable (low settings, 30 fps), and this should also start to make the i3's 4600 graphics appear useless and overpriced, which hopefully will force Intel into releasing GT3 chips on their i3's.

Judging by the mobile chips (87xxm series) I can generally say that (3dmark 11 score)

(note that clockspeeds vary a bit as these are mobile chips and manufacturers have set them to different speeds).

a10-6800k : 1457 points at 844 mhz
384 gcn cores @ 650 mhz + 900 mhz DDR3 : ~1370 pts
384 gcn cores @ 775 mhz -825 mhz + 900 mhz DDR3 : ~1600 pts
384 gcn cores @ 850-900 mhz + 1000 mhz GDDR5 : ~2135 pts

(notebookcheck)

Looks pretty restricted by memory bandwidth to me. At a similar clockspeed I would expect an increase from ~1450 points to ~1600-1700 points. AMD seems to be much more sensitive to memory bandwidth than nvidia or intel.

 

[NTMBK]

Perhaps, but you are still going to be bandwidth limited. Even the highest apu will have the same shaders as a 7750 and will be slower than a DDR 5 7750 due to bandwidth limitations. Not to mention that the 7750 is quite old now and hopefully faster discrete cards in the same class will be out soon. I agree that it would be nice if Intel released gt3 on the i3, but I still see a cheaper CPU with a discrete card as a better solution on the desktop for gaming.

Have we heard anything about the cache for Kaveri? AMD were experimenting with on-package memory a while ago, but I don't know if anything came of it- a Crystalwell-style on-package cache would really help it out. (Not that it's especially likely, but its the only way around DDR3's limits I can see.)

 

[SiliconWars]

I believe GCN was designed with lowering bandwidth requirement in mind so there should be an uplift just from that. There may be advancements from hUMA as well, but I don't think there is anything like Crystalwell coming. For me they were banking on DDR4 being ready by now (this may also have been the cause of the initial delay).

It's not all bad though, as I said the lower end parts will also likely be bandwidth restricted which just means they will be insanely powerful for such cheap parts. Mobile should be good as well, if the bandwidth is a wall then clocks can be reduced for much less loss overall, but nice savings on the battery.

 

[Haserath]

According to Wiki, DDR4 was already planned to be here(2012 release) and at least 2133 Mhz mainstream and over 3Ghz for enthusiast. So with point to point bandwidth increase and higher speeds, I doubt they planned using anything else.

And according to crucial, DDR4 should be out in late 2013.
http://www.crucial.com/promo/DDR4.aspx

Has anyone heard what Ram type Kaveri is planned to use?

 

[mikk]

No DDR4 for Kaveri. Maybe for their 2015 APU Carrizo.

On the mainstream side of things, AMD will launch the “Carrizo” APU in 2015. It will be based on the “Excavator” CPU core and Bolton-D4 A88X chipset. AMD hasn’t named the GPU yet, only stating it will have an unnamed “next generation” GPU. AMD lists its TDP at 65W, but says it is “scoping” 45W. AMD also says that “Carrizo” will be scoping DDR4, meaning that they haven’t confirmed support but are planning on having it.

http://vr-zone.com/articles/new-con...ineup-kaveri-delayed/47455.html#ixzz2aQaXz0rE
​

 

[JDG1980]

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. This would put IPC roughly on par with Intel's Nehalem, maybe a little less. This would give AMD the option of using lower-clocked parts to save on TDP without sacrificing performance, or maintaining existing clock speeds and increasing performance substantially for the high-end "black edition" parts.

Another significant fact is that Steamroller has separate decoder units for each of the two cores in a module, whereas Bulldozer/Piledriver had to share. This should drastically reduce the CMT penalty, if not eliminate it completely. Kaveri will be a true quad-core part, not quad-core with an asterisk like Trinity/Richland.

 

[ShintaiDK]

AMD never said SR would have 30% higher IPC than PD. Its simply a forum hype.

 

[sm625]

Come on this is AMD. 30% higher IPC would actually put them in competition... we cant have that now can we?

 

[PPB]

Considering Piledriver already was a 15% perf increase in heavy MT workloads at the same clock, I think it should be in the +20-25% IPC territory.

 

[jhu]

Considering Piledriver already was a 15% perf increase in heavy MT workloads at the same clock, I think it should be in the +20-25% IPC territory.

15% in heavily threaded programs? I'm not seeing it. For Povray, it's about 4%. For Cinebench, it's about 1%. This is taking into account FX 8150's turbo of 3.9 GHz and FX 8320's turbo of 3.8 GHz.

 

[Enigmoid]

Considering Piledriver already was a 15% perf increase in heavy MT workloads at the same clock, I think it should be in the +20-25% IPC territory.

Piledriver was a good amount of frequency increase from 8150 to 8350.

 

[PPB]

Check hardware.fr vray and mental ray benchs. Actual production renderers, ya know

 

[Sweepr]

Check hardware.fr vray and mental ray benchs. Actual production renderers, ya know

7.7% faster on average @ 4GHz according to them (applications, not games).

 

[jhu]

Check hardware.fr vray and mental ray benchs. Actual production renderers, ya know

It's your job to support your claim, not anyone else's.

 

[PPB]

Sigh, considering I already posted those charts (altough with the intention of showing Cinebench's irrelevance in showing rendering performance) here we go (again):

It also shows Haswell as a nice performance upgrade from Ivy bridge, +20% perf at same clock whereas that POS Cinebench only shows less than 10%.

Seems some people need to be spoon fed with information, google won't bite

 

[NTMBK]

Seems some people need to be spoon fed with information, google won't bite

If you make a claim in an argument, it is your job to provide evidence to support that claim; if you give no evidence, there is no reason to believe your claim.

 

[AtenRa]

CPI(Cycles Per Instruction, from now on ill use this instead of IPC) may be up to 15% and im expecting a nice boost in Gaming. MT Scaling and performance will also gain a lot.

But, we may see lower clocks so overall performance may only raise 10-15%

 

[NTMBK]

CPI(Cycles Per Instruction, from now on ill use this instead of IPC) may be up to 15% and im expecting a nice boost in Gaming. MT Scaling and performance will also gain a lot.

But, we may see lower clocks so overall performance may only raise 10-15%

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