Kaveri APU Features 20% CPU and 30% GPU Performance Uplift Over Richland

[Abwx]

Why did you not give the rest of the quote?

Because i also gave the full link where everybody
can read by himself the whole article....

"When pressed for details after the briefing, however, Kozak clarified that Kaveri should only be equivalent in terms of combined CPU and GPU compute power. If one measures x86 performance on its own, Kozak said, "we'll lose." However, Kozak expects Kaveri's integrated graphics, bolstered with Mantle support, to be better than the latest version of Intel's HD Graphics."

Richland's igp is already ahead of HD4600, so I dont see anything game changing there.

If he mention better GFX than Intel it s surely because
he s talking of something else than HD4600 perfs.

 

[JDG1980]

"When pressed for details after the briefing, however, Kozak clarified that Kaveri should only be equivalent in terms of combined CPU and GPU compute power. If one measures x86 performance on its own, Kozak said, "we'll lose."

I don't think anyone seriously expected Kaveri to match an i5-4670K in terms of raw CPU horsepower. That would require at least a 50%-60% increase in IPC, which was never in the cards. A more reasonable guesstimate would be a 20%-30% IPC increase plus elimination of the CMT penalty, which would put Kaveri roughly on par with Nehalem. Catching up to Sandy Bridge will probably have to wait for Excavator. The good news for AMD is that Intel hasn't really made any major IPC jumps since SB; it's reasonable to assume that the P6 architecture probably doesn't have all that much more to squeeze out. AMD may never quite catch up, but we should expect the IPC gap to narrow substantially in the next couple of years.

 

[sniffin]

I don't think anyone seriously expected Kaveri to match an i5-4670K in terms of raw CPU horsepower. That would require at least a 50%-60% increase in IPC, which was never in the cards. A more reasonable guesstimate would be a 20%-30% IPC increase plus elimination of the CMT penalty, which would put Kaveri roughly on par with Nehalem. Catching up to Sandy Bridge will probably have to wait for Excavator. The good news for AMD is that Intel hasn't really made any major IPC jumps since SB; it's reasonable to assume that the P6 architecture probably doesn't have all that much more to squeeze out. AMD may never quite catch up, but we should expect the IPC gap to narrow substantially in the next couple of years.

I think the lack of serious gains for Intel is less about not being able to squeeze anymore out and more about their unwillingness to do so. They made Haswell slightly wider and the result is higher load power usage. At this point AMD aren't their focus.

If Steamroller catches AMD up to Nehalem that's really promising for them. It'll be the first significant stride they've made in a very long time, and would put their CPUs in the good enough category. AMD don't need to catch Intel, they just need to be close enough for it to not really matter. With the exception of enthusiasts, Nehalem level performance is good enough for everybody.

GCN is so far ahead of Intel's graphics tech that almost as good x86 performance should propel them far ahead in anything that can be accelerated by gpu tech.

 

[nenforcer]

Catching up to Sandy Bridge will probably have to wait for Excavator. The good news for AMD is that Intel hasn't really made any major IPC jumps since SB; it's reasonable to assume that the P6 architecture probably doesn't have all that much more to squeeze out. AMD may never quite catch up, but we should expect the IPC gap to narrow substantially in the next couple of years.

As much as I'd want or hope for this the latest roadmaps from AMD suggest there will be no Socket AM3+ FX chips from AMD at all in 2014. It looks like Kaveri in January on Socket FM2+ is going to be it.

We're going to have to wait until 2015 for anything Excavator and by then Intel will have Haswell Refresh out and be working on either Broadwell and Skylake at .14nm.

 

[JDG1980]

As much as I'd want or hope for this the latest roadmaps from AMD suggest there will be no Socket AM3+ FX chips from AMD at all in 2014. It looks like Kaveri in January on Socket FM2+ is going to be it.

We're going to have to wait until 2015 for anything Excavator and by then Intel will have Haswell Refresh out and be working on either Broadwell and Skylake at .14nm.

That's why I said "the next couple of years", not 2014 specifically. Anyway, the mainstream Intel chips (everything except LGA 2011) max out at 4 cores, so AMD isn't really at a substantial disadvantage on that front even if they stick to APUs - as long as Steamroller manages to get rid of the CMT penalty.

My gut feeling is that there will likely be no Steamroller chip with more than 2M/4C, but that Excavator will probably see a server and/or FX release, though it may come after the corresponding APU. I don't think any future architectures will be on AM3+; it's more likely that AMD would create a new socket. It would presumably be designed for server use first and foremost, with the lower-binned chips going to enthusiasts. Excavator widens the execution units substantially, and that seems unnecessary if AMD is only going to focus on the low end. At this time it seems that AMD is primarily focused on the console contracts (PS4/XB1), and secondarily on low-power chips for portable devices, with servers and enthusiasts being an afterthought. But in the long run, if they want to turn enough profits to sustain R&D, they're going to have to go after the high-margin server market. And APUs just aren't going to cut it there.

As for Broadwell and Skylake, I don't see this changing the equation much. Intel is no longer focused on squeezing out the last possible bit of IPC; they are more concerned now with reducing power consumption and improving the GPU side. I don't expect to see more than 10% additional IPC from Skylake over Haswell, and clock speeds aren't going to go up - they peaked with Sandy Bridge, and all of Intel's process nodes below 32nm have been optimized for power consumption, not high clocks. We will be seeing a greater convergence between Intel and AMD with regards to mainstream processors: Intel will improve their graphics to approach (if not quite catch) AMD, and AMD will increase their IPC to approach (if not quite catch) Intel. And both sides will obsess over reducing power consumption for the mobile market, which is apparently the only thing that matters now. No, I'm not bitter...

 

[inf64]

Prohardver has a slide explaining two top desktop Kaveri's specs

Max boost for 7850K is 4Ghz. It's probably ST boost but I wonder if it might be all core boost too due to new Turbo core based on temp. monitoring?

So in best case it's 4Ghz Kaveri against 4.1-4.4Ghz Richland. In order to be an upgrade over Richland and to match AMD's generational roadmap performance projections, it needs to perform at least 10-15% batter while having ~0.9x the clock speed of Richland.
In order this to happen we need:
1)for base clocks: 3.8xY=4.1x1.125 (12.5% performance increase is assumed as mean value between 10 and 15% AMD projected)=> Y=1.21x or 21% IPC jump
2) for ST boost clocks: 4xY=4.4x1.1 (Assuming worst case of ~10% better performance) => Y=1.21x or again 21% IPC increase
3)for average clocks: 3.9xY=4.2x1.125 => 1.21x or again ~20% IPC increase

So to sum it up, Kaveri needs to have around ~20% higher IPC with these launching clocks in order to justify AMD's claims and to be upgrade to Richland (10-15% better).

 

[NTMBK]

AMD tried balls out performance with the fx 9-series, have you forgotten how much it has been panned...

AMD also tried balls out performance with Hawaii, which went pretty damn well.

 

[Ventanni]

AMD also tried balls out performance with Hawaii, which went pretty damn well.

I, for one, would love an 8 core streamroller chip, but both you and I know full well that you can't compare amd's gpu competitiveness with nvidia to their cpu competitiveness with Intel.

 

[NTMBK]

I, for one, would love an 8 core streamroller chip, but both you and I know full well that you can't compare amd's gpu competitiveness with nvidia to their cpu competitiveness with Intel.

Who says I'm talking about CPU? An APU which gives CPU performance comparable to a Nehalem quad core (which Kaveri hopefully does) and gives great graphics performance would be very, very nice. A 100W APU with a big fat ESRAM cache and roughly 7790 or 7850 performance would be amazing for a gaming HTPC, or Steambox.

 

[monstercameron]

Who says I'm talking about CPU? An APU which gives CPU performance comparable to a Nehalem quad core (which Kaveri hopefully does) and gives great graphics performance would be very, very nice. A 100W APU with a big fat ESRAM cache and roughly 7790 or 7850 performance would be amazing for a gaming HTPC, or Steambox.

trust me when I say I want that aswell, right now I am running a 965BE and an gtx650 but to replace that with a one chip kaveri apu of similar levels of gpu performance would be awesome.

 

[blackened23]

I think the lack of serious gains for Intel is less about not being able to squeeze anymore out and more about their unwillingness to do so. They made Haswell slightly wider and the result is higher load power usage. At this point AMD aren't their focus.

No. It's about mobile. In other words, it's about selling a product that matters to the market that is increasingly going mobile for computing products, that is what intel is focused on. Now every generation is bringing 10% IPC increase pretty consistently, but with a MUCH larger efficiency jump. That's pretty much what AMD should be focused on as well with Kaveri, desktop sales are diminishing by double digits every quarter so desktop IPC isn't a sustainable business.

That's why Haswell wasn't a huge jump on desktop, but was a huge (efficiency) jump for mobile - the core Haswell is getting 12+ hours of battery life in macbook airs. That's what intel is focused on, not desktop IPC. It would make sense for AMD to do the same, but their prior APUs have been very good for the LGA form factor with the mobile parts being severely cut down and not so good. We'll see what mobile Kaveri does in that respect...I fully expect LGA kaveri to be very good for HTPC use as Richland was, but I have doubts on the mobile Kaveri parts. I'll hold judgement until it actually is released, though.

 

[Abwx]

Yet another geekbench entry...

Kaveri 3.5 vs FX8350 3.5 :

http://browser.primatelabs.com/geekbench3/compare/223722?baseline=243160

Kaveri vs Richland , both seems to be at 3.5 and more interestingly
are using a single channel with 4GB wich invalidate the MT scores
but still allow a comparison of ST scores :

http://browser.primatelabs.com/geekbench3/compare/89728?baseline=223722

 

[parvadomus]

Looking at that benchmarks, all AMD needs to be competitive (vs i5s) is a 3M APU.

 

[CHADBOGA]

Yet another geekbench entry...

Kaveri 3.5 vs FX8350 3.5 :

http://browser.primatelabs.com/geekbench3/compare/223722?baseline=243160

Kaveri vs Richland , both seems to be at 3.5 and more interestingly
are using a single channel with 4GB wich invalidate the MT scores
but still allow a comparison of ST scores :

http://browser.primatelabs.com/geekbench3/compare/89728?baseline=223722

Kaveri's improvement over Richland clock for clock is 16.8%, but Kaveri is going to be released running at a slower clock speed. D:

 

[NTMBK]

Kaveri's improvement over Richland clock for clock is 16.8%, but Kaveri is going to be released running at a slower clock speed. D:

It's unverified Geekbench, don't panic just yet.

 

[cytg111]

Kaveri's improvement over Richland clock for clock is 16.8%, but Kaveri is going to be released running at a slower clock speed. D:

Nice 2 second effort FUD - or not.
Maybe if you average the score.. but look at the peaks, factor out the 4 core+ workloads, and you'll get some pretty hefty improvements, ST looks good.

 

[Abwx]

Another comparison, Kaveri 3.5 being about equal to a 4.1 Richland,
that s 17% better IPC , both configuration seems to run a single 4GB
memory channel.

http://browser.primatelabs.com/geekbench3/compare/82001?baseline=223722

 

[DownTheSky]

It seems to be lacking multicore performance. If you look at the charts single core perf improvement is bigger than multicore. Yet AFAIK AMD stated improvements in multicore scaling.

 

[Abwx]

There s a single memory channel , the same 6800K has 10-15%
better scores with a dual channel and we dont know to wich extent it would improve Kaveri s numbers if ever it has a somewhat better controler , so far thoses scores should be relevant only for ST comparisons since the configurations are about the same but still the OS are W7 for Kaveri and W8 for Richland.

 

[SlowSpyder]

Nice 2 second effort FUD - or not.
Maybe if you average the score.. but look at the peaks, factor out the 4 core+ workloads, and you'll get some pretty hefty improvements, ST looks good.

And the 8350 gets the added benefit of L3 cache. Against Richland Kaveri may be a few percentage points faster per clock than against the 8350.

 

[CHADBOGA]

It's unverified Geekbench, don't panic just yet.

How does one panic, when they are simply underwhelmed? :hmm:

 

[JDG1980]

Pure synthetic benchmarks are garbage. Let's see some real-world tests like LAME, 7-Zip, or Starcraft II, or at least something like Cinebench that roughly approximates actual workloads.

When HardOCP tested Vishera vs. Zambezi, Vishera actually came off worse in the pure synthetics. But in every single real-world test, it was ahead by at least a couple of percent.

I suspect a lot of what we're seeing in Geekbench might be due to the lack of L3 cache in Kaveri. How this affects applications in the real world remains to be seen.

 

[ElFenix]

7-Zip

are zip programs really in that heavy usage all the time? i always figured i was more disk limited than anything with those, anyway.

 

[NostaSeronx]

are zip programs really in that heavy usage all the time? i always figured i was more disk limited than anything with those, anyway.

7-zip ultra with x+3x threads.

 

[USER8000]

The Kaveri sample testing was an ES,and it might not have things like the CPU Boost activated. The leaks say the top desktop A10 model runs at 3.7GHZ base clockspeed and 4GHZ boost clockspeed. So if the 3.5GHZ Kaveri ES is matching a 4.1GHZ A10 6800K in multi-threaded benchmarks,that is around a 9% improvement in multithreaded performance for the A10 7850K. If the ES sample is not boosting that means a 3.5GHZ Kaveri CPU is matching a 4.4GHZ Richland. If that is the case single threaded core IPC is a big improvement.

The thing is if the laptop models ship at clockspeeds similar to Richland,we might see much bigger improvements in that market.