AMD CPU speculation... and expert conjecture

[Cazalan]

Now, if only there were someone on this forum to argue this same exact point for the past four years...

...oh wait, that's me. Glad AMD finally came around.

Hah! Doesn't every company that just released a new product say that the best solution is the one they just released?

Funny though when you read the BIOS guide there's still a mention or two of the 3 module support and GDDR5. These weren't fully scrubbed from the document (not yet anyway).

Hardware is becoming more like today's games. I wanted to get the latest Sim City but I knew it would have launch problems and didn't buy it. And a year later many of these are finally getting implemented. Kaveri is a lot like that, in that HSA software will take some time. Mantle will take some time. But I think they're headed in a good direction.

 

[Cazalan]

You see the trend compared against Richland: ~27% in gaming, mostly due to the GPU. ~135% in compute, again, due to the GPU.

Less then 8% in multicore CPU performance, and the same single-core performance [about 60% IB].

So matches what everyone else is seeing.

I like S|A but boy to they make some odd comparisons sometimes. That i7 is 2x the cost.

I suppose it could be worse. AT used a 3x more expensive Iris Pro.

 

[de5_Roy]

^^ there's no performance equivalent sku from intel at kaveri's price point. even against present-price a10 7850k. compared against core i7 4770k or 4770R, 7850k seems like better value. core i3 is a better competitor against kaveri. it feels weird writing that, since core i3s are undercutting kaveri apus at their current prices. used to be the other way.

s/a still takes the prize for worst comparison. they benched hd4600 against a radeon hd 7790. the 7790 won. >_>

edit: after reading the s/a analysis, i think carrizo's exc cores won't clock high either, since amd will be using high-density cell libraries(...?) for the igpu. and carrizo might use bulk if glofo still can't figure things out. amd isn't blaming glofo.. but glofo has been consisten about them screwing up with anything with a gpu. the delays, redesign and sr cores were collateral damages this time.

 

[juanrga]

I think Juanrga did a great job with all his work on his blog, the only ones to blame was GloFo and possibly AMD for being to stubborn to stick with them.

Juanrga numbers where rather high in the end, but his work done on comparisons and estimations were good enough, at the end of the day it was not juanrga estimated numbers, but the numbers provided by AMD as claim that Kaveri will have X% of improvements where just wrong.

I am still working in the review and will update my article with the benchmark results, but first results are very promising. See above message

http://www.tomshardware.co.uk/forum/352312-28-steamroller-speculation-expert-conjecture/page-220#12445948

x264. I predicted 102 and Kaveri scores 94.45. Percent error = -7%
JTR. I predicted 4310.46 and Kaveri scores 3631. Percent error = -16%
C-Ray. I predicted 38.96 and Kaveri scores 37.00. Percent error = -5%
Himeno. I predicted 845.64 and Kaveri scores 958.10. Percent error = +13%

My predictions assumed a 4GHz CPU. Kaveri is finally 3.7GHz. If I had know the real frequency before, I would offer better estimations. For instance, I would predict 94.35 (=102x3.7/4) for x264. The percent error had been reduced to 1% then :lol:

The JTR score is too small. Part of the discrepancy can be explained by the clock reduction, but still there is about a 9% less performance. It could be a compiler regression and the score improve in a posterior test with updated compiler; it could be related to some architecture detail. I don't know.

At the other hand, the Himeno score is much better than I expected. The improvement is even more impressive when we consider the clock reduction. The improvement is of 22%, which implies more than 30% IPC. The Himeno test is very heavily FP oriented. This huge IPC gain is about the same found in Wprime, in games, and in Floating Point GeekBenches. Apparently the FPU units were not improved. I believe that the improvement in floating point performance is due to the improvements made in the L2 cache, but I need to ask for a second opinion.

Up to this moment, I can say that Kaveri performs poor than I expected by about 5--16% in integer workloads, but it is still at the Sandy Bridge i5 level. Kaveri is slightly behind the i5-2500k in C-ray or x264, but outperforms the i5-2500k in John The Ripper.

Unexpectedly, the floating point performance of Steamroller is rather better than I expected (about 13% better) even with a clock deficit of ~10%.

The average of the percent error is -4%. Would I say that Kaveri CPU performs about 4% poor than I predicted?

 

[logainofhades]

The 7850k is selling on newegg for like $190 now. You can get an FX 6300 and an HD 7750 for like $195. :lol:

 

[8350rocks]

Well, I was considering adopting the APU for my itx/OC the crap out it build. Looks like I will have to wait until prices come down and updated bios out of the box arrive. Now if only my microcenter can stock some ITX mobos.

In the meantime. Team Vulcan 2133 RAM. Anybody know of them? Quality?

Overclock the crap out of it all you want, it still won't match the cpu performance of an i3. You will be able to match a PhII at 4.0 or so at best. You may as well go with a 760k and HD 7770 and have better overall performance for nearly the same cost.

This is not entirely true...

Richland was able to hit 5.0 GHz pretty regularly, at that kind of speed, the i3 cannot compete. So, while you may have to overclock the hell out of the APU, you can exceed i3 performance. If kaveri will OC into the mid 4 GHz range, it may come out similar.

FYI 760k has the same handicap as the A10-6800k, no L3 cache. While it is cheaper to buy the 760k, you would still be even better off with something like the 4300/6300 because of the better cache for only a few dollars more.

Essentially FX > APUs, exactly as I said all along...

 

[logainofhades]

Yea, APU's have no appeal really to me at all except maybe in mobile.

 

[noob2222]

@juan

Rofl, I have to give you credit. You never give up.

First absolute error is just that, absolute. It is not a percentage.

2nd. Nice try on finding an overclocked kaveri and pretending its stock.

http://openbenchmarking.org/result/1401152-SO-1401123PL97

also try using the updated i5 figures since the software is newer. Dont skew the results by ignoring it.

X264 stock intel 105.58 vs 90.34 for 3.7 ghz kaveri.

Also quite a discrepency in the c-ray results. Looks like c-ray is heavily compiler dependent.

http://openbenchmarking.org/result/1306303-SO-AMDA1068056

 

[juanrga]

@juan

Rofl, I have to give you credit. You never give up.

First absolute error is just that, absolute. It is not a percentage.

2nd. Nice try on finding an overclocked kaveri and pretending its stock.

http://openbenchmarking.org/result/1401152-SO-1401123PL97

also try using the updated i5 figures since the software is newer. Dont skew the results by ignoring it.

X264 stock intel 105.58 vs 90.34 for kaveri.

Also quite a discrepency in the c-ray results.

The absolute error of two percentages is... a percentage.

But you are entirely right that the link that I used is a Kaveri at @4GHz. I didn't know. I got it from here

http://www.phoronix.com/scan.php?page=news_item&px=MTU3MTQ

and I assumed was stock, because there is no mention of OC LOL

Thanks by the correction!

That C-ray result that you report is another. The mine is a match=native.

I will correct the above numbers. This is the first time that you are useful. LOL

 

[jdwii]

Well i'm done being negative at juanrga if it means anything i wish he was right, i wish the CPU was as powerful as a I5 2500K i would probably buy it then. But for now it has to be in a 500$ budget build.

 

[logainofhades]

Well i'm done being negative at juanrga if it means anything i wish he was right, i wish the CPU was as powerful as a I5 2500K i would probably buy it then. But for now it has to be in a 500$ budget build.

I would have liked to seen at least FX 4300 performance with GPU performance of an HD 7750. The $180 price tag would have been slightly more justified then.

 

[noob2222]

Lets try without the percentages. Predicted 102. Actual 90.3. Starting point of 4.1 ghz richland 87. The starting point isn't 0 or 1 because an increase of 15% of 0 is 0 and 1 is 1.15.

Predicted increase of 15, actual increase of 3. Error of 12/15.

 

[juggernautxtr]

I'd like to see the performance of the cpu done on a high performance soi, bulk has got to be leaky as hell.
really Bulk? you know Glofo is just reaming AMD, and smiling about it.
i'd almost bet we'd see some way better numbers on a good soi.

 

[noob2222]

^^I don't remember seeing any thermal tests.

 

[blackkstar]

The 7850k is selling on newegg for like $190 now. You can get an FX 6300 and an HD 7750 for like $195. :lol:

@juan

Rofl, I have to give you credit. You never give up.

First absolute error is just that, absolute. It is not a percentage.

2nd. Nice try on finding an overclocked kaveri and pretending its stock.

http://openbenchmarking.org/result/1401152-SO-1401123PL97

also try using the updated i5 figures since the software is newer. Dont skew the results by ignoring it.

X264 stock intel 105.58 vs 90.34 for 3.7 ghz kaveri.

Also quite a discrepency in the c-ray results. Looks like c-ray is heavily compiler dependent.

http://openbenchmarking.org/result/1306303-SO-AMDA1068056

Renderers are very easy to play compiler tricks with. I can make FX 8350 faster clock for clock than 3930k with enough tricks.

In fact, I've actually done just that. Had my 5ghx FX 8350 pull ahead about 30% over a 4.2ghz 3930k in Blender. Of course I was running Blender on Gentoo with bdver2 for everything and he was running the official version from blender.org, but at the end of the day, I could have easily put that on a review site, titled a graph "Blender results" and showed the 3930k getting slaughtered.

Open source rendering benchmarks are one of my favorite benchmarks, because you can play with the compiler and see just how good it can perform. Seeing Phoronix results with FX 8350 vs Intel was enough to make me consider completely ignoring going Intel for my last build. Yeah, AMD doesn't do so great in Cinebench, but open source rendering projects like C-Ray and Blender let you compile the benchmark however you like, and it's a huge opportunity to grab free performance.

In fact, it's just about the only reason why I don't jump ship to 3d Studio Max or Maya. Blender rendering is significantly faster, I've learnt the tools, and the render results for baking aren't that much worse if you know what you're doing. Although admittedly it's a lot easier to get good results with something like Mental Ray, but the thought of running Nvidia software on an AMD CPU sounds painful to me.

HSA has a bright future, and as soon as it gets mature enough so that Gentoo has an HSA flag, I'm grabbing an APU or some sort of HSA compatible AMD solution and going Gentoo bonkers with it. It'd be more than worth it to have it set up as a server for all my needs and to be able to SSH with X into the machine for HSA accelerated stuff.

It's just too bad running Gentoo and running home servers isn't more common, HSA and AMD are going to be absolutely amazing for this. I thought my Opteron 165 would be enough to keep me happy, but it's really chugging along here creating databases and stuff.

 

[palladin9479]

a8 7600's memory bandwidth looks so... bad..
http://www.pcper.com/reviews/Processors/AMD-A8-7600-Kav...
it's amazing that the apu performs so well despite that. it's a damn shame that it's constricted this way (<- after seeing how much haswell core i3 gets). injustice! oh, don't use ram slower than ddr3 1600 if you wanna game with the igpu.

Yes and no. The vast majority of "memory benchmarks" are actually cache benchmarks due to the front end prefetcher. Our very own benchmark writer found that out. Intel has a vastly superior cache subsystem, this more then anything else is what gives them such a large lead in single core performance. Look at the memory access latencies if you want to see it in action. The number crunching part of the CPU can't access memory directly, instead it's gotta go through the MMU which is serviced by the prefetcher / cache subsystem. Intel's design is able to ensure what the CPU needs is always present and ready, where as AMD's design stalls out more often. To assist this AMD went with an exclusive cache system to make the window larger, but that also makes the overall latency higher.

None of the effects the iGPU portion as it doesn't use the CPU's cacheing subsystem, only it's MMU. So what you see with Sandra is the performance of the CPU's cache and how it effects the ability to keep the cores fed. I actually suspect that's why AMD went with a wide 2 ALU design vs their older deep 3 ALU design, they figured they couldn't keep all 3 ALU's busy with a single task so just whack one off to save die space. Guess nobody thought about gaming, which is something that actually can use all three ALU's. Intel actually introduced a fourth ALU to the haswell uArch which is why the haswell i3's gained such performance with HT.

http://www.realworldtech.com/haswell-cpu/4/

Good article walking through all the components involved and what they all do.

I plan on upgrading my 100W A10-6800K to the 65W A8-6600 with the focus being to bolt on one of Noctura's 65W ultra-quiet coolers and fully eliminate sound from my living room. Was playing on my living room Richland box (the one I posted pics for awhile back) and as I got into things I could hear the CPU fan spinning up, which annoyed the sh!t out of me.

http://i537.photobucket.com/albums/ff338/palladin9479/PC%20Builds/20130917_232308_zps77588ff8.jpg

http://i537.photobucket.com/albums/ff338/palladin9479/PC%20Builds/20130917_231934_zps992b6846.jpg

Goal is to remove that huge white HSF and replace it with a much quieter one from Noctura. Need to get the power draw lower for that to happen.

 

[juanrga]

Second try.

http://openbenchmarking.org/result/1401145-PL-AMDA1078505

http://openbenchmarking.org/result/1401123-PL-AMDA1078581

http://juanrga.com/en/AMD-kaveri-benchmark.html

x264. I predicted 102 for Kaveri @4GHz and Kaveri @4GHz scores 94.45. Percent error = -7%

Kaveri at stock scores 90.34. Percent error = -11%

JTR. I predicted 4310.46 for Kaveri @4GHz and Kaveri at stock scores 3631. Percent error = -16%

Kaveri @4GHz would score ~3925. Percent error = -9%

C-Ray. I predicted 38.96 for Kaveri @4GHz and Kaveri @4GHz scores 37.00. Percent error = -5%

Kaveri at stock would score ~39.77. Percent error = ~2%

Himeno. I predicted 845.64 for Kaveri @4GHz and Kaveri @4GHz scores 958.10. Percent error = +13%

Kaveri at stock would score 886.24. Percent error = +5%

If I did no mistake the average of the percent errors is -5% (for stock). Would I say that Kaveri CPU performs about 5% poor than I predicted?

Special thanks to noob2222 by finding a stupid mistake that I did. He has always hated my article and accused me of lying (he has accused the cosmology site of lying as well and even trolled against BSN* when published my article). noob2222, how were your exact words against BSN*? Can you repeat them?

I don't expect some "hey! juan your predictions were rather good after all" from noob2222, neither from 8350rocks, jdwii, Hafijur, and others here.

I think I am done with this thread. Time to move to next project.

 

[palladin9479]

Well i'm done being negative at juanrga if it means anything i wish he was right, i wish the CPU was as powerful as a I5 2500K i would probably buy it then. But for now it has to be in a 500$ budget build.

I would have liked to seen at least FX 4300 performance with GPU performance of an HD 7750. The $180 price tag would have been slightly more justified then.

Which 7750? The 2GB DDR3-1600 version or the 1GB GDDR5 version?

The GDDR5 version should beat out anything any APU can produce due to raw memory performance alone. The DDR3 version suffers from a bigger memory bottleneck then current APU's do so it will generally lose. It's about $100~120 USD for the GDDR5 version and $80~90 USD for the DDR3 version. Which is why the A8-6600 is in such a nice price position, but the A10-7850 is pricing itself out of the market. APU's need to be $150~160 or less to really hit their value point.

 

[$hawn]

$190 is just the newness tax...wait for a few weeks, it'll come down to a price that it actually deserves.

My understanding is AMD totally ignored Desktop, and went all hardcore on optimizing for mobile, in the 2-3 GHz range. This would be helpful to servers as well.
The desktop chips are just basically clocked up from there, but unfortunately AMD and GlobalShitteries hit a power wall around 4GHz.

I want to see a top end Mobile Kaveri A10 against a A10-5750M. Looking at the 45W APU's, I'm pretty sure it should be a significant improvement at 35W and lower as well.

I don't know why, but over the past few months, most new notebooks from HP and Dell in India use 20-25W chips only. It's getting really hard to find standard 35W processor notebooks. Most probably, this is the segment AMD is trying to attack.

 

[$hawn]

Globel foundries did steamroller A , tsmc did steamroller B (kaveri)

Global foundries makes fx cpu, richland /trinity SOI

tsmc makes kaveri on bulk but also has production lines for soi only there is no capacity available for soi.

You can bet GF is working on something for amd to get back there future business,

Damn! I didn't know that Kaveri was made by TSMC!

 

[juggernautxtr]

I'd like to see the performance of the cpu done on a high performance soi, bulk has got to be leaky as hell.
really Bulk? you know Glofo is just reaming AMD, and smiling about it.
i'd almost bet we'd see some way better numbers on a good soi.

IIRC amd was on the hook for around 385 million in wafers for 4th quarter 2013, but a whopping 455 million for Q 1 2014 from GF. The fact AMD has not made a stink about it in public tell me they used that as a negotiating chip for a different deal

edit 404mil in q4 and 250 in Q 1

http://blogs.barrons.com/techtraderdaily/2013/11/18/amd-take-or-pay-globalfoundries-deal-may-spell-inventory-build-says-bernstein/

looks like some are out to plain hurt AMD any way they can by some of the comments made.

amd just needs one really good cpu/apu i remember seeing stocks from AMD at $8 dollars cpl years ago,but I know they have 1 8350 sold and a R9 290 with asrock extreme 9 mother board my upgrade is due this year.
my poor 6850 is starting to be a little slow in games.

 

[$hawn]

@tourist

I was right, Kaveri is manufactured by GloFo.

" All are on the same die, a 2.41 Billion transistor chip built on Globalfoundries 28SHP process using 245mm^2 of silicon all told." - from SemiAccurate

 

[juggernautxtr]

@tourist

I was right, Kaveri is manufactured by GloFo.

" All are on the same die, a 2.41 Billion transistor chip built on Globalfoundries 28SHP process using 245mm^2 of silicon all told." - from SemiAccurate

Careful i am rarely wrong.

most of the reviews i have seen said gloflo SHP/bulk

 

[jacobian]

Where are the claims of "optimizing for mobile" coming from? Yes, it looks like Kaveri APU will work better than Richland APU, but mobile Kaveri is half year away. The 45watt TDP A8 is _not_ mobile chip. And 45watt is still too high for mobile these days unless you're build a desktop replacement notebook with a high performance CPU like i7. The real power target for mobile these days is 15-35watts. If AMD was optimizing for mobile, why are mobile Kaveri a half year away? (Indeed, why? I want to know.. the laptop market should be more important IMO).

 

[$hawn]

Where are the claims of "optimizing for mobile" coming from? Yes, it looks like Kaveri APU will work better than Richland APU, but mobile Kaveri is half year away. The 45watt TDP A8 is _not_ mobile chip. And 45watt is still too high for mobile these days unless you're build a desktop replacement notebook with a high performance CPU like i7. The real power target for mobile these days is 15-35watts. If AMD was optimizing for mobile, why are mobile Kaveri a half year away? (Indeed, why? I want to know.. the laptop market should be more important IMO).

No one knows for sure. We are just speculatiing

If you look at Kaveri’s performance improvement over Richland at different wattages, it goes from almost nothing at ~95W, to quite impressive at 45W. If you extrapolate the line downwards, it hints that at 35W, we could expect some pretty decent performance gains.

IMHO the reason for not releasing mobile Kaveri first would be that AMD wanted to do a trial run of sorts with Desktop silicon. I think based on observations of how these chips perform, AMD might make a more better (less leakage) minor stepping before releasing a Mobile Kaveri. This is the first time AMD has made a 28nm CPU, that too, not on SOI, but on Bulk.
Again, the above statements have been pulled out from my rear end, so please don’t treat them as facts, because they are not.