AMD CPU speculation... and expert conjecture

[FALC0N]

When we asked AMD's CTO Joe Macri if a better solution would have been six CPU cores and fewer GPU cores, the answer was a resolute 'no'. Macri said that very little software is able to take advantage of more than four cores. The software that does work well with multiple cores, such as photo and video editing programs, works even better with GPU cores, since algorithms can run even faster and more efficient. Macri does have a good point, but as long as all multi-threaded software hasn't been optimised to efficiently use GPUs, AMD will keep lagging behind Intel.

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.

Seriously, I've been arguing this same exact point FOR FOUR YEARS, been constantly attacked for doing so, and now the AMD CTO agrees with me. VINDICATION!

Also explains why no PD refresh; AMD views it as a dead-end arch.

AMD's CTO agreeing with you isn't vindication. Its the reality that they have fallen years behind intel on the cpu side. They have little choice at this point but to try and change the battlefield and leverage their GPU assets.

If by PD refresh you mean AM3+ refresh, your conclusion is not in evidence. One thing this launch has shown us is that the 28nm half node was not well suited for an enthusiast grade part. Perhaps they are just waiting for the full shrink.

 

[de5_Roy]

Yes and no. The vast majority of "memory benchmarks" are actually cache benchmarks due to the front end prefetcher.

Spoiler

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 don't know what programming language sandra uses, but wasn't ags1's benchmark written in java? does that affect which components the benchmark taxes? i assume sandra's coders know which part of the cpu/apu (vendor-specific, in instances?) to aim for like all synthetic benchmark programs.

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.

homework, cool . will get on that asap. anyway, 4th alu in ~3.5ghz dual core cpus sounds like overkill... does that mean haswell's core i3's htt performance is as strong as amd's cmt?

 

[Cazalan]

I found this rather interesting. AT was able to OC from 3.7Ghz to 4.2Ghz without even bumping the Vcore.

That's pretty rare isn't it?

 

[$hawn]

I found this rather interesting. AT was able to OC from 3.7Ghz to 4.2Ghz without even bumping the Vcore.

That's pretty rare isn't it?

Looks like 4.2-4.3 GHz is the freq at which this particular chip's power draw starts to rise exponentially….any higher than this and it’ll start drawing astronomical amounts of power.
This is lower than PD on 32nm if I’m correct?

 

[noob2222]

Is toms doing a kaveri review or did Chris blow it up?

 

[de5_Roy]

Is toms doing a kaveri review or did Chris blow it up?

incidentally, anand didn't do 7850k review. i wonder if they were playing with an iris pro pc and oculus rift...

 

[Cazalan]

Is toms doing a kaveri review or did Chris blow it up?

incidentally, anand didn't do 7850k review. i wonder if they were playing with an iris pro pc and oculus rift...

Check again. That's where the chart above came from.

Still waiting for a 2nd round of reviews that goes more in depth (Tom's hopefully). Some off comments I've been reading indicate higher minimum fps increases. So while the peak/avg fps might not be going up a lot the gameplay would be more fluid.

 

[Ags1]

Yes and no. The vast majority of "memory benchmarks" are actually cache benchmarks due to the front end prefetcher.

Spoiler

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 don't know what programming language sandra uses, but wasn't ags1's benchmark written in java? does that affect which components the benchmark taxes? i assume sandra's coders know which part of the cpu/apu (vendor-specific, in instances?) to aim for like all synthetic benchmark programs.

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.

homework, cool . will get on that asap. anyway, 4th alu in ~3.5ghz dual core cpus sounds like overkill... does that mean haswell's core i3's htt performance is as strong as amd's cmt?

Yes Headline Benchmark was designed to maximize the benefit of caching - it copies the same piece of data repeatedly, so the likelihood of a cache hit is maximized. The only way to make it more cache-friendly would be to lock the thread to the same core to avoid core hopping, but that's a "no can do" for Java. Still it produces curves that usually correspond to the cache size per CPU, and matches the expected cache performance of eg FX versus Core i. I'd like to see what happens on an Iris Pro system as that breaks the test a bit - I copy blocks up to 32GB thinking that would be enough to exceed the size of any consumer CPU cache, but then Intel came out with 100MB L4 cache... :-(

 

[de5_Roy]

Check again. That's where the chart above came from.

Still waiting for a 2nd round of reviews that goes more in depth (Tom's hopefully). Some off comments I've been reading indicate higher minimum fps increases. So while the peak/avg fps might not be going up a lot the gameplay would be more fluid.

i don't see a second article. the at review was done by ian cutress and rahul garg.
i assume chris' iris pro article, kaveri review ...scratch that, the review just came out (while i was typing the reply)
http://www.tomshardware.com/reviews/a10-7850k-a8-7600-kaveri,3725.html

 

[colinp]

I'd also like to see some in depth reviews and figures. We haven't seen any temp charts yet, have we? Also, some analysis on the frequencies the higher end parts manage to maintain. I'm wondering if there is some throttling going on due to thermal issues. All those transistors densly packed together on a bulk process may be harming performance a little, but hum away quite happily on the lower-clocked A8.

Ultimately, you have a processor architecture that was designed to hit high frequencies on SOI, but now that is no longer an option. I think AMD need to go back to the drawing board and design a proc from the ground up to work at lower frequencies on bulk. Wait, they already have - Jaguar. Any reason that can't be made to work at higher TDPs?

 

[juanrga]

My apologies anand and a few others are saying GF also. Now i have to find a way to blame it on juan. J/k

I have always said that Kaveri is made on 28nm bulk process at GlobalFoundries. You will have to find another person to blame ;-)

 

[Reynod]

ok guys enough mud slinging now.

In fact in the last 15 years I only know of one prediction in a microprocessor release well ahead of the fact that exceeded expectations ... and that was the core2 ... based on the IPC from it's early mobile predecessor.

10 points for naming the part and the article.

 

[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.

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.

In Anandtech's review, 7850k failed to beat an HD 6750. Granted the CPU was a 4770k, but the 6750 beat out the 7850k at higher resolutions even by just the slightest amount. HD 7750 is definitely superior to 6750. For awhile yesterday there was an HD 7770 on sale for the price of a 7750. Keep eye out on combo deals and sales and can easily outmatch a 7850k for the same amount of money.

 

[-Fran-]

Well, Toms put out a little dual graphics round up and the results (with frame pacing enabled) look quite promising.

Maybe, just maybe, the A10 7850K could justify its price point with a cheap R7 240 for more computing (ergo graphics) muscle. I wonder if the R7 260 can work as well in dual graphics. Well, the higher the graphics fidelity, the better the CPU has to be, I guess.

Cheers!

 

[de5_Roy]

i didn't see any temperature measurements, throttling tests at stock in any of the kaveri reivews i read so far. i guess writers and editors are tired from the recent events. detailed analyses on how much kaveri can be pushed on mainstream 3rd party cooling should be interesting.

 

[8350rocks]

I found this rather interesting. AT was able to OC from 3.7Ghz to 4.2Ghz without even bumping the Vcore.

That's pretty rare isn't it?

My 8350 runs @ 4.4 GHz on stock vcore of 1.35...

AMD has been particularly good at accomplishing such feats if you get a good batch of silicon.

However, with bulk silicon and higher density, the leakage will be off the charts and power consumption will rise dramatically once you get out of the "sweet spot". We saw this evidenced with PD architecture, however, the sweet spot was significantly higher, in the realm of 4.6-4.8 on the high end depending on your silicon due to PD-SOIs leakage tolerances. I could likely push my chip harder, but I prefer to keep my case running cooler without going to a more elaborate cooling setup.

 

[colinp]

Another thing to consider: Are they using their RCM tech, and is it tuned for lower frequencies?

 

[de5_Roy]

Another thing to consider: Are they using their RCM tech, and is it tuned for lower frequencies?

good point! i wonder what happened to rcm. after all the hype about 30% less power use, they went silent on trinity release and nothing ever since. meanwhile glofo did claim that their 28nm process has adopted rcm. that's too vague.
the current sr-b and kaveri is really a laptop uarch with hsa enhancements scaled up to desktop use. i credit amd and their revamped power gating for this for the most part. that said, i'd also like to see some o.c. and how their new power management tech works.

one thing is quite clear though, what amd Really pitching here is not kaveri, not steamroller, But, HSA. they're trying to drive the point that HSA has a physical form now, and it is a major achievement for the betterment of pckind and the software ecosystem is coming. meanwhile enjoy your low cost apu (which will take advantage of HSA when it gets here).

all of the reviews lack in depth technical and hardware insights like haswell launch or bulldozer (zambezi and vishera). the most technical info was in tom's review and in s/a's. at focussed more on hsa. that was it. and then a8 7600 stole the show, lol.

 

[griptwister]

I hope we see some Excav for AM3+/AM4...

 

[lilcinw]

I agree that the A8 7600 is the star of this release, especially at 45 watts. It retains a remarkable amount of gaming prowess for shedding 50 watts compared to the 7850K. Tom's review claims that Kaveri was optimized for ~45 watts and the reviews seem to confirm this.

It would be interesting to build a silent HTPC like palladin likes to advocate for around one. Too bad it isn't available yet.

 

[Cazalan]

Another thing to consider: Are they using their RCM tech, and is it tuned for lower frequencies?

When a company doesn't update their website in 2 years that just looks odd. Not even a part time web editor to bump Copyright dates?

http://www.cyclos-semi.com/company/history.html

There was a flood of announcements/articles in 2012 and near silence since then.

 

[jdwii]

Is it me but is the A10 7850K slower as much as its faster in CPU tasks?

 

[Ags1]

I have had a AMD A10-7850K APU result submitted to headline benchmark. Here it is compared to an FX6300. Both at stock, although the FX6300 is watercooled:

http://www.headline-benchmark.com/results/24e1ef3f-be69-4503-9310-1cb3e1222523/38b1dd57-0ee2-4849-8278-82073ce722ca

Looks like caching got a bit better this time around...

Anyone got richland at stock to compare the 7850K to?

 

[-Fran-]

I have had a AMD A10-7850K APU result submitted to headline benchmark. Here it is compared to an FX6300. Both at stock, although the FX6300 is watercooled:

http://www.headline-benchmark.com/results/24e1ef3f-be69-4503-9310-1cb3e1222523/38b1dd57-0ee2-4849-8278-82073ce722ca

Looks like caching got a bit better this time around...

Anyone got richland at stock to compare the 7850K to?

Now that you mention it, I'll submit the A8-3850 testing. I totally forgot about it, lol.

Cheers!

EDIT: http://www.headline-benchmark.com/results/24e1ef3f-be69-4503-9310-1cb3e1222523/e286f2a7-a169-4663-91d6-96fec96cacb6

 

[blackkstar]

I really wonder how this would have turned out if it were a 3m/6c on 22nm SOI.