AMD CPU speculation... and expert conjecture

[airborn824]

AMD's best core is *probably* Piledriver. It has per-core, per-clock performance roughly on par with the Deneb core you have but there are more cores and they are more highly clocked. Whether or not you see a jump in buying an FX-8350 vs. keeping your 965BE mainly depends on if your 965BE is really a limiting factor in what you do. You won't see a huge difference in gaming but if you do video processing you will get a good bump from the extra threads in the 8350.

The most powerful x86 core out there is Intel's Ivy Bridge and it is approximately 15-20% faster clock for clock, core for core than Piledriver. However YMMV a lot since performance depends heavily on if you get a fully-functional core with HyperThreading on vs. off, how much the CPU is able to use Turbo Boost, and what exact programs you are running (e.g. how highly optimized they are for Intel CPUs or not.)

Short answer: Bang for the buck is going to be AMD's latest FX series. In general, nothing Intel has below the i7-3930K is going to be significantly faster than AMD's FX series. Intel is quite a bit faster with their 6-core chips but they charge dearly for them.

Thank you, seems we should wait for a CPU that supports more DDR channels and express 3. maybe steamroller. i am a avid AMD user, mostly because of price to performance and being a business major i am not a fond of the deeds Intel has done to get where they are.

 

[gamerk316]

http://www.guru3d.com/articles_pages/amd_a10_5800k_review_apu,15.html

These APU's are getting faster and stronger. What I found with my benches is the absence of L3 cache hasn't had a profound hit on CPU performance as most believe. The IMC in the APU's along with its L2 is faster than that on Thuban, Zambezi and Vishera. The addition of L3 would help performance by a small margin but at the cost of power and latency so the faster IMC and cache actually mitigates its performance loss.

Duh? L3 isn't THAT much faster then needing to access main memory, and throw in the latency involved to access it, die space, and power draw, it simply doesn't make much sense. Never really thought it was that good a use of die space...

 

[anxiousinfusion]

With CeBIT starting today, what can we expect to be revealed by AMD?

 

[noob2222]

http://www.guru3d.com/articles_pages/amd_a10_5800k_review_apu,15.html

These APU's are getting faster and stronger. What I found with my benches is the absence of L3 cache hasn't had a profound hit on CPU performance as most believe. The IMC in the APU's along with its L2 is faster than that on Thuban, Zambezi and Vishera. The addition of L3 would help performance by a small margin but at the cost of power and latency so the faster IMC and cache actually mitigates its performance loss.

Granted nobody buying this platform will use it with a 7970, that said we got every game in the suite playable at over 60FPS which for its price is some serious potency. The Athlon II x 4 740 represents the best value here, the 65w TPD at $55 with turbo boost of up to 500mhz and sufficient performance to frap record an entire BF3 64 man multiplayer round and maintain an average over 75FPS (min-35, avg-75,max-89) with a HD7970GE is nothing short of bang for buck.

http://www.anandtech.com/bench/Product/675?vs=700

Move down to the game bechmarks and you see where you need l3 cache. The 5800k is upwards of 20% slower at the same clock speed.

http://www.anandtech.com/bench/Product/122?vs=81
http://www.tomshardware.com/reviews/athlon-l3-cache,2416-6.html

Same with athlon II vs a slower clocked Phenom II, still 10% loss for l3 cache.

Sure, there are some games out there that could care less about L3 cache, but when you get one that does, you might as well shoot yourself in the foot for buying a cacheless cpu.

On the other side tho, office use and productivity is nearly identical, l3 or not doesn't matter if thats the intended use. This is where the APU is at its best, in office application usage, not as a dedicated gaming machine.

As far as main memory vs l3 cache ...

http://www.sisoftware.net/?d=qa&f=ben_mem_latency

40 clock cycles is dependent on cpu speed, but at 4.0 ghz thats 10 ns, 12.5ns at 3.2ghz, main memory is 65 ns (full random tests).
I am not sure where you get that main memory is just as fast as l3 cache when its 8.4ns to 75.6ns (nearly 10X slower?!!!!) or at best 13 clocks (3.9 ns) to 7.7 ns on sequential reads on the 2500k. Seeing the sequential times would explain why office applications don't really care, but on random access, your way off.

AMD definately needs to fix their 73 clk random l3 speed.

 

[gamerk316]

I am not sure where you get that main memory is just as fast as l3 cache when its 8.4ns to 75.6ns (nearly 10X slower?!!!!) or at best 13 clocks (3.9 ns) to 7.7 ns on sequential reads on the 2500k. Seeing the sequential times would explain why office applications don't really care, but on random access, your way off.

At least as of the Core 2 Generation, the rule of thumb was:
Data in CPU Register: 0 clock cycle latency
Data in L1 Cache: 1-2 clock cycle latency
Data in L2 Cache: ~20-30 clock cycle latency
Data in RAM: ~80-100 clock cycle latency
Data NOT in RAM: ~100k+ clock cycle latency [Page Faults are expensive!]

Based on what I've seen, L3 is about 60ish clock cycles latency to access. Anyway, in regards to sisofts tests, until I see for sure that they are not running into any pagefaults, my comment stands.

 

[gamerk316]

Also, since TressFX was brought up not too long ago, some early reports over at HardOCP have TressFX incurring a 30 FPS hit...on AMD Hardware.

 

[-Fran-]

Also, since TressFX was brought up not too long ago, some early reports over at HardOCP have TressFX incurring a 30 FPS hit...on AMD Hardware.

Then for Kepler cards it will be massive! And previous gen AMD cards as well (VLIW4 and 5).

Under that premise, I'm willing to bet that the GTX580 will be so near the GTX680 it will make no sense, hahaha.

Anyway, we had a very long L3 discussion a while back. I'm in the camp of "who cares about L3"

Cheers!

 

[griptwister]

This video makes me want to go out and buy a FX 8350 right now... But there is a slight chance steamy will make an appearance in Summer 2013... But those 990FX motherboards are so cheap right now... I think I'll wait for Haswel benchmarks.

http://www.youtube.com/watch?v=eu8Sekdb-IE

 

[anxiousinfusion]

But there is a slight chance steamy will make an appearance in Summer 2013

Pretty much the number one reason I'm keeping up with this thread. I have unreasonable hopes that three AMD products will be released this summer;

1. Steamroller CPUs
2. Sea Islands GPUs
3. 1090FX chipset

Even if only one or two of these make it, I'll still be happy.

 

[palladin9479]

Anyway, we had a very long L3 discussion a while back. I'm in the camp of "who cares about L3"

L3 has always been one of those "if you have extra room, why not" kind of things. Of all the things on-die its providing the least performance for it's respective power / space consumption.

I've explained before that L3 was kind of the "oh sh!t" method of last resort before your forced to stall for main memory access. It's only there for when your prefetcher fails miserably and the data isn't inside L2.

That being said, SB/IB utilizes L3 a bit more then PD/BD do. Intel went with a very small L2 cache size of 256KB which is tiny by today's standards of 512~1MB. They tightly coupled that L2 cache to give it the lowest latency possible and rely on L3 to catch anything that couldn't be fit inside it's small L2.

So yeah APU's not having L3 isn't a bit deal, they have 2MB shared L2 per module. The on board GPU was far more important then the L3 cache for it's market value.

 

[jdwii]

Pretty much the number one reason I'm keeping up with this thread. I have unreasonable hopes that three AMD products will be released this summer;

1. Steamroller CPUs
2. Sea Islands GPUs
3. 1090FX chipset

Even if only one or two of these make it, I'll still be happy.

Amd has pretty much stated none of this will be out this year Steamroller non APU's will be out next year with the 1090FX(if that's what they call it) 8000HD series wont be out until 4th quarter of this year or even next year(most likley before steamroller AM3+ CPU's) if i remember correctly. Richland will be out the second quarter of this year maybe even third quarter since its AMD.

 

[palladin9479]

This video makes me want to go out and buy a FX 8350 right now... But there is a slight chance steamy will make an appearance in Summer 2013... But those 990FX motherboards are so cheap right now... I think I'll wait for Haswel benchmarks.

http://www.youtube.com/watch?v=eu8Sekdb-IE

Interesting...

People can speculate on both sides but he had a very good point about streaming while playing. That's become all the rage these days with how interconnected social media has become. Also goes back to what I've stated in the past about needing to benchmark multiple heavy applications running at once, namely a game along with some sort of rendering software.

 

[griptwister]

Well, If this MSI MoBo is any good... I might just forget an AMD build all together and jump to a i7. I really hope AMD will make me regret my decision though.

 

[de5_Roy]

Amd has pretty much stated none of this will be out this year Steamroller non APU's will be out next year with the 1090FX(if that's what they call it) 8000HD series wont be out until 4th quarter of this year or even next year(most likley before steamroller AM3+ CPU's) if i remember correctly. Richland will be out the second quarter of this year maybe even third quarter since its AMD.

1090 is Long overdue. amd should decide if they're gonna continue with am3+ or make a new socket(with nothing new to offer) and introduce 1090fx on the new socket.
rumor is, there might not be an 8k series. looks like amd will release new gcn cards first, with new specs.
you can always count on glofo to be late.

Interesting...

People can speculate on both sides but he had a very good point about streaming while playing. That's become all the rage these days with how interconnected social media has become. Also goes back to what I've stated in the past about needing to benchmark multiple heavy applications running at once, namely a game along with some sort of rendering software.

how about this one below -
http://techreport.com/review/23750/amd-fx-8350-processor-reviewed/9

 

[palladin9479]

1090 is Long overdue. amd should decide if they're gonna continue with am3+ or make a new socket(with nothing new to offer) and introduce 1090fx on the new socket.
rumor is, there might not be an 8k series. looks like amd will release new gcn cards first, with new specs.
you can always count on glofo to be late.

how about this one below -
http://techreport.com/review/23750/amd-fx-8350-processor-reviewed/9

--Yeah-- I'd have to look at exactly what their doing cause there is no way in hell a four core i5 should do better then an 8 core fx8350 with that many threads. Unless the "Movie Encoder" was single threaded. In that scenario it would be a load of 3~4 cores which explains much. I don't use Windows Live Movie Encoder (many better options freely available) so I can't really speak about what it's doing.

 

[Cazalan]

Pretty much the number one reason I'm keeping up with this thread. I have unreasonable hopes that three AMD products will be released this summer;

1. Steamroller CPUs
2. Sea Islands GPUs
3. 1090FX chipset

Even if only one or two of these make it, I'll still be happy.

There are 3 products due this summer.

1) Richland (due slightly before the others)
2) Temash
3) Kabini

Steamroller (Kaveri) is Q4 at the earliest. Unknown why it's delayed so long past the other 28nm parts but it is a much bigger chip.

Really curious as to how well the 28nm (TSMC) CPUs will overclock compared to the 32nm (GF).

 

[sarinaide]

http://www.anandtech.com/bench/Product/675?vs=700

Move down to the game bechmarks and you see where you need l3 cache. The 5800k is upwards of 20% slower at the same clock speed.

http://www.anandtech.com/bench/Product/122?vs=81
http://www.tomshardware.com/reviews/athlon-l3-cache,2416-6.html

Same with athlon II vs a slower clocked Phenom II, still 10% loss for l3 cache.

Sure, there are some games out there that could care less about L3 cache, but when you get one that does, you might as well shoot yourself in the foot for buying a cacheless cpu.

On the other side tho, office use and productivity is nearly identical, l3 or not doesn't matter if thats the intended use. This is where the APU is at its best, in office application usage, not as a dedicated gaming machine.

As far as main memory vs l3 cache ...

http://www.sisoftware.net/?d=qa&f=ben_mem_latency

40 clock cycles is dependent on cpu speed, but at 4.0 ghz thats 10 ns, 12.5ns at 3.2ghz, main memory is 65 ns (full random tests).
I am not sure where you get that main memory is just as fast as l3 cache when its 8.4ns to 75.6ns (nearly 10X slower?!!!!) or at best 13 clocks (3.9 ns) to 7.7 ns on sequential reads on the 2500k. Seeing the sequential times would explain why office applications don't really care, but on random access, your way off.

AMD definately needs to fix their 73 clk random l3 speed.

I think most of us here have already stated that the 5-10 FPS benefit that L3 may sometimes give is not worth the power consumed and die space used up, there are better means to mitigate that.

The Athlon 750K is faster than the stars based Athlons and uses less power, the L2 is also much faster with less latency than the former Athlon II's. The 750K nets similar performance with a discrete Graphics card as what the old P2 x 4 955/965 and the FX4000 series does within a few percentiles.

What we all will agree on here is that AMD has a lot of work to do on cache latencies and IMC speeds...oh wait there is SteamRoller, 37% latency reductions and 40% faster IMC as the slides say. Lets just wait and see. Already leaks of Kaviri almost doubling up Trinity, the other leaks is that x86 single threaded performance is now actually respectable compared to intels sub $200 offerings which is fair enough.

 

[de5_Roy]

^^ your numbers are all over the place. 47% faster imc?!?! XD

 

[palladin9479]

^^ your numbers are all over the place. 47% faster imc?!?! XD

"47%" on IMC wouldn't mean +47% total performance, more like reduced latency / more simultaneous transactions or a bigger I/O buffer.

 

[de5_Roy]

^^ that looks more believable.

toms' crysis 3 cpu scaling bench:
http://www.tomshardware.com/reviews/crysis-3-performance-benchmark-gaming,3451-8.html

AMD Begins to Send Software Developers Next-Generation APUs.
AMD Starts to Provide Its Next-Gen Fusion APUs to Software Developers
http://www.xbitlabs.com/news/cpu/display/20130304190201_AMD_Begins_to_Send_Software_Developers_Next_Generation_APUs.html

AMD Redoing Radeon HD 7990 Under New Codename - "Malta"
http://www.techpowerup.com/180923/AMD-Redoing-Radeon-HD-7990-Under-New-Codename-quot-Malta-quot-.html

 

[sarinaide]

^^ your numbers are all over the place. 47% faster imc?!?! XD

37%, trying to work off a smartphone no putah and its killing me.

"47%" on IMC wouldn't mean +47% total performance, more like reduced latency / more simultaneous transactions or a bigger I/O buffer.

25% less miss predicts, thats about in margin of error with Intel's, l cache misses reduced 25% is also a big number generation to generation. 5-10% faster scheduling and 30% greater OPS.

It is very hard to predict how it will scale in real world but it will nontheless be faster than Vishera by some margin.

AMD has been vague on the 15% performance improvements, if it is in general computing across the board then Steamroller will represent a massive step forward, if it gets more then its a bonus. Still slower x86 compared with Intel but thats expected when Intel pumps 10:1 more resources into its microarchitecture. On Kaviri side, some leaks suggesting as about double Trinity's performance, considering Trinity is still capable of playing many games not included in suites at 16x10 and 19x10 at maxed or med-high settings. While its not apt for Crysis3 and the like what is certain is AMD's iGPU solution is improving, dual graphics is improving so I do agree with what AMD said, they are only a couple generations away from mainstream entry level gaming graphics on a chip.

 

[Cazalan]

With AMD's luck the switch from GF to TSMC will give the IPC gains they wanted but the process will limit them to ~3Ghz. The fastest TSMC 28nm chip I found was a 3.1Ghz ARM A9 test chip. A die that's tiny in comparison to the larger A10 style APUs.

 

[sarinaide]

With AMD's luck the switch from GF to TSMC will give the IPC gains they wanted but the process will limit them to ~3Ghz. The fastest TSMC 28nm chip I found was a 3.1Ghz ARM A9 test chip. A die that's tiny in comparison to the larger A10 style APUs.

AMD should guard against the smaller is better myopia, simple die shrink doesn't equate necessarily to improved performance, IB was not so much of a evolutionary step and Haswell from what is seen is hardly much more despite the 32-22nm jump. What Intel has done since Nahelem is improve the IMC significantly which has probably yielded the highest benefit to IPC's by affecting general performance the most rather than simple 10nm shrink. AMD's position is somewhat different, their fabs are sourced and limited resources prevent them from aggressively jumping the die sizes. Also AMD cannot gung ho to 22nm find that its not the lithology but rather the metal level architecture thats at fault.

 

[noob2222]

1090 is Long overdue. amd should decide if they're gonna continue with am3+ or make a new socket(with nothing new to offer) and introduce 1090fx on the new socket.
rumor is, there might not be an 8k series. looks like amd will release new gcn cards first, with new specs.
you can always count on glofo to be late.

how about this one below -
http://techreport.com/review/23750/amd-fx-8350-processor-reviewed/9

I love how TR tries to twist the actual numbers around. First off, they used skyrim, amd's worst game in their benches.

Intel pre-encoding 3570k, 104FPS
AMD 8350 pre-encoding, 77 fps

with endocing, Intel, 82fps, amd 62

both dropped by ~20%, so what does it prove? nothing really, pretty much every one of the cpus dropped close to 20% until you look at the dual core cpus, pentium g2120 and I5 655 hit the hardest, go from 78 fps to 37. Conclusion : dual core are not for multitasking at all.

other than that, take a small sample for a program that already heavily favors intel, it still favors intel if you try other things.

 

[sarinaide]

Noob, its because if they didn't bench market Intel, they would buy into it then the 10 Billion USD R&D fund would look like its being used to line pockets and hold many office parties because the 10:1 R&D differential between intel and AMD is clearly not showing.

Intel shot themselves shifting Haswell to a new socket, it made end users irrate when IB a side grade to SB came out with a new chipset with neglible performance gains despite the faster IMC and goodies thrown onto the new chipsets. I feel they are going to take the flack for releasing a new socket for a same process part which ironically is another side grade made more expensive by the platform change. Then there is the iGPU, now many sane reviewers told people to hold their horses and not expect a golden lolipop, well the hype over GT is going to hit the ground hard when it barely manages to compete with Llano. The ultimate realisation here is that Intel has its fails as well.