K8 vs K10 on 790FX - Results

[fotme]

informal: nehalem is not a core μarch, its a completely different proccess.
Penryn is a core μarch shrinked derivative.
Dont forget tick-tock model..

edit>@trueblue: lsdyna, mathematica, pam crash and fp in general are all memory intensive applications, hence imc has an advantage on that. However this is just a 10% of the pie, amd should have been improved to the 90%.

[tux08902]

Netburst->Conroe was a bit different cos Netburst sucked so much that it was easier to get huge performance gains. K8 was much better than Netburst so it's not so easy to get that big performance gain. Though I agree that K10 isn't exactly a succes.

K10 can be a success, but AMD needs to go back to their roots. They need to come up with an impressive pricing scheme. They also need to drop plans for single core chips based on K10. It's completely unnecessary. Not to mention that even dual-core on K10 would be unnecessary at this point. Tri-core and quad-core is what they should offer. The tri-core is something that Intel can't compete with, and dual-core and tri-core would be equally pricey to manufacture and deliver, given that AMD stupidly went for a monolithic design rather than a modular one.

Regardless, K10 seems like a decent option for K8 users. I wouldn't mind getting one so long as the price is right and it works in my board.

[JumpingJack]

There have been a couple of really nice Phenom threads, this being one... kudos.

[JumpingJack]

still, I know an intel engineer and he's been saying for a long time now the increase of conroe-->Nehalem will be similar to that of netburst-->conroe. I know intel has an awesome amount of resources, but still, amd should easily be able to pull 20% across the board over k8 especially with the 4 years it took for them to design k10, and if they wanted that market share and actually be able to call their product an innovation, k10 should have been at least 15+% superior to conroe across the board, and a lot stronger in areas where amd shines like fpu

I disagree, looking over mAJORD data, pretty much in agreement with the reviews that did similar, we are looking at 4 to 18% IPC improvement over K8 -- This is healthy, the K10 design rev from that point of view is successful.

If you compared Conroe (Core 2) over Yonah (Core) you get the same kind of 4 to 20% ish type of improvements depending on the app... any design that achieves ~ 10-15% (on average) improvement IPC wise is a good. What made Conroe shine so much was not the improvement over the prior tech along it's same lineage, but the improvement over the segment product which was along a different lineage... then again, it wasn't too difficult to improve over Prescott.

[informal]

informal: nehalem is not a core μarch, its a completely different proccess.
Penryn is a core μarch shrinked derivative.
Dont forget tick-tock model..

edit>@trueblue: lsdyna, mathematica, pam crash and fp in general are all memory intensive applications, hence imc has an advantage on that. However this is just a 10% of the pie, amd should have been improved to the 90%.

You are wrong..It IS based on 4-issue Core2,but it will have improved core(especially SSE part;Hyperthreading included).The most radical change will be system arch. which will change from FSB based to IMC+QuickPath(pretty much like IMC+DirectConnect in K8&K10).So,yes,it is a "tock" but it's not radically new "core-wise",more radical "system arch.-wise".
Here you go:
http://aycu07.webshots.com/image/373...5101578_rs.jpg

Now can we stop this off topic discussion and pay more attention at the subject of this topic?
To me it seems that Phenom in this present form is very good (solid) improvement over K8 in ALU perf.(around 15%).In pure x87 code,there is no improvement(no surprise there),while in SSE intensive cases we can see some very good improvements.The things that bother Phenom the most at present are some bugs(will get fixed soon) and CnQ2 not properly running in some cases.I reckon that with B3 we could finally see some higher clocked nortbridges in Phenoms,since 2Ghz is fairly low and affects the L3 latency in a bad way.

[awdrifter]

If we look at the K10 as just a tweaked K8, the gains are not bad. But when compared to the competing Conroe and upcoming 45nm (don't know the code names), things doesn't look good.

[donitsi]

I reckon that with B3 we could finally see some higher clocked nortbridges in Phenoms,since 2Ghz is fairly low and affects the L3 latency in a bad way.

Yes i hope for higher clocks too. What do you think about the bugs, will rev b3 be enough? I think we would need at least b4 to get the phenom really "shine"

[RedBull78]

If we look at the K10 as just a tweaked K8, the gains are not bad. But when compared to the competing Conroe and upcoming 45nm (don't know the code names), things doesn't look good.

Just a tweaked K8? hahah I had to laugh. Do you really believe it took like 2-3 years to tweak something based on a same core? I doubt it. K10 will grow in time just be patient and you will see how great the application will respond to it that support multi-cpu.

[RPGWiZaRD]

Thanks for sharing mAJORD.

I agree with some others here, makes you wonder what AMD has been doing all this time, developing K10 or having a coffée break?

Not that I thought there would be much to improve from K8, but a say a ~15% clock for clock performance boost sure is a bit dissapointing as it means it's a bit behind C2D. AMD just needs to drop this old milked cow and start from scratch. At least consumers can look forward to cheap quads this time around.

[EnJoY]

Just a tweaked K8? hahah I had to laugh. Do you really believe it took like 2-3 years to tweak something based on a same core? I doubt it. K10 will grow in time just be patient and you will see how great the application will respond to it that support multi-cpu.

Dude, just look at the results. Single threaded performance is barely better than that of K8. The only reason performance gains of k10 over K8 are even in the double digits are due to having double the cores and more cache.

[informal]

Dude, just look at the results. Single threaded performance is barely better than that of K8. The only reason performance gains of k10 over K8 are even in the double digits are due to having double the cores and more cache.

ALU scores are actually ~15% better than on K8(per core,not multi core).This is really a good achievement,since Core2 over Core was also 15-20%.
SSE units and cache widths in K10 are doubled when compared to K8 and here you'll see a lot more than15% gain(provided you run vector SSE optimized application-Cinebench10 uses scalar SEE for instance,so not much gained there).
Like Jack said,15% better IPC-wise than K8 is a really good result!Or else you can call Core2 over Core a bad IPC improvement too..

[tux08902]

Dude, just look at the results. Single threaded performance is barely better than that of K8. The only reason performance gains of k10 over K8 are even in the double digits are due to having double the cores and more cache.

Care to post the results you're looking at.

[Omastar]

Give EnJoY a break; he's a RAM guy, not a CPU guy.

[RedBull78]

Dude, just look at the results. Single threaded performance is barely better than that of K8. The only reason performance gains of k10 over K8 are even in the double digits are due to having double the cores and more cache.

Give it time man, trust me. I still remember the day Tom's Hardware tested the A64 based on singel channel 754 pin platform and believed it was no match for P4.

But we all know what happend a few months later do we now?

I hope you get what im trying to say. AMD just doesn't go about its business like Intel does. Of course Intel is currently the leader when comes to CPU performance no doubt about that. But, you lack the understanding how AMD offers its products in the retail market. Give it time , that's all im saying. Wait for the 45nm revisions and you will see how well it compares to c2d. *(Not counting the nahalem) , now thats a different ball game which is way too early even for me to speculate.

[PetNorth]

I disagree, looking over mAJORD data, pretty much in agreement with the reviews that did similar, we are looking at 4 to 18% IPC improvement over K8 -- This is healthy, the K10 design rev from that point of view is successful.

If you compared Conroe (Core 2) over Yonah (Core) you get the same kind of 4 to 20% ish type of improvements depending on the app... any design that achieves ~ 10-15% (on average) improvement IPC wise is a good. What made Conroe shine so much was not the improvement over the prior tech along it's same lineage, but the improvement over the segment product which was along a different lineage... then again, it wasn't too difficult to improve over Prescott.

QFT.

And for the records. I'm sure all you guys agree A64 was a success, wasn't it?

Well, I've recolected benches results from two reviews (hardware.fr web, in this case). First, A64 launch (sep-2003, comparing FX-51 vs XP 3200+, both at 2.2ghz); second, Phenom launch (nov-2007, comparing Phenom 9600 X2, two cores vs. A64 X2 4400+, both at 2.3ghz):



Well, Phenom perfomance gain is just fine. K10 arch is OK, IMHO.

What's been the problem? two word: clock frequency.

AMD has failed with mhz, not arch per se (on the contrary, to me arch is fine and elegant), but clock frequency, for whatever reason, some L3 bug, or MB bug, or whatever. It limits right now to launch higher freqs. But it seems they will solve it soon (with B3). And additionally, it seems that Phenom will gain even some more % perf with it


BTW! great work mAJORD!

[Brother Esau]

[QUOTE=PetNorth;2598504]QFT.

And for the records. I'm sure all you guys agree A64 was a success, wasn't it?

Well, Phenom perfomance gain is just fine. K10 arch is OK, IMHO.

What's been the problem? two word: clock frequency.

AMD has failed with mhz, not arch per se (on the contrary, to me arch is fine and elegant), but clock frequency, for whatever reason, some L3 bug, or MB bug, or whatever. It limits right now to launch higher freqs. But it seems they will solve it soon (with B3). And additionally, it seems that Phenom will gain even some more % perf with it


Now thats really the crust of the matter now aint it? Everybody here goes by more is better and instant gratification theory and that just aint true in life in general

[accord99]

Well, Phenom perfomance gain is just fine. K10 arch is OK, IMHO.

What's been the problem? two word: clock frequency.

AMD has failed with mhz, not arch per se (on the contrary, to me arch is fine and elegant), but clock frequency, for whatever reason, some L3 bug, or MB bug, or whatever.

Architecture is definitely a problem, seeing as how as the E6600 and Q6600, for the same hardware.fr benchmarks have nearly the same performance advantage (even when clockspeed is accounted for) over Phenom as the Phenom has over the Athlon X2. So, AMD needs to clock at least 10-15% higher than Intel to be competitive but given the current power consumption, doesn't have the thermal headroom to do so even if the core can scale that high.

[Brother Esau]

With proper cooling I don't see the issue if thats the case As I am sure on Phase you can crack that whip and make that Biatch sing

[mAJORD]

QFT.

And for the records. I'm sure all you guys agree A64 was a success, wasn't it?

Well, I've recolected benches results from two reviews (hardware.fr web, in this case). First, A64 launch (sep-2003, comparing FX-51 vs XP 3200+, both at 2.2ghz); second, Phenom launch (nov-2007, comparing Phenom 9600 X2, two cores vs. A64 X2 4400+, both at 2.3ghz):



Well, Phenom perfomance gain is just fine. K10 arch is OK, IMHO.

What's been the problem? two word: clock frequency.

AMD has failed with mhz, not arch per se (on the contrary, to me arch is fine and elegant), but clock frequency, for whatever reason, some L3 bug, or MB bug, or whatever. It limits right now to launch higher freqs. But it seems they will solve it soon (with B3). And additionally, it seems that Phenom will gain even some more % perf with it


BTW! great work mAJORD!

Thanks, and great work yourself with that table. highlights the point's made.

It's important to note also that infact the gains from K8-K10 on average are probably MORE than K7-K8 if you look purely at CORE enhancments., esp those resulting in some huge (and much needed to compete with core) SSE performance boosts.

Te large gains in gaming + Archiving in the K7-K8 comparison are largly due to the move to IMC + hypertransport. Much as that was the focus of the A64 arch, native quad and enhancing SSE performance were the biggest priorities this round.



More benches to come today. Been running Wprime 1.55, X264 encoding bench, Pov-ray, Systemtool

[tux08902]

Architecture is definitely a problem, seeing as how as the E6600 and Q6600, for the same hardware.fr benchmarks have nearly the same performance advantage (even when clockspeed is accounted for) over Phenom as the Phenom has over the Athlon X2. So, AMD needs to clock at least 10-15% higher than Intel to be competitive but given the current power consumption, doesn't have the thermal headroom to do so even if the core can scale that high.

Architecture is not the problem. K10 is quite a bit more efficient than K8. In fact, it's equally as Core is than NetBurst. AMD needs to scale clock speeds the way they planned to. In my opinion, they should not have released Phenom X4 until they had ironed out production issues and errata.

[JumpingJack]

Architecture is not the problem. K10 is quite a bit more efficient than K8. In fact, it's equally as Core is than NetBurst. AMD needs to scale clock speeds the way they planned to. In my opinion, they should not have released Phenom X4 until they had ironed out production issues and errata.

I think Accord99's point was that the architectural revision did not yield a per clock advantage over the Core microarchitecture.

[saaya]

nice results! thanks for posting all thoe results

[Jacky]

Now I think:
The problem ain't that K10's gains IPC-gains over K8 are bad(the sole problem compared to K8 is the clockspeed) or that it isn't a good cpu-arch. However Intel's perfect execution and a good arch on a lower process node *is* a serious problem.
thank you for the results so far!

[Tony]

Increasing default HTT may have helped also, i see no reason why we could not have had 333 default HTT over 200...many are trying to test this as reports are saying it speeds up L3 thru put, the only issue is HTT over HT3 is very limited at this time.

[Lightman]

Increasing default HTT may have helped also, i see no reason why we could not have had 333 default HTT over 200...many are trying to test this as reports are saying it speeds up L3 thru put, the only issue is HTT over HT3 is very limited at this time.

Good point!

But even if we would want to set HTT to 333 using HT2.0 we can't do it through BIOS on AM2+ boards. I've tried on GA-MA790FX and everytime I set 1GHz link or belove = no POST