AMD 65nm and 45nm

[coldpower27]

Originally posted by: customcoms
Some of you aren't even paying attention to what is going on with AM2 b/c of C2D; people are hitting 200-400mhz MORE with an AM2 processor equivlant to its S939 counterpart and a good motherboard (like the DFI Infinity and Lanparty which have just been released). Now, anything is possible with the switch to 65nm; the switch from 130nm to 90nm went far more smoothly for AMD than Intels switch to 65nm, but the latter is a harder transisition. I seriously doubt a huge performance increase from this switch (maybe on the order of 5% like the winchester and venice), but maybe higher clockspeeds will be attainable (like I said, AM2 is already oc'ing better). The extras silicon in revG is probably for 4x4 systems, and some tweaks like the venice SSE3 instruction set addition.

However, big picture, this isn't going to be able to regain the performance crown. It will probably keep AMD highly competitive on the low end for the time being (until we get more and cheaper C2D's and their mobo's). At the moment, the biggest cost appears to be ram and psu prices, at least in my book (ram is getting expensive, and some psu's prices doubled in the last month.

Just for reference, I was planning on building a C2D system as a "budget" rig for my family but that might change with cheaper/better boards and cheaper processors available on AM2.

The Socket AM2 processors are of lower transistor density then the equivalent Socket 939 processor, that partially explains why they can clock higher, not to mention they are fairly late in the life of the 90nm process.

Windsor-512 = 183mm2, while the Manchester Core is 154mm2. They have roughly the same transistor counts.

Going to 65nm will likely reduce power consumption, but how much of an increase in headroom at this point is unknown. Though as long as AMD has a better price then Intel then they are doing allright.

Considering AMD's 130nm to 90nm shrink was one of the most boring, as it was only an optical shrink and on lower binned SKU's while Intel had to make their 90nm process give at least the same headroom or more from the get go.

The 65nm transistion is looking to be more difficult for AMD then 90nm if the rumors of tweaks to the core are accurate but we will see, it's a little more hten an optical shrink.

 

[coldpower27]

Originally posted by: jiffylube1024

Originally posted by: Frackal
I don't think so. He's saying a C2D performs at an average of Clockspeed x1.2 that of K8, and I'm saying that's a very lowball estimate. It is at least +30% faster per clock vs the X2's.

For instance, from AT's Conroe review:

CPU Bound gaming performance, x6800 vs. FX 62:


Advantage (Intel) 22.8% 55.0% 43.9% 30.9%


Not one bench is even at 20%. The average across the board shows that it is nearly 40% faster, and the x6800 is only slightly higher clocked than the FX-62

You've also gotta factor all the things the X6800 has going for it: ginormous cache, high FSB and a tad more clockspeed.

While I'd agree that the 4MB L2 Conroes are in the order of ~20%+ faster than similarly clocked AMD's with dual 1MB cache, it will be interesting to see if 65nm AMD can close this gap or at least narrow it significantly.

Actually the high FSB isn't much of an advantage considering it barely has barely half the memory bandwidth available compared to the Socket AM2 memroy bandwidth figures. Real world wise vs theoretical figures.

I don't think adding a ton more cache to AMD will help with their IMC advantage, as hitting main memory won't hurt them that much. It also looks like there will be no additional increase in cache for 65nm they will remain at 2x512KB and 2x1MB. It's better to reserve the larger 2MB of LV3 for the 65nm K8L, whenever that maybe coming down the road.

Well the 4MB of Shared Cache, is part of Intel's arsenal, if your going along that line of thinking, we can call out that it is a bit disappointing AMD is behind even with their IMC advantage.

However, Conroe XE doing well in gaming is not surprising considering it's very powerful in integer related tasks.

 

[Polish3d]

Originally posted by: inspire

Originally posted by: Frackal

Originally posted by: Viditor

Originally posted by: Frackal

Originally posted by: harpoon84
Viditor - You can't compare 65nm Conroe to 65nm K8 - different architectures, different clockspeed ceilings.
Look how much 130nm to 90nm gained for AMD over 3 years - a mere 800MHz.

I think 65nm K8 is just a dress rehearsal for the real 'next big thing' - K8L.

Even a 3GHz K8 will not be competitive with the top end C2Ds. As a rough rule, C2D x 1.2 = eqv K8

E6600 = 2.88GHz K8
E6700 = 3.2GHz K8
E6800 = 3.51GHz K8

A 65nm K8 vs Conroe is about as exciting as a 65nm Presler P-D vs AMD X2 - ie. not very.

Bring on K8L!

No way, it's more than that. It's more like +30% average at a minmum. At the LOW end its usually 20% faster per clock. I'm sick of people lowballing this kind of thing. Hell, in some games its ~50% faster per clock, and in SuperPi it's 60% faster

Frackal, I think you're confusing performance with clockspeed. We aren't discussing performance in this...

I don't think so. He's saying a C2D performs at an average of Clockspeed x1.2 that of K8, and I'm saying that's a very lowball estimate. It is at least +30% faster per clock vs the X2's.

For instance, from AT's Conroe review:

CPU Bound gaming performance, x6800 vs. FX 62:


Advantage (Intel) 22.8% 55.0% 43.9% 30.9%


Not one bench is even at 20%. The average across the board shows that it is nearly 40% faster, and the x6800 is only slightly higher clocked than the FX-62

Actually, no. 20% is a good estimate if you take all the individual benchmarks done in that AT review and look at the relative difference in performance between the FX-62 and the X6800 and weight them evenly to compute the average. I don't know how you come up with "Not one bench is even at 20%" - in the General Performance PC Worldbench 5 bench on Nero, the FX-62 actually posted a gain over the X6800 (2.3%).

I have a spreadsheet I made that incorporates all the benches from the AT review from a couple months ago when harpoon84 and I got into a debate about this exact same issue. I uploaded it somewhere, but it's been so long that I don't have the link anymore...

I was speaking specifically of the game benchmarks with respect to "not one is at 20%"


And as far as the PC WorldBench goes, is that purely a CPU limited benchmark? I suspect not, so it is not measuring a pure clock for clock difference.