AMD Zambezi news, info, fans !

[mongoled]

If it turns out that IPC is lower, will terrance be reinstated with an aplology and JF be banned because he blatently lied to this community?

This is just hypothetical

Thanks................

[PerryR]

Something like AMD FX => Amazing Modern Design F-up eXtremely? Time will tell.

Here's hoping the design is different for desktop.....But, not much from what I've seen.

[undone]

To much annoying and frustrating comment behind this thread. I have stated before pls everyone don't only focus on benchmark, there's too much you should care about, guess why amd delayed and delaying, why those naming scheme and scheduling changing suddenly, if these chips on roadmap aren't even retail, why bother?

[halofanman]

I Predict that the 8150/8170 will be able to reach 5.3ghz on air with a decent cooler,
turbo core on the 8170 is 4.7GHz, and that is with a stock air cooler. (stock water-cooling is just a rumor)

[m411b]

Most of this may just be speculation and most of it may not! I'm sure most is not and most is! But one thing is for damn sure; The more I read this thread, the more, and more I am doubting AMD! I really hate to say it because I've been an AMD fan since 1998 and have NEVER purchased anything Intel(save for EEEPC), and am really not looking forward too.

I can see this going REALLY bad for AMD if they don't produce just short of a miracle..

[Pestilence]

I Predict that the 8150/8170 will be able to reach 5.3ghz on air with a decent cooler,
turbo core on the 8170 is 4.7GHz, and that is with a stock air cooler. (stock water-cooling is just a rumor)

Ahh that just takes me back to the netburst days. Yeah lets clock it higher and higher and to hell with IPC

[AKM]

With the Bulldozer arch being radically different it's perhaps not so surprising that performance improvements over Family 10h are very hit and miss.

[informal]

Still something doesn't add up. Before the Interlagos was "scaled down" clock-wise( from roughly 2.8Ghz base to 2.3Ghz base we will have now),integer and spec fp rates were listed at roughly 48% and 77% better than 2.3Ghz MC system. This would lead us to believe integer performance per core and clock was slightly better (~3%-counted in non-perfect scaling with more cores and more clock and 80% scaling AMD lists for integer cores) while FP performance was noticeably better,between 15 and 20% per clock at same core count.
The sisoft results for 6282SE suggest 2.5Ghz base clock and 14% better integer and 76% better fp throughput than 2P MC @ 2.5Ghz.
We are left with few possibilities:
1) 6282SE/6220 results are not real.(ie. they are fake) This may very well be the situation here .
2) 6282SE/6220 results are "kinda" real. The problem is the Turbo is not working. I'm personally leaning towards this. We have 2.5Ghz in both integer and multimedia tests(and for both platforms 2.5Ghz). This equates to 7-8% better per core result in integer and 32% better per core result in Multimedia. 80% scaling in integer throughput was used to calculate how much better per core is Bulldozer : 232/0.8/16=18.12Gops for 1 BD core @ 2.5Ghz Vs 202/12=16.83Gops for 1 MC core @ 2.5Ghz;for fp it was simpler, just divide the result by number of "cores" for both and you get to 32% faster for Bulldozer per core(or per FMAC if you like ). Note that 6220 posts almost identical results as 6282SE,of course adjusted for difference in core count and clock(2x less cores and 20% less clock speed).
3) Results are not fake per se,but they are not accurate(bios/platform/CPUs are not final). It means we can't use them as reference point for anything. This one may be true too.

Number 1 and 3 are the ones we might consider the most. Number 2 is there since the results are there,but hinges on hope that Turbo was disabled.

[AKM]

If it turns out that IPC is lower, will terrance be reinstated with an aplology and JF be banned because he blatently lied to this community?

This is just hypothetical

Thanks................

What if IPC will be lower in some cases and higher in others?

[xsecret]

Why not :

4) 6282SE/6220 results are real and close to shipping performances.

?

[radaja]

Why not :

4) 6282SE/6220 results are real and close to shipping performances.

?

that would just be bad,and i hope thats not the case

[Pestilence]

Found this over at H. B2 Stepping. If this is correct the 2500K should be on par with the 8150

Anyone with an X6 care to run Cinebench at 3.1Ghz for us for a comparo?

[xsecret]

X6 3.3 GHz : 5.8

And I can confirm I have the same score with a final platform.

[nex_73]

And I can confirm I have the same score with a final platform.

If you now have the processor, haven't you signed a NDA?
Hmmm..
Motherboard and what bios & microcode AGESA version?
What chipset drivers version?
What SB driver version?
OS?
HT Link/NB Freq?
Mem Freq and timings?
hehe, lots of questions there...

[BeepBeep2]

4% multithreaded IPC difference in Cinebench R11.5 vs Sandy Bridge.
3.6 Ghz FX-8150 vs 3.4 Ghz 2600K, they should be near equal. 4.2 Ghz Turbo vs 3.8 Ghz Turbo, AMD should win.

Some of the benchmarks look bad, I agree. Less than Thuban performance in Cinebench looks bad, I agree. Getting really close to Sandy Bridge though, quite impressive. The AMD processor has no IGP so I can see why they are priced a little lower than Intel 2500/2600K. For those of us with discrete GPUs, I (pray) hope we get a platform win.

[informal]

Why not :

4) 6282SE/6220 results are real and close to shipping performances.

?

Yes this could be true too,but it doesn't bode well with integer throughput- at all. If this is true then each Bulldozer core is 11% slower in single thread mode (without the scaling penalty!) and close to 40% slower per thread when you just count in pure performance per thread and per clock : 202Gops 24C MC @ 2.5Ghz Vs 232Gops 32C Interl. @ 3Ghz(2.5Ghz base and Turbo working=>3Ghz) ; 202/24=8.41Gops per core @ 2.5Ghz vs 232/32/1.2=6.04Gops per core @ 2.5Ghz; 8.41/6.04=1.4x or 40% slower. That's why I think Turbo wasn't enabled for integer tests at all. Floating point test won't trigger turbo anyway so it's relevant for that segment of performance (which excels versus MC by the way: 32% faster per core at the same clock with effectively having 50% less FPUs,on paper).

If the 6220 results didn't have Turbo enabled,then compared to this poor showing we would have different picture.

2P 6220 @ 3Ghz and no Turbo . Integer: 138Gops,MultiM. : 315Mpix/s (15% slower than what 6282SE effectively gets with same core count!) . 1P,which equates to Zambezi 8C @ 3Ghz without Turbo would get 138/1.95=71Gops and 315/1.95=161Mpix/s. (if we take 6282SE results we land at 180Mpix/s !).
Leaked Zambezi at 2.8Ghz (with Turbo to 3.3Ghz for all int cores?? ) got : 54.64Gops and 147Mpix/s. According to above ,this is 40% lower in integer benchmark than Opteron 1P @ 3Ghz if you count that Zambezi had Turbo on or 21% lower per clock if Zambezi's Turbo didn't kick in. For fp the difference is between 3 or 14 %,depending which Opteron you take for comparison (6220 or 6282SE).

Now for FX 8150 @3.6Ghz with 3.8Ghz (assumed) Turbo for all cores we would have : 71x3.8/3~=89Gops and 161x3.6/3=193Mpix/s ( or 216Mpix/s if you take 6282SE results as reference point).
For reference 1100T @ 3.3Ghz gets : 65Gops and 115Mpix/s. So if Opteron numbers are true (but ran without Turbo) 8C FX8150 should be : 89/65=1.37 or 37% faster than 1100T in multithreaded integer benchmark and 193/115=1.67 or 67% (or 216/115=1.87 or 87%) faster than 1100T in Multimedia(AVX/SSE) benchmark. Per core and per clock this boils down to : 11% faster in integer (80% scaling counted in for Bulldozer and adjusted for clock difference of 3.8Ghz Turbo for 8150 and 3.3Ghz def. for 1100T) and 15%/28% faster in AVX/SSE test ( depending which Opteron you take as reference,6220 or 6282SE).

So if Turbo was off for the opteron results, Bulldozer does look pretty good.For throughput benchmarks it is between 35% and >65% faster,depending if it's pure integer or FP/SIMD.

[Formula350]

i was looking at things like laptops where the price is already summed up based on all the specs and the market kinda dictates prices much more than individual component competition. and keep in mind that the top end it mentions dual gpu, so thats another 50-70$, then beep pointed out you missed the os for 100$ right there. bringing the total to $800+
things seem to be about 20% over those prices (which are system prices, not user needs upgrade by pc knowledgeable friend prices)

Have you been looking at Best Buy, or what? hahah

Anyways, the OS part is moot if you're talking about pre-built, especially laptops. So you can't factor that in at all.

I didn't look through every one, but a number of these Radeon models are ringing bells for me with being the ones integrated into the APU. I never did look at how my discovery panned out compared to what AMD really ended up releasing, but not everything matches up on there either. So, if it's dedicated or not is a toss up: Uber Long NewEgg Search Link

This is the only one that comes up as specifically stating it has a dedicated card though.

All but 2 of those aren't under $700 :\

With the Bulldozer arch being radically different it's perhaps not so surprising that performance improvements over Family 10h are very hit and miss.

Wonder if maybe the software-designed core layout might be partially to blame? I expect what you said to be pretty much the case though.

[Particle]

1100T @ 3.1 with turbo off yields 5.4 for me, but that's with the NB at 2.8

[Formula350]

If you now have the processor, haven't you signed a NDA?
Hmmm..
Motherboard and what bios & microcode AGESA version?
What ...

Depending on the specifics of the NDA, you are allowed with in the contract to say a certain amount. Specifics like everything you're asking, pretty much a no-no. Saying something like "Yea, production sample A vs Pre-production D yield the same results", well that's a bit vague and doesn't exactly reveal much. Not to say it ISN'T in the NDA as something that can't be talked about though. *shrug*

1100T @ 3.1 with turbo off yields 5.4 for me, but that's with the NB at 2.8

Which is where you'd want to do it at, along with DDR3-1866 if at all possible, too. I mean, if you really want to remove all variables.

[informal]

Well if Opteron 6220 results are true ,as xsecret claims,then C11.5 result should roughly follow the same route as the Sisoft's MM benchmark.I say roughly since I have no idea what is the ratio of memory and SIMD instructions in these tests. If it does have similar ratio,then instead of 5.24pts one should have 1.67x the result of 1100T if he would to run C11.5 on FX8150 @ 3.6Ghz.Or in numbers : 9.85pts.
9.85pts is dangerously reminiscent of this (early slide detailing Scorpius platform and Zambezi advantage over Thuban in 3 benchmarks;slide was from Dec 2010 and was pointing roughly at 10pts in C11.5 for 8C Zambezi @ unknown clock).

[Oese]

well, the problem is, one integer "core" might indeed be weaker then one thuban core. But one module might be stronger. taken into account the thubans are benched at faster clocks, they may have same ipc per clock overall 4 core vs 6 core..

assuming bd has 0.85x the ipc per "core", it would equal 1.53 (0.85x1.8 scaling factor) ipc per "module". now 1.53x4 = 6.12, so it might be a bit faster then a six core thuban clock for clock.

given it is designed for higher clocks, i could imagine this very well.

the thing is the "core count". bulldozer as an 8 core has less ipc, as an 4-core it has 50% more ipc. Finally, ipc doesnt come out of the blue but from execution units and decoder/prefetch efficiency. AMD should have launched it as a 4-core and all would be good...

only question remaining is then, when will the 0.85 and when will the 1.63 be used. I think, for desktops like for servers multithreaded workloads are importnat for "snappiness" and "smoothness", whereas benchmarks dont need that much flexibility..

[Formula350]

Well if Opteron 6220 results are true ,as xsecret claims,then C11.5 result should roughly follow the same route as the Sisoft's MM benchmark.I say roughly since I have no idea what is the ratio of memory and SIMD instructions in these tests. If it does have similar ratio,then instead of 5.24pts one should have 1.67x the result of 1100T if he would to run C11.5 on FX8150 @ 3.6Ghz.Or in numbers : 9.85pts.
9.85pts is dangerously reminiscent of this (early slide detailing Scorpius platform and Zambezi advantage over Thuban in 3 benchmarks;slide was from Dec 2010 and was pointing roughly at 10pts in C11.5 for 8C Zambezi @ unknown clock).

I think you rrrrrreally need to move on from comparing everything to the server variants. They might be mostly the same, but I suspect there is enough of a difference for them to perform rather far apart from each other.

I also suck at math, but that link says 1.5x greater than 1100T, which if you use XSecret's 5.8 score you get 8.7m not 9.85. Which goes right back to comparing server stuff, as now even comparing unknown systems brings in way too many variables I think, especially when it comes to BIOS versions between board makers.

[Oese]

Found this over at H. B2 Stepping. If this is correct the 2500K should be on par with the 8150

Anyone with an X6 care to run Cinebench at 3.1Ghz for us for a comparo?

i think here the turbo clock is shown?

[informal]

I think you rrrrrreally need to move on from comparing everything to the server variants. They might be mostly the same, but I suspect there is enough of a difference for them to perform rather far apart from each other.

I also suck at math, but that link says 1.5x greater than 1100T, which if you use XSecret's 5.8 score you get 8.7m not 9.85. Which goes right back to comparing server stuff, as now even comparing unknown systems brings in way too many variables I think, especially when it comes to BIOS versions between board makers.

No.The link says 1.78x higher than Thuban in render. Check again.It's the middle bar (yellow) one. It represents Render performance . You can see it being roughly 1.78x higher than 1100T's. Media performance (first bar from the left to right) is based on PCmark's subtest called TV and Movies(this one is mostly single threaded and in this one Zambezi is around 15% faster). Last bar is the one showing 1.5 or 50% faster and is based on 3dMark06's CPU subtest. It fits good since this subtest doesn't scale that well with cores,but scales good with clock. Just read the footnote,it's all there.
And yes,in that sisoft suite,8C Zambezi without turbo would perform nearly identical to 1P BD Opteron 8C without turbo and at the same clock. On this I based my math and double checked it. You can compare Lisbon (2P 6C Opteron and divided by 2x) scores to Thuban X6 at the same clock and see it for yourself.

[xsecret]

You should be aware that Zambezi Turbo mode is not that simple. There is 2 "main" levels of Turbo. For example, a 3.6 GHz base CPU can reach 3.9 GHz with all cores used as long as TDP remain under a specified value AND can reach 4.2 GHz in single core mode. So, depending on the usage on a 8-threads application, you can be at 3.6 GHz or 3.9 GHz.