Bulldozers first screens

[Lanek]

Can windows 7 even be installed on a Pentium?

and what is your point ? he speak about the result .... see a bit of sarcarsm ( the system of his mom who is still under win 98 will match this score ) ... , i have point in other post my 2600K with 1 single core test made 2.04pnts under Cinebench ... why a 6 cores Bulldozer will do 2.86 pnts when a Phenom 1095T do 5.73 pnts ....... you don't think the scores mess somewhere ? AMD have release a completely new architecture who are 2 times slower of the old ?

[madcho]

if the R11.5 scores are under 800mhz, with only 6 score, that's just amazing.

Else it could be fake ... or an alpha bugged cpu ... who knows.

[informal]

Core 1
2x128b AVX or 2x128b un-recompiled SSE
or
128-bit FP command

Only one legacy 128-bit FP FMAC per Thread

Solus Corvus already corrected you(after I did even before that). You need to read that blog post again.The whole point of co-processor model is that one core can have the whole FPU(2xFMAC) in case of second core having no fp instructions scheduled.Yes ,even in legacy code and that's the beauty of AMD's approach.

As for the benchmarks from chiphell,this is what Charlie said at SA:

Ax steppings had serious problems. B0 might have too, but that was supposed to fix many. If it is pre-B1, I would not count on it to be a reliable benchmark.

-Charlie

[w0mbat]

It's not 800Mhz, it's not 2800Mhz

where do u got that info from?
cause a butterfly told me otherwise.

[kl0012]

and what is your point ? he speak about the result .... see a bit of sarcarsm ( the system of his mom who is still under win 98 will match this score ) ... , i have point in other post my 2600K with 1 single core test made 2.04pnts under Cinebench ... why a 6 cores Bulldozer will do 2.86 pnts when a Phenom 1095T do 5.73 pnts ....... you don't think the scores mess somewhere ? AMD have release a completely new architecture who are 2 times slower of the old ?

In fact, Bulldozer's module max floating point instruction throughput on current software (without AVX and FMA) is equal to thus of one Phenom's core - two 128-bit fp ops per cycle. Bulldozer module is more flexible - it can start any combination of ops per cycle (such as MUL+MUL or ADD+ADD) while Phenom core is tied to MUL+ADD. On the other hand Bulldozer has higher latencies for fp ops and various FP-pack/blend/copy ops are executed on one of two fp-pipes while Phenom has special unit (fp-misc) for such type of instructions. So it is possible that 6-core bulldozer will have equal performance to 3-core Phenom on the same freq in apps with many fpu code.

[informal]

In fact, Bulldozer's module max floating point instruction throughput on current software (without AVX and FMA) is equal to thus of one Phenom's core - two 128-bit fp ops per cycle. Bulldozer module is more flexible - it can start any combination of ops per cycle (such as MUL+MUL or ADD+ADD) while Phenom core is tied to MUL+ADD. On the other hand Bulldozer has higher latencies for fp ops and various FP-pack/blend/copy ops are executed on one of two fp-pipes while Phenom has special unit (fp-misc) for such type of instructions. So it is possible that 6-core bulldozer will have equal performance to 3-core Phenom on the same freq in apps with many fpu code.

My prediction is that one FMAC will have around 20-30% higher performance than one Thuban core ,in non recompiled MT software.Single thread fp performance should be a lot higher than that(2xFMAC in this case).In FMA optimized code, there should be substantial jump,maybe up to 50%.
You can see from leaked donanimhaber slide that 8 core (probably <3.5Ghz) model has approx. 1.88x the performance of 1100T in Cinebench 11.5.That's non recompiled legacy fp workload in which you have 8 128bit FMACs working versus 6 Thuban cores(each of which is Mul+Add). This roughly corresponds to 1.3x the fp power of Tuban core,roughly at the same clock.

edit: Someone asked about stepping or revision of BD ES in question. It has W8K44 at the end so it is a B0 for sure. Since Charlie wrote pre-B1 was useless for benchmarking you can now see why the scores are the way they are.

[-Boris-]

My prediction is that one FMAC will have around 20-30% higher performance than one Thuban core ,in non recompiled MT software.Single thread fp performance should be a lot higher than that(2xFMAC in this case).In FMA optimized code, there should be substantial jump,maybe up to 50%.
You can see from leaked donanimhaber slide that 8 core (probably <3.5Ghz) model has approx. 1.88x the performance of 1100T in Cinebench 11.5.That's non recompiled legacy fp workload in which you have 8 128bit FMACs working versus 6 Thuban cores(each of which is Mul+Add). This roughly corresponds to 1.3x the fp power of Tuban core,roughly at the same clock.

But has that Donanimhaber slide any credibility? I see no reason why it should be considered real.

[w0mbat]

@informal: yes, afaik pre-B1 has a multiplier issue.

[informal]

But has that Donanimhaber slide any credibility? I see no reason why it should be considered real.

I personally think it's real.

[Dresdenboy]

somehwere between 800 and 2800Mhz?

It's clearly not 800Mhz.. Performance would never be that good.

The lowest frequency P-State can be be programmed to be as low as 500 MHz.

[kl0012]

My prediction is that one FMAC will have around 20-30% higher performance than one Thuban core ,in non recompiled MT software.Single thread fp performance should be a lot higher than that(2xFMAC in this case).In FMA optimized code, there should be substantial jump,maybe up to 50%.
You can see from leaked donanimhaber slide that 8 core (probably <3.5Ghz) model has approx. 1.88x the performance of 1100T in Cinebench 11.5.That's non recompiled legacy fp workload in which you have 8 128bit FMACs working versus 6 Thuban cores(each of which is Mul+Add). This roughly corresponds to 1.3x the fp power of Tuban core,roughly at the same clock.
.

Especially I did not want to connect my post to any leaked data or benchmarks. The question was whether it is possible that 6-core Buldozer is slower then 6-core phenom in some cases. Based on data published by AMD - yes, it's possible.

[-Boris-]

I personally think it's real.

Well I can't say I have any proof for the opposite. But I think we should be careful when using the numbers from that slide.
I do believe though that the decoupled FPU will boost FPU performance quite good at 1-4 threads. I think the extra FPU power in Kuma compared to Brisbane and Windsor is the main reason it performed 20-30% better in games and Cinebench. I don't know if it will be as big difference this time but it will help alot.
That combined with turbo is the main reason I think BD won't suffer much from games being poorly threaded.

EDIT: Link to my claims about Kuma performance if anyone doubt my word for it.
http://www.bit-tech.net/hardware/cpu...n-cpu-review/5

[mAJORD]

where do u got that info from?
cause a butterfly told me otherwise.

Well, I'll answer with a question. do you really think that's 800Mhz?

I'm not objecting to the notion of sampels getting stuck in CnQ, but the numbers show it's a bit more complicated than that.

Maybe the windows thread-throwing act is playing havoc with these B0 (apparently) samples with CnQ

As a result of whatever's going on I feel projecting actual performance nothing short of Impossible.. You certainly can't just assume it's a solid 800Mhz and start multiplying scores

[Hornet331]

Better FP performance won't lead to faster games, since they are usually INT heavy, but other improvments as better/faster cache/branch predictors do wonders in games.

[informal]

Better FP performance won't lead to faster games, since they are usually INT heavy, but other improvments as better/faster cache/branch predictors do wonders in games.

Games are mostly using integer SIMD,which in BD's case is also improved and resides in fp co-processor. But other parts of the core are beefed up too,so they will also help.

[-Boris-]

Better FP performance won't lead to faster games, since they are usually INT heavy, but other improvments as better/faster cache/branch predictors do wonders in games.

Has it been a shift from FPU to INT in games recently? The bad FPU was the reason K6 sucked in games, K6 had very good INT. 3Dnow fixed this in games that supported it. It's also said that it was the superior FPU that made K7 perform better in games. Nowdays I guess SSE and so on has been taking over, but that is still calculated by the FPU right? I'm a bit unsure on this.

[w0mbat]

@mAJORD: i dont know if its 800mhz, more or less. my x6 here has 800/1500/2200/2800 as possible (w/o turbo). im just saying its not running at max speed cause pre-B1 are supposed to have a issue with the multi. of cause im not there to proof it and i can be wrong.

[madcho]

Has it been a shift from FPU to INT in games recently? The bad FPU was the reason K6 sucked in games, K6 had very good INT. 3Dnow fixed this in games that supported it. It's also said that it was the superior FPU that made K7 perform better in games. Nowdays I guess SSE and so on has been taking over, but that is still calculated by the FPU right? I'm a bit unsure on this.

i think games use both, but i can't be categoric, on the one most used.



But there is a loooooot of branch in games, that sure. And Phenom II sucks in it

[-Boris-]

But there is a loooooot of branch in games, that sure. And Phenom II sucks in it

And from what I've heard that might be one of the most improved parts with BD.
Can anyone verify this?

[XRL8]

Two thoughts.
Anyone told this guy to disable all power saving features in both bios and windows + trying others OSES ? Or Another mainboard,there are currently avaialble both AM3 and Am3+ boards with support out there.
Or maybe in the earlier ES`s AMD intentionnaly broke multi management to be sure that no one leakes it, and even if they do, no one will believe them.They did similar things on GPU front.

[Olivon]

There's clearly a problem in this leak:

[informal]

Like wombat said,looks like multiplier issue(and who knows what else).


edit: ok,don't ask how i got to this but C11.5 score is roughly 2.66x lower than what it should be at that clock . I believe the multiplier issue is such that only one core out of 8 is boosted to its clock of 2.8ghz,rest are siting at idle clock(some may figure out how i got to 2.66x now ). This brings us back to what would 3.5Ghz X8 get in C11.5. Answer :~10.6,dangerously close to DH slide . As for 3dmark cpu score ,i'm not sure how it scales with more cores or clock speed. According to same DH slide,X8 should have 50% higher score than 1100T ,so around 8600pts.

[FlanK3r]

now qestion, what P-state is it...? The lowest 800 MHz or some diferent (1200/1400/1500/1600 MHz)?

[Mats]

A bit brainstorming here:

If only one core runs at full speed, and CPU-Z reports a 8P system, maybe Windows interprets the CPU the same way?

Maybe Server 2008 Enterprise (8P support) would help? Or 2003..

IDK what happens if you run W7 on a 8P system, does even load Windows, or does it lock down the unsupported cores to min P-state?

[NEOAethyr]

Default pwr states are:
4x
8x
12x
user defined
turbo

But we know those can all be changed.
Default ref clock would be 200mhz of course unless something is messed up, and it's possible it could mean 100 or 150mhz.

2600mhz supposed stock setting.
That would be a 13x default ratio (user define pwr state thing above).

Given the sore it got on pi though I would guess it's more like 2000mhz comparing to a thuban.
I rather doubt I needs 600mhz more to equ to a thuban.
That would make it almost worthless for an upgrade, then again you got a new fpu the setup and 8x cores.
But still...

I suppose it is possible that it's running 2600mhz, but the bios is absolutely bare bones then, that would be the only thing I could think of.

Perhaps it's that along with the mem controller timings not wokring 100% in the bios, say tcr of 2t and some extremely slack timings at some low ram speed.

We are all used to halfway responsibly tweaked machines, which those shots may be showing us something alot more slack.

At any rate, those shots are not valid enough to give us a good idea of how it compares.
It's a beta bios with a beta cpu I guess you could say.

And I bet the bios'es stay that way until a week before launch of close to it.

Edit:
Windows non server only supports 2 sockets I think it was, but supports many cores per socket.
But if the bios was buggy, and the dmi pooling was screwed maybe it might think it's a multi socket system I don't know.
Then again I don't know if dmi even makes a real difference, on a single cpu and single core setup you didn't need a dmi pool at all.