Found another one: Ukraine store accidentally ships FX-8120 and it gets tested!

[Idontcare]

Looks like its faster than 2500K and nearly as fast as 2600K, am i missing something?

Its priced below those.

Yeah, that has to be overclocked to 4GHz to do that, while not OC'ng the 2500K and 2600K that you are comparing it to.

Comparisons need to be stock-to-stock or OC'ed-to-OC'ed, not mixed and matched.

 

[StrangerGuy]

The problem is that the test file parameter is unknown (file size, type and compression factor). As far as I can remember, the highest numbers I've seen are ~4.7GB/s on my friend's Core i7 920 @3.6GHZ, and ~4.5GB on Core i7 2600K @4.2GHz (with a 300MB high compression file).

Nope it's with the built-in benchmark. It runs forever.

 

[Silverforce11]

If it fights between 2500k and 2600k for cheaper i'm not fussed and my next rig may finally have an AMD CPU.

Multithread perf is more important to me than single core ipc, as all the new games ive been playing support a lot of threads and tend to be GPU limited. It comes down to media encodes/compression as that's the only other intensive thing most users do with their PC. Web browsing etc even a netbook can handle.

 

[formulav8]

Yeah, that has to be overclocked to 4GHz to do that, while not OC'ng the 2500K and 2600K that you are comparing it to.

Comparisons need to be stock-to-stock or OC'ed-to-OC'ed, not mixed and matched.

Isn't the 8150 around a 4.0ghz stock clock or is it 3.8ghz? If its 4ghz then it should compete ok in the $200-$250 area. Its definitely NO FX though.

At least they have a new foundation to build on. They seem to already have a tweaked successor to BD coming in 1H 2012 if GF yields don't kill it.

The biggest problem is AMD did NOT get the clockspeeds they were wanting. They expected a specific area in clockspeed; and at higher speeds it will compete quite well with Thuban and Intel. But they were apparently caught Way off guard about all of the teething issues GF is currently having at 32nm. And thus, it is what it is. They have to release something. Whether its good or bad they got to release a product they've spent almost 4 years on and not let all of the r&d go to waiste completely. imo

 

[StrangerGuy]

Isn't the 8150 around a 4.0ghz stock clock or is it 3.8ghz? If its 4ghz then it should compete ok in the $200-$250 area. Its definitely NO FX though.

At least they have a new foundation to build on. They seem to already have a tweaked successor to BD coming in 1H 2012 if GF yields don't kill it.

The biggest problem is AMD did NOT get the clockspeeds they were wanting. They expected a speicific area in clockspeed and at higher speeds it will compete quite well with Thuban and Intel. But they were apparently caught Way off guard about all of the teething issues GF is currently having at 32nm. And thus, it is what it is. They have to release something. Whether its good or bad they got to release a product they've spent almost 4 years on and not let all of the r&d go to waiste completely.

FX-8150 is 3.6GHz base and 4.2GHz turbo.

I don't know about you guys but a 4.5GHz 8 core BD barely beating a 4.3GHz 2500K in Winrar is downright pathetic.

 

[-Slacker-]

Could it possibly be a bad windows\scheduler\BIOS issue, in other words, would a better driver or perhaps a software patch improve BD's performance?

Here's to hoping.

 

[Idontcare]

The biggest problem is AMD did NOT get the clockspeeds they were wanting. They expected a speicific area in clockspeed and at higher speeds it will compete quite well with Thuban and Intel. But they were apparently caught Way off guard about all of the teething issues GF is currently having at 32nm. And thus, it is what it is. They have to release something. Whether its good or bad they got to release a product they've spent almost 4 years on and not let all of the r&d go to waiste completely. At least that is my opinion on the matter

We can't blame GloFo at this point, we don't know whether this was a "failure to DFM" on AMD's design part or if this was a "failure to meet the parametric targets" on GloFo's process technology part.

The same as occurred with 90nm Prescott. People railed that it was the 90nm process itself that was at fault, until the mobile Dothan parts came out on the exact same process (but different microarchitecture) and the same 90nm process looked golden.

If this is a case of GloFo failing to deliver to their parametric models (leakage specs, driver current specs, etc) then that is a disaster for their foundry model as other customers look for this as signs of credibility going forward with nodes that are currently under design.

If this is a case where AMD failed to properly design for manufacturing (DFM), as Nvidia did with Fermi on TSMC's 40nm, then that is more your garden variety "they done goofed" scenario. GloFo is going to have lots of customers that done goof, all foundries have that, they can't hold their hands and design the chips for the customer so that it can mesh with the reality of the fab environment.

The jury is out still whether or not Bulldozer is AMD's Prescott, but the preliminary info's are not painting a favorable picture.

 

[Black96ws6]

If it fights between 2500k and 2600k for cheaper i'm not fussed and my next rig may finally have an AMD CPU.

Multithread perf is more important to me than single core ipc, as all the new games ive been playing support a lot of threads and tend to be GPU limited. It comes down to media encodes/compression as that's the only other intensive thing most users do with their PC. Web browsing etc even a netbook can handle.

Then you want a 2600k. At stock it beats an OC'd Bulldozer in both single and multi-threaded apps.

An OC'd 2600k would downright embarrass it. And this is coming from an AMD owner*

(*assuming these benchmarks and all the others that have been popping up now that the press kits are out are legitimate)

 

[formulav8]

We can't blame GloFo at this point, we don't know whether this was a "failure to DFM" on AMD's design part or if this was a "failure to meet the parametric targets" on GloFo's process technology part.

The same as occurred with 90nm Prescott. People railed that it was the 90nm process itself that was at fault, until the mobile Dothan parts came out on the exact same process (but different microarchitecture) and the same 90nm process looked golden.

If this is a case of GloFo failing to deliver to their parametric models (leakage specs, driver current specs, etc) then that is a disaster for their foundry model as other customers look for this as signs of credibility going forward with nodes that are currently under design.

If this is a case where AMD failed to properly design for manufacturing (DFM), as Nvidia did with Fermi on TSMC's 40nm, then that is more your garden variety "they done goofed" scenario. GloFo is going to have lots of customers that done goof, all foundries have that, they can't hold their hands and design the chips for the customer so that it can mesh with the reality of the fab environment.

The jury is out still whether or not Bulldozer is AMD's Prescott, but the preliminary info's are not painting a favorable picture.

None of us knows exactly who the problem is. But it can definitely be a yield problem. Just look at the problem Llano is having.

Also, you can't exactly compare Dothan and P4. Dothan only had to go to 2ghz or so on the 90nm process. P4 had to go MUCH higher in clockspeed. So it could still have been the 90nm process that was the P4's problem. Dothan had completely different requirements that the 90nm process could deliver. But when it came to the P4's requirements it may not have delivered. Of course this is only my opinion.

 

[bronxzv]

(Sorry for all the images, but I think this is really what everyone wants to see)

if this test is genuine we see clearly that Bulldozer is L1D$ write bandwidth castrated, it will choke any high throughput code, even more with two threads running on a module since the two L1D$s are write-through and they will fight for the lackluster L2$ write bandwidth

this looks incredibly unbalanced for a design pretending to support 256-bit FMA4

 

[dma0991]

This is starting to turn out to be a disappointment if BD can't even be good at MT workloads. I wouldn't mind too much it being bad at ST tasks but to be bad at both ST and MT tasks is just terrible.

I have a feeling that the legit reviews that are being published 12th Oct onwards are not going to bring any optimistic news to those who are planning to buy BD. :|

 

[formulav8]

if this test is genuine we see clearly that Bulldozer is L1D$ write bandwidth castrated, it will choke any high throughput code, even more with two threads running on a module since the two L1D$s are write-through and they will fight for the lackluster L2$ write bandwidth

this looks incredibly unbalanced for a design pretending to support 256-bit FMA4

Are there actual apps to see where 'exactly' the weak point is on a cpu and what scenerio?

 

[classy]

Yeah, that has to be overclocked to 4GHz to do that, while not OC'ng the 2500K and 2600K that you are comparing it to.

Comparisons need to be stock-to-stock or OC'ed-to-OC'ed, not mixed and matched.

Thats not overclocked. The 8120 turbos to 4g. The 2600K turbos to 3.8. But seeing the Cinebench which is improved, but all the other ones seem lower that reported other leaks makes me kinda not fully trust these results. But we are closer now to the truth.

 

[bronxzv]

Are there actual apps to see where 'exactly' the weak point is on a cpu and what scenerio?

you see this typically with profilers analyzing performance events like cache misses

from my experience with high throughput code the bottleneck is always the memory hierarchy, particularly the L2/L3 cache bandwidth since it's generally unpractical to do cache blocking for the L1D cache and if you have too much system memory accesses you can't achieve high throughput anyway i.e. your code is really not much more than a memory bandwidth benchmark such as stream in this case

for example on Sandy Bridge you can get up to 80% speedup with AVX-256 vs SSE 128-bit when using mostly registers, up to 50% speedup for workloads fitting in the L1D cache though on realworld cases you are limited by the L2 cache bandwidth and the speedups are typically only 10%-20%

 

[BlueBlazer]

if this test is genuine we see clearly that Bulldozer is L1D$ write bandwidth castrated, it will choke any high throughput code, even more with two threads running on a module since the two L1D$s are write-through and they will fight for the lackluster L2$ write bandwidth

Or it could possibly be that cache aliasing bug which I've mentioned earlier. :hmm:

Could it possibly be a bad windows\scheduler\BIOS issue, in other words, would a better driver or perhaps a software patch improve BD's performance?

Perhaps that's the reason why AMD has been highlighting Windows 8 in their presentation. There could be some changes to the Windows kernal (including a patch?) to support Bulldozer. :hmm:

 

[bronxzv]

Or it could possibly be that cache aliasing bug which I've mentioned earlier. :hmm:

probably not, a cache bandwidth test typically use a linear range of addresses with a constant stride and a single live stream so it isn't a scenario prone to cache aliasing issues

now, real world code will be another story with more than one memory stream per thread and two threads per module, the effective L1D$ available write bandwidth for each thread will be less than half what is shown here (the two L1D$ are write-through so each write must go to the shared L2), clearly not matching the bandwith required for well optimized AVX-256/FMA4 kernels

 

[alyarb]

I think the cache performance alone gets GloFo off the hook. You can't forgive 2 years of "look how wide this CPU will be" rhetoric when the cache is slower than Deneb. You would think they have a battery of simulation systems and/or FPGAs to test the individual behavior of new systems before they are set in stone. They said they would focus on branch prediction too and yet IPC is decreased across the board. Performance in almost every category seems to have gone down compared to Thuban.

 

[formulav8]

you see this typically with profilers analyzing performance events like cache misses

from my experience with high throughput code the bottleneck is always the memory hierarchy, particularly the L2/L3 cache bandwidth since it's generally unpractical to do cache blocking for the L1D cache and if you have too much system memory accesses you can't achieve high throughput anyway i.e. your code is really not much more than a memory bandwidth benchmark such as stream in this case

for example on Sandy Bridge you can get up to 80% speedup with AVX-256 vs SSE 128-bit when using mostly registers, up to 50% speedup for workloads fitting in the L1D cache though on realworld cases you are limited by the L2 cache bandwidth and the speedups are typically only 10%-20%

Thanks for the info! Makes sense

 

[Nemesis 1]

Terrance was right . The more he talked the slower BD got . He was banned when BD was still going to destroy intels SB-E platiform . He was banned unjustly and way to early . He hadn't talked BD down to 2500k performance as of the time of ban . Not to worry terrance I covered the 2500k for ya.

 

[formulav8]

Terrance was right . The more he talked the slower BD got . He was banned when BD was still going to destroy intels SB-E platiform . He was banned unjustly and way to early . He hadn't talked BD down to 2500k performance as of the time of ban . Not to worry terrance I covered the 2500k for ya.

:\ :whiste:

 

[MentalIlness]

Here you go, just for a little fun.

 

[mrSHEiK124]

Wow. PO'd.

 

[Black96ws6]

Don't forget Coolaler benched this exact same chip about a month ago, and got about the same results:

He said at the time it was the B2 latest version...

 

[Ryun]

I'm getting anxious waiting for the official reviews (not that I haven't been already). On the one hand, these results are atrocious. They're worse than Deneb clock-for-clock, core-for-core.

On the other hand, there a number of respected enthusiasts in the know who remain excited about the chip.

I told myself that I would pick up BD if it had better IPC than my Conroe C2D. I'm still hoping that's the case come Oct. 12th.

 

[frostedflakes]

Higher IPC than Core 2 Duo? Not really setting the bar high there, lol.