AMD Trinity CPU section to be up to 20% faster than Llano

[CAT-THE-FIFTH]

More details:
http://www.techpowerup.com/154105/AM...5-Chipset.html

It was mentioned by this website that a Trinity sample they had performed better than the Llano sample they had:

http://www.phoronix.com/scan.php?pag...ty_early&num=1

However,this was under Linux.

[kalniel]

That's more like it.

[scaryjim]

As I said on the main site article about this, the 20% has been very specifically cherry picked in terms of task to highlight the improvement. Unless they can hike the clockspeeds significantly and get MS to release a scheduler patch for Win 7, I can see Trinity benchmarking quite disappointingly in CPU-intensive benchmarks. The real question is whether the further-enhanced GPU can make it an overall better package than the incumbent Llano APUs. I suppose if they can give the top end Trinity the same CPU performance as an A8-3850, 30% better IGP performance and keep to the same TDP, that wouldn't be too bad a showing...

[CAT-THE-FIFTH]

As I said on the main site article about this, the 20% has been very specifically cherry picked in terms of task to highlight the improvement. Unless they can hike the clockspeeds significantly and get MS to release a scheduler patch for Win 7, I can see Trinity benchmarking quite disappointingly in CPU-intensive benchmarks. The real question is whether the further-enhanced GPU can make it an overall better package than the incumbent Llano APUs. I suppose if they can give the top end Trinity the same CPU performance as an A8-3850, 30% better IGP performance and keep to the same TDP, that wouldn't be too bad a showing...

I think your calculations on CPU performance are a tad off:

http://forums.hexus.net/2160214-post4.html

Look at the second link in the OP:

http://forums.hexus.net/cpus/212854-...han-llano.html

Under Linux a 2.5GHZ Trinity CPU compares quite well to a 2.9GHZ A8-3850 under Linux.

The FX4100 for example is not really that much faster than an A8-3850 in multi-threaded applications such as video encoding:

http://www.guru3d.com/article/amd-fx...mance-review/5

A 20% increase would put the CPU section on par with a Phenom II X4 970BE or 975BE in multi-threaded applications. This means it will be much faster than a Core i3 2100 or even a Core i3 2120 in many applications such as video encoding.

It also indicates single core performance has improved over the A8-3850 and FX4100 as CMT scales more poorly than normal cores.

On top of this the GPU is meant to be 30% faster too. This is all done on the same 32NM process and I suspect the maximum TDP will not exceed 100W as the CPUs need to work in FM1 motherboards.

A 20% improvement would put the quad core Trinity CPU on the level of an FX6100 series CPU or one of the higher clocked Phenom II X4 CPUs.

Phoronix actually tested some Trinity samples under Linux. They were comparing a 2.5GHZ Trinity quad core CPU against a quad core 2.9GHZ A8-3850 under Linux.

[scaryjim]

Sorry CAT, but that website doesn't say it performs better - it says that the figures are "nice" in comparison to an A8-3850. That could mean anything. Given what we know about bulldozer single threaded performance and clock-for-clock performance against Phenom II, I'm very sceptical of a 2.5GHz quad piledriver besting a 2.9GHz Llano. The article doesn't give any actual figures to back up the statement, and without knowing exactly what the different specifications were and how the benchmarking breaks down, it's impossible to extrapolate anything about the CPU component of Trinity.

It would also beg the question of why, if they had samples of an enhanced CPU architecture that has > 16% better IPC than Athlon II / Llano (i.e. 2.5Ghz quad is better than a 2.9GHz quad), did they release an earlier version that has worse IPC? it would also suggest that Piledriver has at least 20% better IPC than bulldozer, which goes against AMDs own releases on the subject.

Make no mistake - if, when Trinty gets released, it does kick Llano in the nads that hard, I'll be over the moon. But it will raise an awful lot of very hard questions about the delivery of bulldozer...

[CAT-THE-FIFTH]

Sorry CAT, but that website doesn't say it performs better - it says that the figures are "nice" in comparison to an A8-3850. That could mean anything. Given what we know about bulldozer single threaded performance and clock-for-clock performance against Phenom II, I'm very sceptical of a 2.5GHz quad piledriver besting a 2.9GHz Llano. The article doesn't give any actual figures to back up the statement, and without knowing exactly what the different specifications were and how the benchmarking breaks down, it's impossible to extrapolate anything about the CPU component of Trinity.

Knowing that website and the reviewer,it would more or less indicate not worse. The slide mentions digital content creation. Guess what?? The CPU which the faster APUs are targetted against is the Core i3. The A8-3850 at least in those sorts of applications is generally ahead of a Core i3 2100(even my Q6600 at 3GHZ is faster than my Core i3 2100 with things like HandBrake). Add 10% or even 20% to that and it will still be ahead. Plus of course since CMT does not scale as well as full cores,this means single core performance has to improve.

It would also beg the question of why, if they had samples of an enhanced CPU architecture that has > 16% better IPC than Athlon II / Llano (i.e. 2.5Ghz quad is better than a 2.9GHz quad), did they release an earlier version that has worse IPC? it would also suggest that Piledriver has at least 20% better IPC than bulldozer, which goes against AMDs own releases on the subject.

Make no mistake - if, when Trinty gets released, it does kick Llano in the nads that hard, I'll be over the moon. But it will raise an awful lot of very hard questions about the delivery of bulldozer...

AMD has commercial contracts to deliver NOW. These customers cannot afford any more delays. Interlagos is an upgrade to existing AMD server CPUs so on that side they will have released it anyway and these run at lower clockspeeds too.

However,AMD is not Intel. This is why they use essentially the same dies for both their commercial and consumer CPUs to maximise what they can sell. Look at both the Phenom and Phenom II for example. AMD took months to get the B3 stepping of the Phenom out.

It is also why AMD even disabled perfectly decent Phenom II X4 CPUs and sold them as Athlon II X3 CPUs. You remember there was even that X2 5000+ which was based on Deneb cores too.

It is likely the best binned CPUs are sold as Interlagos CPUs and the rest are sold as desktop Bulldozer CPUs. A factor people have not looked at though is that both the 4 core and 6 core Bulldozer CPUs consume significantly less power at idle than a Phenom II X4 and on top of this consume less power at load too. For OEMs this alone will be good news.

Maybe AMD could have just stuck with Interlagos now and kept the Phenom II going. However,I suspect that buying more CPUs from GF rather than less would drop the per unit cost of a Bulldozer CPU and a lot of these CPUs will end up in OEM boxes I suspect.

[CAT-THE-FIFTH]

Anyway,look at the following benchmarks in Excel,Winrar and CS5:

http://www.legionhardware.com/articl...fx_4170,4.html

http://www.techspot.com/review/418-a...apu/page8.html

Both reviews are done by the same person. The conditions of the video encoding benchmarks have been changed but not the ones on the page linked.

Even,if Trinity were to have the same IPC as a 4.2GHZ FX4170, a nominal 2.8GHZ Bulldozer based CPU would still be faster than an A8-3850 in these common applications.

[scaryjim]

The benchmark you mention certainly use common applications, but the tests aren't really commonly performed tasks - Monte Carlo simulation? Radial Blur filter? Compressing a 720p video file? Not things that I do every day - in fact not even things I do any day. And since APUs are targetted at the entry level consumer market, they're not things a Trinity APU is likely to find itself doing very often. It's no good saying "this one task is faster in this program, so this program will inherently be faster". The real bottleneck in Excel and Photoshop is the user, and a new CPU isn't going to change that.

Besides, have you looked on the next page of those reviews? In Handbrake an A8-3850 coupled with DDR3-1866 MHz RAM is only 3% behind the PII X4 980BE. However, the FX-4170 - a 4.2GHz processor - was 12% behind the PII 980BE - leaving it almost 10% behind the A8-3850. So even with a 15% IPC increase a 4.2GHz piledriver would just about match a 980BE, putting it a tiny shade ahead of the A8-3850. And that's at 4.2GHz, not 2.5GHz.

What the benchmarks show is that in particular tasks the bulldozer architecture does well, but we already knew that - it has a massive parallel int throughput - just take a look at the Dhrystone CPU test in the guru3d review you linked to. FX-4100's parallel int throughput (@3.6GHz) is higher than the PII X4 975BE (also @ 3.6GHz). So in ideal circumstances the parallel int throughput of bulldozer *has* gone up compared to Phenom II. So we can assume that in the benchmark tasks you've highlighted Bulldozer is able to make use of it's very high int throughput. Compare that to the Whetstone (FPU) chart, and you'll see where AMD have come unstuck: given ideal FPU circumstances the FX4100 falls well behind the PII 975BE, and is scrapping with the A8-3850, which has a 700MHz (~ 20%) clock speed shortfall. Any task which involves FPU usage, or fails to make optimal use of the parallel int throughput, is going to perform less well.

Bulldozer is an odd beast really - there's obviously a lot of potential sitting there untapped. As I said before, I'd love Trinity to fix all of the problems and come out 20% faster than Llano, at the same clocks, in a wide range of tasks. But what I suspect is that we'll get a 20% increase in those parallel int tasks where bulldozer is already matching / beating PII on IPC, coupled with the same performance, or possibly even slightly worse, on all the other tests - bulldozer is already looking at up to 20% worse IPC in flop workloads compared to Llano, and the roadmap only claims ~15% improvement for piledriver...

[CAT-THE-FIFTH]

The benchmark you mention certainly use common applications, but the tests aren't really commonly performed tasks - Monte Carlo simulation? Radial Blur filter? Compressing a 720p video file? Not things that I do every day - in fact not even things I do any day. And since APUs are targetted at the entry level consumer market, they're not things a Trinity APU is likely to find itself doing very often. It's no good saying "this one task is faster in this program, so this program will inherently be faster". The real bottleneck in Excel and Photoshop is the user, and a new CPU isn't going to change that.

No good?? Now you are just changing what you are saying!

Firstly you were saying that the Trinity CPU section would NOT be any faster than the Llano one. I countered by stating that even if IPC did not even increase over Bulldozer,it seems at least in three common applications, such as Excel,Photoshop and things such as unzipping a file with winrar(very common) at 2.8GHZ. The FX4100 and FX4170 have 100MHZ to 200MHZ Turbo too. I made the assumption that the 2.8GHZ part had 200MHZ Turbo.

You were the one saying it would not be any faster than Llano so in this case in three common applications it is not so.

Besides, have you looked on the next page of those reviews? In Handbrake an A8-3850 coupled with DDR3-1866 MHz RAM is only 3% behind the PII X4 980BE. However, the FX-4170 - a 4.2GHz processor - was 12% behind the PII 980BE - leaving it almost 10% behind the A8-3850. So even with a 15% IPC increase a 4.2GHz piledriver would just about match a 980BE, putting it a tiny shade ahead of the A8-3850. And that's at 4.2GHz, not 2.5GHz.

The conditions on the video page are different between the reviews as I mentioned before. Look at the difference between the scores. Have you noticed how the scores don't tally for HandBrake.
A Phenom II X6 1100T is 5% slower in one chart than a Core i7 2600K and in the other 20% and the scores I quoted do tally due to the numbers being exactly the same.

The slide said 20% over an A8-3850. Also,you said before that AMD was only cherry picking "digital media content creation" benchmarks. As you know AMD has used video encoding and rendering benchmarks before for defining these tasks in presentations.

Using your assumption that Trinity is faster than Llano ONLY in such applications applications it would place it at the same level as a higher clocked Phenom II X4.

When I linked to the FX4100 review of the Guru3D,it shows the FX4100 as being around the same as an A8-3850 for video encoding in the applications tested. So 20% on top would mean it would be nearer to a higher clocked Phenom II X4.

This is the problem with what you said. If performance as increased in "digital media content creation" benchmarks,per core performance has to have increased over an A8-3850. CMT does not scale as well as full cores.

I suspect that in Photoshop CS5,applications like winrar,excel and video encoding the higher end Trinity CPU section is going to be a decent improvement over an A8-3850. The only unknown is game performance.

So basically outside of game performance the higher Trinity CPUs is going to be not slower than a high end Llano one.

What the benchmarks show is that in particular tasks the bulldozer architecture does well, but we already knew that - it has a massive parallel int throughput - just take a look at the Dhrystone CPU test in the guru3d review you linked to. FX-4100's parallel int throughput (@3.6GHz) is higher than the PII X4 975BE (also @ 3.6GHz). So in ideal circumstances the parallel int throughput of bulldozer *has* gone up compared to Phenom II. So we can assume that in the benchmark tasks you've highlighted Bulldozer is able to make use of it's very high int throughput. Compare that to the Whetstone (FPU) chart, and you'll see where AMD have come unstuck: given ideal FPU circumstances the FX4100 falls well behind the PII 975BE, and is scrapping with the A8-3850, which has a 700MHz (~ 20%) clock speed shortfall. Any task which involves FPU usage, or fails to make optimal use of the parallel int throughput, is going to perform less well.

Bulldozer is an odd beast really - there's obviously a lot of potential sitting there untapped. As I said before, I'd love Trinity to fix all of the problems and come out 20% faster than Llano, at the same clocks, in a wide range of tasks. But what I suspect is that we'll get a 20% increase in those parallel int tasks where bulldozer is already matching / beating PII on IPC, coupled with the same performance, or possibly even slightly worse, on all the other tests - bulldozer is already looking at up to 20% worse IPC in flop workloads compared to Llano, and the roadmap only claims ~15% improvement for piledriver...

hmmmm... suspicious footnotes - 20% CPU performance increase over Llano is "Based on AMD projections using digital media workload"... being that specific suggests that the general CPU performance boost over Llano is going to be very small - which will comes as little surprise gvein the relative performance of the 2 module FX-4100 (I assume that quad core here means 2 modules). Looks like AMD have decided that mainstream CPU performance simply doesn't need improving for the next few generations - which means they'll fall even further behind Intel in CPU terms, which Intel is then bound to use for marketing purposes.

I'm not disagreeing with AMD mind you - for the vast majority of modern day-to-day workloads a 2GHzish dual core is adequate - but AMD are already behind in the marketing stakes and they could do without giving Intel any more ammunition...

The problem is you were making the assumption that Piledriver won't be an improvement over an A8-3850 in any way which is a bit daft. You were also making the assumption that there won't be ANY improvement overall over Bulldozer.

If this is the case AMD would have just stuck with the first generation Bulldozer cores wouldn't they?? No doubt new steppings would improve power consumption and stock clockspeeds.

The Flex FP unit has problems and to assume that AMD will not try and remedy this is also a bit daft too.

Piledriver is the core which not only will replace Husky in their higher end APUs but also is the replacement for Bulldozer.

[scaryjim]

No good?? Now you are just changing what you are saying.

Firstly you were saying that the Trinity CPU section would NOT be any faster than the Llano one. ...

Not exactly - I said I could see it benchmarking dissapoointingly in CPU intensive benchmarks. In the same way that Bulldozer benchmarked disappointingly - i.e. a lot faster than PII / LLano in some scenarios, but overall very little improvement.

You were the one saying it would not be any faster than Llano so in this case in three common applications it is not so.

No, in three compute-intensive scenarios it is faster. I never use Excel to run Monte Carlo, or Photoshop to do a Radial Blur. I suspect that's true of most home users.

What I said in the other thread (on the news article) was (emphasis added):

the general CPU performance boost over Llano is going to be very small

As in the performance in most real world scenarios. AMD is aiming APUs at low-cost integrator boxes and laptops, and the people buying those will not regularly be performing compute-intensive tasks. WinRAR is one possible application where Trinity will have much better performance, but then how many people regular compress and decompress large files with WinRAR? Most people I know use Windows' built in zip compression, and use it on much smaller, and easier to compress, files than the benchmarks were done with. In the real world, the difference is not going to be so significant or noticable.

Note I also didn't say it wasn't going to be faster - I said the increase was going to be small.

... Look at the difference between the scores. Have you noticed how the scores don't tally for HandBrake. ...

Yes, but that doesn't mean you can't make some assumptions - even if they aren't entirely borne out. Something odd is going on with that test as the i7-920 is below the X4 980BE in the second review.

None the less, the FX-4170 is significantly further behind the X4 980BE than the A8-3850. If you adjust for the differences between the X4 980BE, the X6 1100T and the i5 2500, you can still see that the FX-4170 falls behind in that test: not as much as I suggested, but enough that it still needs to make up ground.

The problem you were making the assumption that Piledriver won't be an improvement over an A8-3850 in any way which is a bit daft.

As I've already pointed out, that's not the case. There's obviously an improvement in some areas, probably related to the massive int throughput that we know bulldozer has*. I just don't think it's going to be a significant across-the-board improvement: and that's a logical conclusion based on the comparison of K10.5 processors to Bulldozer: bulldozer was *not* an across the board improvement on Phenom II, and whilst piledriver will be an enhancement of bulldozer, the Stars core in Llano are meant to be an improvement on Athlon II / Phenom II cores.

Now, *IF* part of the problem with Bulldozer is low level bugs that can be fixed in a core revision (rather than a full redesign), then Trinity could yet see significant [b]across-the-board[/i] improvements over Llano. if it does, I'll be over the moon.

But I'm not going to get all excited about a few vague, unverified reports about Trinity performance being "nice", or by an AMD press release claiming 20% increased performance but hastily footnoting it with regards to "projections" (i.e. the 20% increase hasn't been tested) and specific applications. Once we've got a released product and hard figures, I'll decide whether I should be getting excited

*side note: due to the discreet nature of bulldozer's int cores, in certain (heavy) workloads it scales just like a full multi-core. Benchmarks have already shown that it's actually BD's single-threaded performance that is below par, as oppose to HT, which has good single-threaded performance but tends to suffer under heavy threading...

[CAT-THE-FIFTH]

Not exactly - I said I could see it benchmarking dissapoointingly in CPU intensive benchmarks. In the same way that Bulldozer benchmarked disappointingly - i.e. a lot faster than PII / LLano in some scenarios, but overall very little improvement.

No, in three compute-intensive scenarios it is faster. I never use Excel to run Monte Carlo, or Photoshop to do a Radial Blur. I suspect that's true of most home users.

You have just contradicted your first point and your second point.

So now you go from it won't increase that much to if it does it is pointless. So if the case why make a comment about CPU performance if it does not bother you.

What I said in the other thread (on the news article) was (emphasis added):

So image editing,excel,file unzipping and some video encoding will be the most intensive tasks most home users will be doing though.

Gaming is the other task. Maybe performance will be lower or slightly better but with an HD5570 level IGP or slightly slower I suspect it won't be an issue.

So out of interest were you expecting like a 50% to 100% improvement over an A8-3850??

As in the performance in most real world scenarios. AMD is aiming APUs at low-cost integrator boxes and laptops, and the people buying those will not regularly be performing compute-intensive tasks. WinRAR is one possible application where Trinity will have much better performance, but then how many people regular compress and decompress large files with WinRAR? Most people I know use Windows' built in zip compression, and use it on much smaller, and easier to compress, files than the benchmarks were done with. In the real world, the difference is not going to be so significant or noticable.

Note I also didn't say it wasn't going to be faster - I said the increase was going to be small.

Loads of people use Winrar and 7 zip(at least the ones I know). Winrar is a common application. So again if image editing,excel,video encoding and unzipping files are not areas worthy of improvements in performance then what are??

Gaming only??

What then??

AMD have a thread where you can post questions:

hardforum.com/showthread.php?t=1645024

Maybe you can ask them whether they should try and increase performance in areas you find relevant.

Yes, but that doesn't mean you can't make some assumptions - even if they aren't entirely borne out. Something odd is going on with that test as the i7-920 is below the X4 980BE in the second review.

None the less, the FX-4170 is significantly further behind the X4 980BE than the A8-3850. If you adjust for the differences between the X4 980BE, the X6 1100T and the i5 2500, you can still see that the FX-4170 falls behind in that test: not as much as I suggested, but enough that it still needs to make up ground.

Look at the review of the FX4100 on Guru3D. In the video encoding applications it is broadly similar to an A8 3850. However,again the slide 20% improvement over an A8-3850 which means higher clocked Phenom II X4 performance level.

Again,you are making the assumption that Piledriver will be no real improvement over Bulldozer.

As I've already pointed out, that's not the case. There's obviously an improvement in some areas, probably related to the massive int throughput that we know bulldozer has*. I just don't think it's going to be a significant across-the-board improvement: and that's a logical conclusion based on the comparison of K10.5 processors to Bulldozer: bulldozer was *not* an across the board improvement on Phenom II, and whilst piledriver will be an enhancement of bulldozer, the Stars core in Llano are meant to be an improvement on Athlon II / Phenom II cores.

Now, *IF* part of the problem with Bulldozer is low level bugs that can be fixed in a core revision (rather than a full redesign), then Trinity could yet see significant [b]across-the-board[/i] improvements over Llano. if it does, I'll be over the moon.

Again,you are making the assumption that AMD will not fix the issues the flex FP unit. On top of this you are comparing two different architechures and trying to say performance is not going to improve with later versions of the newer one which is daft.

Your statement:

"Looks like AMD have decided that mainstream CPU performance simply doesn't need improving for the next few generations - which means they'll fall even further behind Intel in CPU terms, which Intel is then bound to use for marketing purposes."

The funny thing what is "mainstream" performance though??

So image editing,excel,file unzipping and some video encoding will be the most intensive tasks most home users will be doing though. But according to you increases in these areas are pointless.

So again is it gaming only you are interested in??

But I'm not going to get all excited about a few vague, unverified reports about Trinity performance being "nice", or by an AMD press release claiming 20% increased performance but hastily footnoting it with regards to "projections" (i.e. the 20% increase hasn't been tested) and specific applications. Once we've got a released product and hard figures, I'll decide whether I should be getting excited

*side note: due to the discreet nature of bulldozer's int cores, in certain (heavy) workloads it scales just like a full multi-core. Benchmarks have already shown that it's actually BD's single-threaded performance that is below par, as oppose to HT, which has good single-threaded performance but tends to suffer under heavy threading...

As opposed to getting excited and then jumping on the Bulldozer suckz and hence Piledriver is just as bad bandwagon!

Your the one who was going on how Trinity CPU performance will be around the same . Yet when shown common everyday applications which Bulldozer is generally faster than an A8 3850 at the same clockspeeds you now say they are irrelevant.

I still don't understand what areas you want it to be much better in though.

[scaryjim]

*sigh*

You've said a lot CAT, and to be honest most of it doesn't need answering. Mostly because you keep coming back to one basic fallacy:

Your the one who was going on how Trinity CPU performance will be around the same.

No, I don't. I keep saying that I believe we won't see 20% increases *across the board*. I'm not even going to think about making statements about what performance *will* be like, because, plain and simple, we don't know. We don't even have any figures *at all* for Trinity yet. What's actually happening here is that I'm suggesting caution in the face of some very vague, unsupported statements about Trinity, and you're jumping up and down shouting in defiance.

At this point I'm not assuming anything. I'm extrapolating from what we already know. In a range of benchmarks, Bulldozer performs well in some and poorly in others. Piledriver will be an evolutionary enhancement of Bulldozer, so unless there are bugs that can be fixed without completely redesigning the CPU core, it is reasonable to extrapolate that Piledriver will also perform well in some tests and poorly in others. Read your own comments and the reviews they link to:

The FX4100 for example is not really that much faster than an A8-3850 in multi-threaded applications such as video encoding

So a 3.6GHz bulldozer quad-core isn't that much faster than an A8-3850, but a 2.5GHz piledriver (30% slower clock speed) will be...?


----------------------------------------------------------------------------------------------
Bottom line here, I wasn't commenting on Trinity performance. I was commenting on the rumours, based on known information. So rumours-wise we've got an AMD projection of 20% better than A8-3850 in specific applications, and we've got one person on the internet who says he's seen some "nice" figures for a 2.5GHz Trinity engineering sample.

The known information - or at least the most relevant available information - is the comparison of bulldozer CPUs to A8-3850. The FX-4100 is the nearest available configuration to a Trinty CPU (2 modules / 4 cores) and at 3.6GHz it performs very similarly to the A8-3850 in a wide range of benchmarks: you've read the guru3d review, so you know that already.

So to extrapolate from the known and the rumours: Trinity will be up to 20% faster than a8-3850, which in turn has very similar performance the FX-4100. If Trinity launches at 2.5GHz, it will need an IPC improvement of 44% just to match FX4100, let alone push signficantly past it. If we believe that a 2.5GHz Trinity having "nice" figures compared to A8-3850 means generally better, then we have to accept that AMDs first revision of bulldozer has managed to increase IPC by 44% - three times AMDs published performance projection - and it's done it despite losing the L3 cache.

So no, I don't buy that.

I'm sure Trinity will improve IPC compared to FX-4100, and I'm sure that in certain tasks it will beat A8-3850 by 20%, but I'm also sure that it won't do that at 2.5GHz. That means it's not going to be a messiah CPU, but it may mean that performance in its weak areas is no worse than A8-3850 (something you can't say for FX4100). What I wanted, and this applies to bulldozer more than piledriver because we don't know what AMD have been able to fix in piledriver, is a CPU that performed strongly in all areas. I'd've been happy with a 10% performance increase if it had been in every benchmark. But it wasn't, and in some benchmarks bulldozer did worse than Deneb and its derivatives. Which means that when piledriver comes out, whether that's Trinity or a future FX CPU, it's still going to be compared to Deneb, because for some benchmarks (and therefore workloads) Deneb is still AMDs best CPU...

[shaithis]

At this point I'm not assuming anything. I'm extrapolating from what we already know. In a range of benchmarks, Bulldozer performs well in some and poorly in others. Piledriver will be an evolutionary enhancement of Bulldozer, so unless there are bugs that can be fixed without completely redesigning the CPU core, it is reasonable to extrapolate that Piledriver will also perform well in some tests and poorly in others. Read your own comments and the reviews they link to:

Isn't extrapolation a form of assumption?

We all know where BD falls down and many are awaiting a Windows update, BIOS revision, chip revision etc. IMO, it's real-world performance is currently being hampered and there is plenty of evidence out there to support that theory....are you taking that into account in your extrapolation? Especially when the biggest concern (Windows scheduler) is what seems to be hampering the lightly-threaded performance figures....i.e. the weak point you keep prodding.

[scaryjim]

Isn't extrapolation a form of assumption?

No, extrapolation uses existing knowns to predict future unknowns: assumption uses existing rumours to wildy guess at future unknowns

The difference is more subtle than this, but basically extrapolation is about calculating based only on known factors.

We all know where BD falls down and many are awaiting a Windows update, BIOS revision, chip revision etc. IMO, it's real-world performance is currently being hampered and there is plenty of evidence out there to support that theory....are you taking that into account in your extrapolation? Especially when the biggest concern (Windows scheduler) is what seems to be hampering the lightly-threaded performance figures....i.e. the weak point you keep prodding.

I can't take any of that into account in my extrapolation, because it's all assumed*, not known

We all know where BD falls down (lightly threaded workloads, non-AVX FP loads), but to be perfectly honest we don't know why it falls down there - we can make educated guesses about the FPU not functioning correctly, about schedulers not being optimised for the modular approach, indeed I've made a wild but supportable accusation recently that the "quick" branch predictor is very poor and causing the pipeline to be flushed more often than at should be - but they are what they are: guesses. Or, if you prefer, assumptions.

Now, if it turns out that we're right about one or more of things, and if they get fixed between bulldozer and piledriver, we move into the area of being able to realistically consider 40%+ real-world IPC increases, but you'd be making a lot of assumptions on what AMD have been able to fix and patch in a relatively short space of time. I'm looking at the known facts - so stuff that AMD have released officially, or published figures from real hardware benchmarking by a reputable source.

As I said some time ago, if you look at guru3d's Dhrystone test BD int IPC is actually a shade higher than PII int IPC in ideal circumstances. If the FPU is "broken" and they're fixing it in piledriver - and assuming it's optimised at least as well as the int pipelines, then we could see > 30% improvement. I can't remember the exact figures, but I think the early look at Windows 8 gave up to 10% improvement due to scheduler optimisations. It's all stuff that could add up to making piledriver buldozer-done-right (in the way GF110 was Fermi done right). But, at this moment in time, that's based on a lot of speculation and assumption, which we have little to no evidence to support. So I'm going to continue being cautious and a little skeptical

*EDIT: To be fair, there's strong evidence that improvements to the Windows Scheduler do improve lightly threaded workloads, but from memory it's only really enough to bring it to parity with similarly clocked PII processors at best...?

[shaithis]

No, extrapolation uses existing knowns to predict future unknowns: assumption uses existing rumours to wildy guess at future unknowns

Assuming everything else stays constant....

That's a big assumption. Especially in the face of all the data and info that people have collected from testing and analysing the chip....

[scaryjim]

Ah, but it's a smaller assumption than assuming that a whole host of different things change. For instance, to factor in a change to the Windows Scheduler (which is the factor we have most evidence for) you still have to assume that a) the scheduler will be patched, b) that it'll be patched prior to the release of Trinity, c) that it will have a positive effect on the performance of Trinity, d) the exact level of that performance increase ... and that's just for one of the assumptions we considered above.

Ockham's Razor favours my version