Same frequency Bulldozer, Thuban and SandyBridge test

[mmarq]

http://en.inpai.com.cn/doc/enshowcont.a ... ageid=8095

To me better test than most of tests that go around. It this there seemed to be a clear intension of comparing oranges with oranges. The only remark is that BD DRAM is 1866 but with not so good timings, while the others were 1333 with good timings. Perhaps the tester was taking the specs to its ultimate consequences, after all SB and Thuban default spec is DRAM 1333.

Summary:
The original FX-8150 3.6Ghz wins 11 tests out of 16, exceptions 5 -> encryption, mem bandwidth, wPrime and the last 2 games on the list for SB. The original FX-8150 3.6Ghz loses only 1 of the tests for Thuban(but not SB) wPrime 2.05.. wonder what this can do... http://www.cpu-world.com/CPUs/Bulldozer ... -8170.html

The FX-8150 underclcoked to 3.3Ghz, the same clock of Thuban X6, Deneb X4, and SB 2500k in this test, OTOH only wins 5 tests -> Sandra CPU arithmetic, multimedia and multi-core efficiency, WinRAR and 7-ZIP compression. Notably since its the FUD running around the NET is lack of IPC (potential), BD at the same clock winning the arithmetic tests for all other CPUs, in my POV closes the matter (hint: its not potential IPC ; hey! Realworldtech, is not AGU its AGLU and that 3th L1-D port has an use)

To note that the crises2 game, as well the other games, the decisive factor is the GPU, matter of fact Crises2 shows this clearly, the differences are so tiny that we can say fall into a noise margin, all CPUs have the same result.



This shows clearly that JF-AMD was not lying when he posted around in several forums, BD general scaler x86 IPC might in fact have increased a little over K10.. throughput might in fact increased more than a little. Not showing because cache system and IMC are not optimized, or in another view, already optimized to take on it a lot of higher clock.


























BD a fail or a victory.. only missing high clock ?.. anyone take his own conclusion.

[Pietro sk]

This looks fair.

Wow about multicore eff.

[keithlm]

The BD at the same frequency as the i5-2500 abuses it and makes it beg for more.

Actually there isn't that much of a difference... personally I would call it a tie.

(But I was just using the same standards and slang that Intel fans would use if the results were exactly the opposite.)

[AussieFX]

Surely those reviewers didn't tell me lies...

[NotaRocketScientist]

It looks like BD is not so bad as we were led to believe from the review sites. I was waiting for someone to do their own review and post results. BD fits in right where it was made to fit in. From some of the reviews, BD even smacks around the i7. I don't see the logic in people buying a $999 processor and even higher if you buy a motherboard vs. a processor that costs $300 and gives just the same performance.

[keithlm]

It looks like BD is not so bad as we were led to believe from the review sites. I was waiting for someone to do their own review and post results. BD fits in right where it was made to fit in. From some of the reviews, BD even smacks around the i7. I don't see the logic in people buying a $999 processor and even higher if you buy a motherboard vs. a processor that costs $300 and gives just the same performance.

Haven't you noticed the popular benchmarking trend... there are about 4 or 5 benchmarks that some people treat as being of paramount importance. To those people it doesn't matter if similar benchmarks contradict those results.

Before BD was released it was predicted that if it didn't win those popular benchmarks that it would be considered a failure by many reviewers. (If it had won every benchmark available except for those popular benchmarks it would still be considered a failure. In reality it is about equal in the non-popular benchmarks.)

[Kedas]

You do realize that the 2500k is a 'cheap' part of 95W while FX8150 is AMD's best and using up 125W.

[keithlm]

You do realize that the 2500k is a 'cheap' part of 95W while FX8150 is AMD's best and using up 125W.

You do realize that the FX8150 falls between the 2500k and the 2600k in performance don't you? Regardless of all of the people trying to pretend otherwise.

(And unless you're running a folding farm the amount of power used is a tired and useless argument made by people that realize they need something additional to support their failed opinion.)

[seronx]

You do realize that the 2500k is a 'cheap' part of 95W while FX8150 is AMD's best and using up 125W.

You do realize 2500K is actually 45-65W @ 3.3GHz for the CPU where the FX8150 is 125W @ 3.9GHz All Core Turbo

3.3GHz for the FX-8150 would be around 90-100W

The 2500K includes the GPU

It would be more fair to compare Sandy Bridge to Falco...I mean Trinity if going by APU standards

[mmarq]

Yes that 45-60W might be true for SB... but Intel doesn't count NB, interconnect(that Ring-bus), IGP... and maybe even the L3(though doubt it)... as part of the CPU for that power, while FX-8150 is Orochi a server chip, with plenty of baggage not needed for desktop...

http://www.lostcircuits.com/mambo//inde ... mitstart=8

all Core iX processors receive auxiliary power through the 3V and the 5V rail as described in detail in our Core i7 Power Plays article.Therefore, the measurements only show the actual core section in those CPUs as opposed to the AMD and older Intel CPUs that receive the full power through the auxiliary 12V supply lines. Without knowing more about the supply rails it is hard to comment much further.

ppl have heated debates about this... and it really makes me wonder, really wonder!... if even 2600k is a 95w category chip, then where is a 3.8-4.0ghz 2800 or 2900 bin ? .. is it because they really can't do it without trashing current yields ? .. meaning those 2500-2600 are already higher than marketed accounting 3.3v and 5v lines ?

[gallier2]

Yes that 45-60W might be true for SB... but Intel doesn't count NB, interconnect(that Ring-bus), IGP... and maybe even the L3(though doubt it)... as part of the CPU for that power, while FX-8150 is Orochi a server chip, with plenty of baggage not needed for desktop...

http://www.lostcircuits.com/mambo//inde ... mitstart=8

all Core iX processors receive auxiliary power through the 3V and the 5V rail as described in detail in our Core i7 Power Plays article.Therefore, the measurements only show the actual core section in those CPUs as opposed to the AMD and older Intel CPUs that receive the full power through the auxiliary 12V supply lines. Without knowing more about the supply rails it is hard to comment much further.

ppl have heated debates about this... and it really makes me wonder, really wonder!... if even 2600k is a 95w category chip, then where is a 3.8-4.0ghz 2800 or 2900 bin ? .. is it because they really can't do it without trashing current yields ? .. meaning those 2500-2600 are already higher than marketed accounting 3.3v and 5v lines ?

I heard that several times here that the power of the AMD and Intel chips are counted differently, but hard data to corroborate that was pretty scarce. So what I did was to find what the hardware documentation of both parties specified, here is what I found.

I did not find comparable data, but looking at the electrical specifications for intel Sandy Bridge 2700K, I discovered that they allow a lot of current for their CPU.
- 200A on Vdd with a voltage between 0.7 and 1.3 (140W to 260W)
- 2A on Vpp (?) with a voltage around 1.0 (2W)
- 32A on V?? with a voltage of 1.1V (35.2W)
- 32A on V?? with a voltage of 0.9V (28.8W)
140+2+35.2+18.8=196W
260+2+35.2+18.8=316W
So, intel has a maximum (electrical) wattage specified of 316W. To convert that in TDP one should subtract a little bit of this value for the real work done by the processor. So let's assume a mean value of 250W.

I haven't found comparable values for AMD processors (if someone knows how to get the processor datasheets, I'm a taker). Therefore I looked at the AM3+ specification sheet, and there, the electrictal power is not specified but only the TDP, but the interesting point of that document, is that the specified TDP is always for all voltage rails. The sheet contains a long and boring section where they specify the breakdown for each rail. This means for example, a 130W TDP socket, will have to dissipate 80+25+25 on the different voltage sections.

This is the nearest I have found to an irrefutable smoking gun that Intel, again, presents their data in a unhonest way.

[mmarq]

This is the nearest I have found to an irrefutable smoking gun that Intel, again, presents their data in a unhonest way.

Worst than that, you seem to know what you are looking for.. for those that really don't, they are so bombarded with propaganda that is hard to keep focus on what you are looking for!... looking for concrete Intel data is time consuming and a very boring experience, means digging trough a lot of worthless junk.

about the "power" issue, its clear that Intel is after low power with IB and some with SB.. perhaps they don't want to push things and trash yields -> its quite different when you tweak for high performance and high power and low power(low power is a quite easier to do right). AMD and GF perhaps the problem was that. Orochi didn't helped being a server chip that doesn't need high clock and process gate-first is a pig to work with because high temperature applied to those "first" formed gates tends to trash the things.. it needs another run to put a "protection" on top of those gates after they are formed, to form the other features.

Another point is that TurboBoost 2 permits already going above the specked TDP for a while, and that is something to take account with, when you bin for higher frequency. Hope TurboCore 3 permits similar.. and IF vishera sheds the all server baggage not needed... if this is already possible with a B3 http://www.cpu-world.com/CPUs/Bulldozer ... -8170.html a proper desktop would permit perhaps 200Mhz more clock, and close to 50mm2 chip reduction (no bulk coherence engines, less 4 HT links PHYs etc)

then comparing with IB which is 22nm, the 28nm SOI is already a very very good size reduction for a optical shrink...

28nm http://www.globalfoundries.com/technology/32-28nm.aspx

by Hans de Vries from chiparchitect

Code: Select all

 ------------  Intel 45nm   Intel 32nm   AMD 45nm  IBM/AMD/FS 32nm *IBM/AMD/FS 28nm   

SRAM cell size  0.346 μm2    0.171 μm2     0.37 μm2      0.149 μm2      0.120um2
NMOS drive cur. 1.36 mA/μm   1.55 mA/μm    1.21 mA/μm    1.55 mA/μm
PMOS drive cur. 1.07 mA/μm   1.21 mA/μm    0.78 mA/μm    1.22 mA/μm

 

* ->my addition

... 28nm "SOI" SHP as i "suspect" it will be(can be wrong of course, but it seems 22nm SOI was canned, and after 28nm will be 20nm UTB-SOI for both SHP and LP) for vishera [perhaps].. if not, Steamroller then for sure... something like 20-25% size reduction, then a proper BD desktop chip CAN BE (314-50)/1.22 = 216mm2 and clock above 4.5Ghz base...

[mmarq]

If 20nm can represent at least the "usual" 40% reduction in size compared with 28nm, another point for not needing bobcat CPUs, is that 20nm UTB-SOI can be very well tweaked for low power... some of its best features, same basic process CAN BE tweak for low power and or high performance... and then at 20nm a proper BD desktop chip with 16Mb of L2/L3 that could be in 2 years time, would be something like 216*0.4= 86mm2 -> that is almost the size of the now 40nm bobcat -> something like a trinity fusion chip could be around the 50mm2 size ... and less than 10W in a super binned part like there will be 17W march/april 2012 ...

[mmarq]

I did not find comparable data, but looking at the electrical specifications for intel Sandy Bridge 2700K, I discovered that they allow a lot of current for their CPU.
- 200A on Vdd with a voltage between 0.7 and 1.3 (140W to 260W)
- 2A on Vpp (?) with a voltage around 1.0 (2W)
- 32A on V?? with a voltage of 1.1V (35.2W)
- 32A on V?? with a voltage of 0.9V (28.8W)
140+2+35.2+18.8=196W
260+2+35.2+18.8=316W
So, intel has a maximum (electrical) wattage specified of 316W. To convert that in TDP one should subtract a little bit of this value for the real work done by the processor. So let's assume a mean value of 250W.

The same might apply to 2500-2600k.. its only a question of tweaking power management, not allowing going above TDP or something, putting that power budget on the base frequency. after all its only 100Mhz more...

[hyc]

So, intel has a maximum (electrical) wattage specified of 316W. To convert that in TDP one should subtract a little bit of this value for the real work done by the processor. So let's assume a mean value of 250W.

No, you cannot subtract anything for the work done by the processor. Power in == power out, and it all turns into heat in the end.

[Kedas]

A good comparison from a design point of view would be Gulftown (i7-980X) against FX-8150.
They have about the same number of transistors (both 1.2B) and about the same power consumption, both on 32nm (although not the same process technology)
Maybe Gulftown should be set back from 3.3Ghz to 3Ghz to cancel out some cherry picking from intel (since I assume there is more than a market reason to have a much bigger price.)

[gallier2]

So, intel has a maximum (electrical) wattage specified of 316W. To convert that in TDP one should subtract a little bit of this value for the real work done by the processor. So let's assume a mean value of 250W.

No, you cannot subtract anything for the work done by the processor. Power in == power out, and it all turns into heat in the end.

Damn, I should have put a smiley, that snippet was tongue in cheak, of course the real work (in physics term) is negligible. The electrons moving around don't need that much cinetic energy.

And granted it was quite unclear because I took the mean afterwards, but the means was not justified by the work done, but by the voltage range specified.

[mmarq]

A good comparison from a design point of view would be Gulftown (i7-980X) against FX-8150.
They have about the same number of transistors (both 1.2B) and about the same power consumption, both on 32nm (although not the same process technology)
Maybe Gulftown should be set back from 3.3Ghz to 3Ghz to cancel out some cherry picking from intel (since I assume there is more than a market reason to have a much bigger price.)

No they have not, while the i7 980x is ~14MB total cache, BD is close to 17MB total cache.. and BD(orochi) has more 4 I/O links and more misc I/O and more bulky "directory" kind of coherency engines... and more buffers allover, including a larger Branch Target Cache(buffer)...



See the picture ?.. that comparatively small rectangle on the bottom has 213million trans., actually more because there are there [in the rectangle] elements of the system request interface and interconnect that don't belong to a module "per se"... and there are 4 of those modules, then 4 large L3 blocks, then the rest... why ppl buy into this kind propaganda floating around i can't find a reason!... what is the catch ? where does it come from ?

http://www.theregister.co.uk/2011/02/24 ... ore_isscc/

The Bulldozer module has 2 MB of L2 cache memory and has a total of 213 million transistors, according to McIntyre.

Lets suppose since 2 MB of L2 is almost half the size of a module, that each of those L3 blocks has ~100Milion trans... then there are 5 HT links, 2 misc I/O, bulky coherency engines, DCCLs for interlagos kind of chips, the all Northbridge (NB), last but not least the very large DDR3 DRAM interface(larger because of bulky cascaded trans.)

Modules: 213*4 = 852Mil
L3 : 100*4 = 400Mil
The rest: = 700-740Mil

Isn't that consistent with the image, if you take out the modules and the L3 it still leaves a lot of space ?

In the end BD is not 2B trans. .. but is very close.

[Kedas]

In the end BD is not 2B trans. .. but is very close.

AMD confirmed/corrected (a few days ago) that the transistor count is 1.2B not 2.0B
http://techreport.com/discussions.x/22100

You say they don't have the same cache size etc., yeah, so what, it's not like both weren't allowed to choose the best size according to their own ideas, if the other idea is better then they did chose poorly. I see it more like you get 1.2B transistors make the best of it within 130W.

[mmarq]

Well i think that test was done without the OS scheduler fix.. so results reflect this.

So the main point of this thread is this



In a very rough approach, both BD and SB have 8 threads:

BD 76.1 points at the same clock, means 9.513 points per thread, for SINGLE THREAD per module, and since thread can scale to 70%(for now) in CMT, means a single thread can catch at least 30% more of the shared resources and SINGLE THREAD 9.513+(9.513*0.3) = 12.367 points NOW

Since the OS fix gets ~10% more, but lets say 7% overall for this, then BD will have 12.367*1.07= 13.233 points per single thread core cluster....

And Piledriver might get 10% more, but lets say 7% overall(cache/pre-fetch issues, path tweak, IMC fixes) 13.233*1.07 = 14.159 points per single thread core cluster for PD
-------------

SB 69.64 points at the same clock, means 8.705 points per thread, for SINGLE THREAD per module, and since thread can scale to 35%(Intel props over 40% running around, so i'm being nice to them here, proping instead single thread for each core) in SMT(hyperthreading), means a single thread can catch at least 65% more of the shared resources and SINGLE THREAD 8.705+(8.705*0.65) = 14.363 points NOW
-------------

With the OS fix, actual BD per "core/cluster" will only be 14.363/13.233 = 8.5% of the single thread performance of a SB core...

With Piledriver most probably that advantage will be reduced to 14.363/14.159 = 1.5% of the single thread performance of a SB core...

So i don't now why ppl have such hated debates about single thread, when it needs only a small clock advantage to completely overrun SB.. clearly with PD will be i think...

Or IB is really 20% over SB, which is the always the same Intel says, including when from Nehalem to SB, which is clear now its not even half of that for most workloads... AMD doesn't even have to push clocks much higher to start to claim wins allover the spectrum...

Is this going to happen as example in tests on AT or other "regular" review sites ?... not even in dreams!..

AMD to Anandtech:
what about some test with GCC and code that is equal tuned for both ? - ummm.. Nope!

what about some tests which the code include XOP and or FMA4 ? - ain't gonna happen!

what about some nice multithreading test not tweaked for Hyperthrading ? - fogetaboutit !

is there actually any test that AMD can show its features and win ? - yes, at power consumption, he have already a nice PSU to make sure this happens, and with a large margin measured at the wall!..

[Smartidiot89]

2500K lacks Hyper-threading

[mmarq]

In the end BD is not 2B trans. .. but is very close.

AMD confirmed/corrected (a few days ago) that the transistor count is 1.2B not 2.0B
http://techreport.com/discussions.x/22100

You say they don't have the same cache size etc., yeah, so what, it's not like both weren't allowed to choose the best size according to their own ideas, if the other idea is better then they did chose poorly. I see it more like you get 1.2B transistors make the best of it within 130W.

This last part is a BD 4.2Ghz base... or more if possible, and ASAP...

But still can't understand the catch ?... if so, and as it says in the article you point, then the "estimative" for a module being 213Million is clearly VERY exaggerated and still not include any of the system request interface for the module, neither any other more things... then the size makes sense again for 1.2B...

OTOH find very strange AMD doesn't know exactly how many trans the chip has without making estimatives ... doesn't the all computerized tools of DFM already count those to the last single one of them ?

was AMD inventing an excuse for a much larger die than expected ?

What is the catch with that count ?... in the end it doesn't matter a bit to users !

[mmarq]

2500K lacks Hyper-threading

gee!.. you might be absolutely right... shows how little i've been looking upon Intel lately, not my fault, tests around showing only less than 10% improvement "at the same lock" over Nehalem taken as reference lostcircuits results.. is the culprit!... actually i posted a long post with all that in here long time ago (sorry not digging for it now)...

Nevertheless my wonder "with that rough approach" was how much multithreading has to do with those numbers, how much does that Sandra test multithreads?... is it more than 4 threads ? .. i know it multithreads and and also has SMP support(MP/MT support).. but how much ?... because if its only 4 threads then:

BD: 76.1/4 = 19.03 points
SB: 69.64/4 = 17.41 points

.. which means BD core/cluster as is, no OS fix, no PD, is already 9.3% higher single thread IPC than a SB core, and we have every reason to doubt that is so.

If its 8 threads and scales linearly, but the code has double path and puts only 4 threads on Intel targets(detects if HT is on or off), then:

BD: 76.1/8 = 9.513 points
SB: 69.64/4 = 17.41 points

.. which means BD core/cluster as is, no OS fix, no PD, is 83% LOWER single thread IPC than a SB core, and we have every reason AND THEN SOME to doubt VERY MUCH that is so.

If its 8 threads for both scaling similarly, then:

BD: 76.1/8 = 9.513 points
SB: 69.64/8 = 8.705 points

.. which means BD core/cluster as is, no OS fix, no PD, is again 9.3% higher single thread IPC than a SB core...suggesting that CMT is almost 100% kick over a single core( from -83 to +9.3) .. which is also very highly doubtful.

And we might add that Sandra Packs are not in any way optimized for AMD... the other way around might be much more correct actually.. so i think my first numbers are much more close to the truth than either of this last ones, and BD has more single thread (IPC) potential than ppl are screaming about.

No! that BD single thread IPC is not awesome, no by a long shot now, perhaps not even a champion or something(single thread) any time in future, no matter how we stretch the views with possibilities.. but does it really matter ?... isn't multithreading scaling orders of magnitude better than IPC ? .. ISN'T IT THE FUTURE ?... isn't this what this BD design reveals ?

[ actually if i were granted a view and an opinion for future iterations, it would be SMT for each core/cluster of a module and 4 threads for each module(yes FX with 16 threads).. and possibility of spMT in the future, making single thread performance completely irrelevant.. but of course the proverbial 800pound gorilla putting a full brake at MT future might weight much in the decisions also... ]

[abinstein]

Why do people who buy enthusiast grade CPUs care about a few watts of power consumption?
If one really cares about power, then compare these chips in server settings on whole system power draw.
All I see who against the conclusions from these numbers are apologists who demand every possible excuses that Intel's inferior solution could have.

BTW marketing crap on non-{Fusion,Bocat,Bulldozer} or immaterial products have been and will be moved to the General board on sight.

[Stimpy]

The only fair test in a server environment is the power consumption for the entire platform.
A customer will typically go to a vender based on system requirements. I.e. the business will tell the IT department they need a server with X amount of performance (support Y number of users with an average response time of Z).
The vendor will then put together a platform that meets those specifications. The IT department will then look at the power consumption of the server.

The CPU is really only going to be a small proportion of the overall power draw. You'll have to consider, power supply efficiency (you will have two of them in a good server).
Memory/Chipset/internal cooling/disk drives/network cards/GPU cards.

In the server world we should not have to care about the CPU power consumption. That should be left to the system designers.