AMD Readies FX-8370, FX-8370E Microprocessors.

[Enigmoid]

3D PM is one of the very few tests where a PII x6 can demolish a 8350.

93 vs. 73 points singlethread despite significantly lower clocks on the PII. 430 vs. 400 for MT.

Looking at the 2500k vs. x6 1100T BE 3D particle movement is exactly where it should be.

http://www.anandtech.com/bench/product/203?vs=288

It simply looks like AMD's recent architectures haven't place much emphasis on FP while intel has. Atom is ungodly strong on that test.

HW i3-4360 @ 3.7 ghz gets 123 in the ST test. 33.2 points/Ghz.

Now look at that again on this weird benchmark. Haswell clock for clock is only 24% faster than kabini. Haswell clock for clock is 23% faster than atom. In fact looking at something like the Q9400 gives 87 points or 32.7 points/Ghz. There is almost no IPC gain C2Q to HW (seemingly like Fritz chess).

Why?

Because it looks like there are a ton of cache misses or stalls in the pipeline and it appears that nothing intel has done recently has improved this. Its one of those test where HT is able to gain 80% (4770k vs 4670K).

Its not a great benchmark but neither is it really unfair. Its just quirky.

Edit: That said the FX CPUs tend to give great performance at good prices. If they can drop the tdp and upgrade the platform they are looking at something much more compelling. 8370 and 8370E are steps forward.

 

[Phynaz]

They are professionals, how could they not know what amateurs know ..??.

One could only be surprised, unless of course that they are ignorant of what is published on other sites, also how could they come to this conclusion based on a single bench that nobody knows .?

Would be interesting to know who is the editor/conceptor of this bench and what it does actualy, as said my opinion is that it s a memory bandwith bench that is not representative of actual FP perfs.

For the Google challenged http://www.borandi.co.uk/3DPM

 

[monstercameron]

3D PM is one of the very few tests where a PII x6 can demolish a 8350.

93 vs. 73 points singlethread despite significantly lower clocks on the PII. 430 vs. 400 for MT.

Looking at the 2500k vs. x6 1100T BE 3D particle movement is exactly where it should be.

http://www.anandtech.com/bench/product/203?vs=288

It simply looks like AMD's recent architectures haven't place much emphasis on FP while intel has. Atom is ungodly strong on that test.

HW i3-4360 @ 3.7 ghz gets 123 in the ST test. 33.2 points/Ghz.

Now look at that again on this weird benchmark. Haswell clock for clock is only 24% faster than kabini. Haswell clock for clock is 23% faster than atom. In fact looking at something like the Q9400 gives 87 points or 32.7 points/Ghz. There is almost no IPC gain C2Q to HW (seemingly like Fritz chess).

Why?

Because it looks like there are a ton of cache misses or stalls in the pipeline and it appears that nothing intel has done recently has improved this. Its one of those test where HT is able to gain 80% (4770k vs 4670K).

Its not a great benchmark but neither is it really unfair. Its just quirky.

what point is there using such a weird benchmark?

 

[monstercameron]

They are professionals, how could they not know what amateurs know ..??.

One could only be surprised, unless of course that they are ignorant of what is published on other sites, also how could they come to this conclusion based on a single bench that nobody knows .?

Would be interesting to know who is the editor/conceptor of this bench and what it does actualy, as said my opinion is that it s a memory bandwith bench that is not representative of actual FP perfs.

Edit: For those wondering why AMD doesn't score so high... AMD architecture loves integer (whole number) operations, especially when parallel. This benchmark was written in floating point (FP) operations, which is the usual standard for scientific programming depending on the algorithm.

http://www.overclock.net/t/1413580/3d-particle-movement-benchmark

without the ad-hominem...

At PJs request I attempted to make a version more memory dependant, but the nature of these algorithms is such that they are best ported to GPU. So GPU version inbound soon, using C++ AMP to make it vendor agnostic.

amd cpus might struggle but apus could theoretically get a massive boost.

 

[Enigmoid]

what point is there using such a weird benchmark?

Not much. But it its typical of what you would actually program or run then it may be relevant.

For small research groups you would be surprised at the computing waste that goes on. Many programs can be run orders of magnitudes faster but aren't because that would require a recompile and substantial rewrite of the program; the pains of legacy and nobody wants to or has the time and skills to do so.

 

[Abwx]

There is almost no IPC gain C2Q to HW (seemingly like Fritz chess).

Why?

Because it looks like there are a ton of cache misses or stalls in the pipeline and it appears that nothing intel has done recently has improved this. Its one of those test where HT is able to gain 80% (4770k vs 4670K).

Its not a great benchmark but neither is it really unfair. Its just quirky.

Edit: That said the FX CPUs tend to give great performance at good prices. If they can drop the tdp and upgrade the platform they are looking at something much more compelling. 8370 and 8370E are steps forward.

I cant answer , as you point it this bench yield a ton of odd results, from the underhelming C2D to HW progress one could assume that recent FP related instructions are not used at all.

On the the other hand CB ,3DS and Povray show very good gainS for the latter so it s obvious that 3DP doesnt provide anything usefull to estimate FP perfs even within a same brand.

 

[Abwx]

http://www.overclock.net/t/1413580/3d-particle-movement-benchmark

without the ad-hominem...


amd cpus might struggle but apus could theoretically get a massive boost.

They are completely off and are obviously clueless, just repeating hearsay.

Kabini is better than BT in FP in both ST and MT , on MT FP 8350 is as good as a 2600K and often better, yet with 3DP the 2600K does 63% better in MT , a pure joke of a bench, let s compare these two CPUs with CB 11.5, Povray and 3DS max :

http://techreport.com/review/23750/amd-fx-8350-processor-reviewed/12

http://www.hardware.fr/articles/880-7/rendu-3d-mental-ray-v-ray.html

 

[bunnyfubbles]

some serious tinfoil needed for this thread

 

[Abwx]

I did read that 6 months ago 30% of all CPUs/APUs sold by AMD were FXs,
a gross calculuus point to the FX detaining 10% of the DT market so it s no surprise that new revisions are released, i suspect that AMD didnt expect this line to be so successfull almost two years after it was released, there might be a lack in guidance as the numbers would had justified a new SKU using Steamroller cores, even if frequency had to be reduced to 4GHz levels since they would had gained in TDP.

 

[Enigmoid]

I did read that 6 months ago 30% of all CPUs/APUs sold by AMD were FXs,
a gross calculuus point to the FX detaining 10% of the DT market so it s no surprise that new revisions are released, i suspect that AMD didnt expect this line to be so successfull almost two years after it was released, there might be a lack in guidance as the numbers would had justified a new SKU using Steamroller cores, even if frequency had to be reduced to 4GHz levels since they would had gained in TDP.

I think that more than Steamroller FX CPUs AMD needs to revamp AM3+. They can probably drop at least 10-20W off load/idle, almost the same as if they moved to steamroller. Steamroller would be nice though.

 

[DrMrLordX]

some serious tinfoil needed for this thread

Benchmarks are weird, and Bulldozer + Piledriver are even more weird. Most applications are compiled with the existing computing paradigm in mind; in other words, people write software for the dominant player in the previous generation, either consciously or unconsciously (much of what is written is determined by the available compiler options, and if most of them produce code that favors Intel processors by default, then people will code for Intel processors without even really thinking about it). There is a lot of software out there written either with Core 2 Duo/Quad or Nehalem in mind.

Anyone, AMD or Intel, who attempts to produce hardware running this software must keep this fact in mind when producing their product. Any attempt to break away from the existing computing paradigm brings with it the prospect that existing software will not run well on said new product. Take a look at what Intel went through with the Pentium 4. Now, fast-forward years in advance and look at how the P4 fares against, say, Athlon 64s in modern benchmarks. Notice any differences? We've had generations of software optimizations for Core 2 and Nehalem. All those little pieces of the Netburst legacy that remained in Intel's designs are making a difference, albeit years later than Intel needed it for Netburst.

The Athlon 64, which did an excellent job of running code more-or-less optimized for the Pentium III, has not aged as well.

Just as it was Intel's responsibility to provide software authors every possible tool to improve performance on their exotic Netburst designs, it is imperative that AMD provide the same support for the construction cores (especially Piledriver). The first people that should have seen something wrong with that 3D PM bench should have been an AMD PR guy, who should then have contacted both Anandtech and borandi to sort out what went wrong.

Also, where are the SPEC numbers for the processors in question (FX, Haswell, Kabini, Bay Trail)? If we really want to discuss differences in FP performance, then maybe that would be a good place to start?

 

[WhoBeDaPlaya]

Was playing around with a bunch of FX 8320s this weekend, and got to benching the various systems around the house (all running the exact same setup of Windows), including a Clevo w230ss with i7 4810MQ.
I was impressed at how much power it has!

Cinebench R15 [multi, single]

i5 2450P @ 3.6GHz (3.9GHz single-thread)		486	137
i5 2500K @ 4.5GHz (no turbo)				603	156
i7 4770K @ 4.55GHz (no turbo)				902	182
FX 8320	(stock)						561	88
i7 4810MQ (stock)					636	140
Phenom II X6 1045t (stock)				414	87

x264 Stress Test v2

i5 2450P @ 3.6GHz (3.9GHz single-thread)		2.12 fps
i5 2500K @ 4.5GHz (no turbo)				2.58 fps
i7 4770K @ 4.55GHz (no turbo)				3.64 fps
FX 8320	(stock)						2.75 fps
i7 4810MQ (stock)					2.66 fps
Phenom II X6 1045t (stock)				1.72 fps

Yes, this was run on Win7 so the FX might have been hobbled, but it's the OS I use.

 

[Abwx]

I think that more than Steamroller FX CPUs AMD needs to revamp AM3+. They can probably drop at least 10-20W off load/idle, almost the same as if they moved to steamroller. Steamroller would be nice though.

Looking at Hfr measurements the plateform power comsumption is 189W(the GC account for 18W out of these 189W) at the main for 116.4W on the 12V rail, that s 44W losses at the ATX connector level and is quite a lot since the losses of the VRMs are still not accounted, so they have a low hanging fruit that is waiting to be harvested.

From the 7850K review it looks like a 95W 8C is doable at 3.7Ghz, i have asked a few question about this to the reviewer since he kindly answer in their forum to any one that seek some info , i ll post his comments once he give his opinion.

Most applications are compiled with the existing computing paradigm in mind; in other words, people write software for the dominant player in the previous generation, either consciously or unconsciously

I downloaded 3DPM , the copyright is owned by Ian Cutress, so this is an Anandtech in house bench.

 

[DrMrLordX]

I downloaded 3DPM , the copyright is owned by Ian Cutress, so this is an Anandtech in house bench.

Ian Cutress = borandi?

Regardless, my point still stands, though you can always contact Ian yourself and hash it out with him.

 

[Abwx]

Ian Cutress = borandi?

Regardless, my point still stands, though you can always contact Ian yourself and hash it out with him.

I ll wait for the 8370E review to eventualy ask him a few infos, would be great that he post on the forum sometime, that s what i appreciate at Hardware.fr were the reviewers will always answer to the members at the bottom of the articles pages but also in their forum.

 

[jhu]

Anyone, AMD or Intel, who attempts to produce hardware running this software must keep this fact in mind when producing their product. Any attempt to break away from the existing computing paradigm brings with it the prospect that existing software will not run well on said new product. Take a look at what Intel went through with the Pentium 4. Now, fast-forward years in advance and look at how the P4 fares against, say, Athlon 64s in modern benchmarks. Notice any differences? We've had generations of software optimizations for Core 2 and Nehalem. All those little pieces of the Netburst legacy that remained in Intel's designs are making a difference, albeit years later than Intel needed it for Netburst.

The Athlon 64, which did an excellent job of running code more-or-less optimized for the Pentium III, has not aged as well.

Would you provide links to more recent benchmarks between Pentium 4 and Athlon 64? From my own tests, Pentium 4 still lags behind Athlon 64 with regard to IPC.

BTW, Core2 and Nehalem have their roots with the P6. Pentium 4 has some optimizations that don't apply to Core 2/ Nehalem.

 

[Erenhardt]

I was reading in Anandtech CPU review that some benchmarks are designed to take additional loop on amd CPUs compared to intel.

 

[DrMrLordX]

Would you provide links to more recent benchmarks between Pentium 4 and Athlon 64? From my own tests, Pentium 4 still lags behind Athlon 64 with regard to IPC

.

We could possibly build a useful data set from some articles that at least deal with the question at hand tangentially. Okay, here's my best shot:

First, we have this small repository of Cinebench 11.5 scores. Two scores worth noting here:

Dirk .44 points Athlon 64 3700+
Playitlouder .42 points Athlon 64 3200+

Now here is a YouTube video of a guy pulling .52 points with a Pentium 4 640 (HT enabled) @ 3.2 ghz.

Now all we need are some Cinebench 2000/2003 numbers to round out this little competition. How about this article here. Notice the following results:

Pentium 4 640
Single-thread: 280
Multi-thread: 337

Athlon 64 3800+
Single-thread: 339

Athlon 64 3200+
Single-thread: 283

Hmm! Somewhat inconclusive, though if we assume that the numbers for the 3700+ and 3800+ would be fairly close . . . well, let's take a look here:

Athlon 64 3700+
Single-thread: 336

Athlon 64 3800+
Single-thread: 338

Okay, so we can say that, at least when dealing with the 2.4 ghz s754 3700+(not the 2.2 s939 3700+, per se) that the 3700+ and 3800+ were very close in Cinebench 2003.

To narrow in on the relevant data:

Cinebench 2003
Pentium 640 (HT enabled): 337
Athlon 64 3700+(s754, 2.4 ghz): 336

Cinebench R11.5
Pentium 640 (HT enabled): .52
Athlon 64 3700+(s939, 2.2 ghz): .44

If we want to correct for the loss of clock speed moving to the 939 version of the CPU, perhaps we can give the s754 version of the processor a score of .46 (look at playitlouder's score of .40 @ 2 ghz).

Even if we give the Athlon 64 an R11.5 score of .46, the 640 has leapfrogged it over the years. In Cinebench 2003, it was mostly a dead heat, but in R11.5 the Pentium 4 now has a ~%13 advantage.

Some people look at this and scream bias, but that brings me back to my original point. Was there a conspiracy to slowly but surely make the Pentium 4 beat the Athlon 64 in future versions of Cinebench? Probably not. They (those maintaining Cinebench) probably looked at the huge installed base of Intel hardware out there and the plethora of Intel-provided/optimized compiler tools and proceeded accordingly. It just so happens that changes in Cinebench software favored the Pentium 4 over the years.

BTW, Core2 and Nehalem have their roots with the P6. Pentium 4 has some optimizations that don't apply to Core 2/ Nehalem.

This is true, though Intel did carry over their bus architecture to Core 2, among other things. There were some things they obviously didn't want/need seeing as how they were no longer pushing such a long instruction pipeline.

 

[Abwx]

.

Was there a conspiracy to slowly but surely make the Pentium 4 beat the Athlon 64 in future versions of Cinebench? Probably not. They (those maintaining Cinebench) probably looked at the huge installed base of Intel hardware out there and the plethora of Intel-provided/optimized compiler tools and proceeded accordingly. It just so happens that changes in Cinebench software favored the Pentium 4 over the years.

Check anand s old reviews, at equal frequency a core 2 does 10% better than the A64 in the previous version of CB, with the newly released CB11.5 the gap increased to 28%, so obviously something happened.

Same with CB R15, the AMD and Intel CPUs get about the same difference in ST but on MT the gap increased in favour of Intel when comparing to 11.5, so it s possible that bandwith requirements increased but on another hand CB used ICC starting from 11.5.

 

[Lepton87]

.

We could possibly build a useful data set from some articles that at least deal with the question at hand tangentially. Okay, here's my best shot:

First, we have this small repository of Cinebench 11.5 scores. Two scores worth noting here:

Dirk .44 points Athlon 64 3700+
Playitlouder .42 points Athlon 64 3200+

Now here is a YouTube video of a guy pulling .52 points with a Pentium 4 640 (HT enabled) @ 3.2 ghz.

Now all we need are some Cinebench 2000/2003 numbers to round out this little competition. How about this article here. Notice the following results:

Pentium 4 640
Single-thread: 280
Multi-thread: 337

Athlon 64 3800+
Single-thread: 339

Athlon 64 3200+
Single-thread: 283

Hmm! Somewhat inconclusive, though if we assume that the numbers for the 3700+ and 3800+ would be fairly close . . . well, let's take a look here:

Athlon 64 3700+
Single-thread: 336

Athlon 64 3800+
Single-thread: 338

Okay, so we can say that, at least when dealing with the 2.4 ghz s754 3700+(not the 2.2 s939 3700+, per se) that the 3700+ and 3800+ were very close in Cinebench 2003.

To narrow in on the relevant data:

Cinebench 2003
Pentium 640 (HT enabled): 337
Athlon 64 3700+(s754, 2.4 ghz): 336

Cinebench R11.5
Pentium 640 (HT enabled): .52
Athlon 64 3700+(s939, 2.2 ghz): .44

All of that is because of HT, compare dual core versions and p4(pentium D) will look as bad now as it did back then. Also compare them in games, same thing, A64X2 demolishes Pentium D

 

[DrMrLordX]

Check anand s old reviews, at equal frequency a core 2 does 10% better than the A64 in the previous version of CB, with the newly released CB11.5 the gap increased to 28%, so obviously something happened.

Same with CB R15, the AMD and Intel CPUs get about the same difference in ST but on MT the gap increased in favour of Intel when comparing to 11.5, so it s possible that bandwith requirements increased but on another hand CB used ICC starting from 11.5.

Yes, something did happen. Software changed over different versions, and the people coding/maintaining the benchmark decided to take advantage of some compiler optimizations. I haven't seen any evidence that Intel paid people off or soft-bribed anyone to make these changes happen. They might have, but nobody has come forward to talk about it.

The real problem here is that there is no AMD compiler or visible program run by AMD to convince groups that maintain these oft-cited benchmark programs to at least make an alternative binary for AMD processors. Bulldozer and Piledriver really could have used some help in that department. It has been up to the enthusiast community to carry the banner for AMD. By this point, their arms must be getting tired.

All of that is because of HT, compare dual core versions and p4(pentium D) will look as bad now as it did back then. Also compare them in games, same thing, A64X2 demolishes Pentium D

Well, yes, but HT was and still is very much a part of the Netburst architecture. HT couldn't put the P4 noticeably ahead of A64 in Cinebench 2003, but it does the job in R11.5.

Here's an interesting roundup from 2011 showing a whole mess of CPUs, old and new, in some benchmarks that were new to 2011. To complement it, let us also examine a review of the Pentium D 965EE.

Cinebench 2003
Pentium D 965EE
Single-threaded: 79.7
Multi-threaded: 36.9

Athlon 64 4800+:
Single-threaded: 76.6
Multi-threaded: 41

Cinebench R11 (not 11.5, but close)
Pentium D 965EE: 989
Athlon 64 4800+(90nm): 1036

The 965EE did not gain any ground . . . in fact, it lost some. I remember Smithfield having die-interconnect problems which severely hampered performance, but I find it interesting that we see Presler getting worse over time in contrast with Prescott's improvements. Maybe things would look different had the roundup used 11.5 instead of 11?

edit: For my part, I think this line of inquiry can be put to rest. Various comparisons between the awkwardness of Netburst and AMD's CMT architecture may be interesting, but the subject has gone too far off-topic for me to want to continue it any further. Let me conclude by saying that it is possible that Piledriver will wind up looking a lot better from a historical perspective in the future, especially if/when software becomes more heavily-threaded, but I do not think the compiler advantage Intel enjoys will do anything to help Piledriver in the future.

I am sure chips like the upcoming FX-8370 will be great for "everyday tasks" far into the future, regardless, provided it has good I/O to back it up.

 

[ShintaiDK]

Cinebench is mainly due to SSE2 in "modern" versions.

The FX (And the uarchs behind) will only be worse day by day. In "everyday" task the FX is already beaten by entry/lowend chips.

The only way for the FX series to even mature in any gentle way would be that there came a revolution in multithreading that would be completely unheard of. And that all companies right away jumped that wagon so 1-2 thread performance would become irrelevant more or less.

 

[FlanK3r]

bullshits...

Its about experience, Im every day testing CPus (AMD, Intel), boards and OC and tweaking is my hobby. I know some tricks how to get more points with FX in 3D Marks example and whats wrong ussually in some reviews if the score is not legit with others.

The problem of FXs is FPU, cache latencies, some single thread (not all situations)... But you and me know, the FX series 15h is ending. No more 15h in future. 4 years old cyclus ending. FX is effective in multithread as multiplication of performance effective. Of course, not overall.
The idea is good, but results is in half hearted .

Im sure AMD team know it and working some months at new microarchitecture.

 

[monstercameron]

bullshits...

Its about experience, Im every day testing CPus (AMD, Intel), boards and OC and tweaking is my hobby. I know some tricks how to get more points with FX in 3D Marks example and whats wrong ussually in some reviews if the score is not legit with others.

The problem of FXs is FPU, cache latencies, some single thread (not all situations)... But you and me know, the FX series 15h is ending. No more 15h in future. 4 years old cyclus ending. FX is effective in multithread as multiplication of performance effective. Of course, not overall.
The idea is good, but results is in half hearted .

Im sure AMD team know it and working some months at new microarchitecture.

as in more than one?

 

[FlanK3r]

I know Shintai from older days. He dont like AMD many years and criticizes "all" from AMD. I agree with some stuffs but not with all. I like AMD slightly more but I like Intel CPus too of course (5960x is really fan for benchamrking ).