FX 8370 Review

[Abwx]

Povray does not use FMA. The issue, I suspect, is that Povray for Windows is compiled with MSVC. Also, it's running on Windows. Things look a bit different with gcc/icc on Linux

Seems that it s GCC compiled and that FMA flags are enabled for BD, i did read that recent beta versions are reverting to FMA4 for all AMD CPUs, the 8150 has only FMA4 and it can explain why it performs better IPC wise than the 8350 wich use perhaps FMA3, possible that Computerbase used an earlier version for the 8150 and an updated one using FMA3 for the 8350, that said only these instructions could explain that a i5 HW has better MT FP IPC than a i7 IB, the single thread chart show HW having about 30% better ST FP IPC than a SB and both have AVX so it s due to other instructions.

 

[jhu]

Seems that it s GCC compiled and that FMA flags are enabled for BD, i did read that recent beta versions are reverting to FMA4 for all AMD CPUs, the 8150 has only FMA4 and it can explain why it performs better IPC wise than the 8350 wich use perhaps FMA3, possible that Computerbase used an earlier version for the 8150 and an updated one using FMA3 for the 8350, that said only these instructions could explain that a i5 HW has better MT FP IPC than a i7 IB, the single thread chart show HW having about 30% better ST FP IPC than a SB and both have AVX so it s due to other instructions.

Again, Povray does not use FMA. It doesn't even use AVX to it's full capacity because that's not how it was written. At best it uses AVX-128 as scalar entities (which, to Povray, is no different than SSE2).

As for Haswell vs. Ivy Bridge, Haswell really is that much faster on architecture optimized code. I was surprised by this too. But this has nothing to do with FMA or AVX2 because, again, POVRAY DOES NOT USE THESE INSTRUCTIONS.

 

[Abwx]

Again, Povray does not use FMA. It doesn't even use AVX to it's full capacity because that's not how it was written. At best it uses AVX-128 as scalar entities (which, to Povray, is no different than SSE2).

As for Haswell vs. Ivy Bridge, Haswell really is that much faster on architecture optimized code. I was surprised by this too. But this has nothing to do with FMA or AVX2 because, again, POVRAY DOES NOT USE THESE INSTRUCTIONS.

Changes between 3.7.beta.RC3 and 3.7.RC4

- Added AMD optimizations for noise (needs FMA4)

http://www.povray.org/beta/

Edit :

Computerbase Povray bench version :

POV-Ray 3.7 RC7

http://www.computerbase.de/2014-09/amd-fx-8370e-im-test/6/

How could you explain that the FX8150 has better MT IPC than a 3770K in their povray s numbers.?.

 

[Ramses]

If your gaming you'd be as well as buying an i5 tbf

I like AMD but you'd be as well as buying haswell i5

I can absolutely guarantee that if/when I start seeing things being annoyingly slow, games or otherwise, I'll hop right online and buy up whatever it takes to make them not-slow.
That hasn't happened yet, but I can likewise guarantee that when this happens, next week, or five years from now, I will necro this thread back from the dead and comment on it.
So place your bets now gentlemen..

 

[Ranulf]

People are seriously using Dead Rising 3 as a reliable benchmark for anything?

 

[jhu]

http://www.povray.org/beta/

Edit :

Computerbase Povray bench version :

http://www.computerbase.de/2014-09/amd-fx-8370e-im-test/6/

How could you explain that the FX8150 has better MT IPC than a 3770K in their povray s numbers.?.

I just did an objdump on the binaries I compiled with gcc (-march=bdver1 and -march=bdver2). No FMA instructions present in the binary.

 

[jhu]

http://www.povray.org/beta/

Edit :

Computerbase Povray bench version :

http://www.computerbase.de/2014-09/amd-fx-8370e-im-test/6/

How could you explain that the FX8150 has better MT IPC than a 3770K in their povray s numbers.?.

Did you see the link for the tests I ran? Here it is again. I'll even give the relevant parts:

FX 8350 (4m/8t, 4.1 GHz turbo):        1985.83 pps ; 121.09 pps/module/GHz
Core i5 3317U (Ivy Bridge)
  (2c/4t, 1.7 GHz, 2.4 GHz turbo):      573.68 pps ; 119.52 pps/core/GHz
Core i5 4570 (Haswell)
  (4c/4t, 3.2 GHz, 3.4 GHz turbo):     1540.24 pps ; 113.25 pps/core/GHz

An AMD module has 2 moderate IPC hardware threads. An Intel core has 1 high IPC and 1 low IPC threads. But add them up and the AMD module wins the IPC race for POVray at least. But look at the IPC for Haswell: 1 thread on Haswell is 95% of the performance of 2 threads on Ivy Bridge. Pretty impressive results for Haswell.

But that's not always the case. For example in Blender Cycles rendering

Core i5 3317U (stock, 2.4 GHz turbo, 2C/4T)
Ubuntu 14.04, Blender 2.71, official Blender binary
Time: 5 minutes, 20 seconds
648000 samples/s, 270000 samples/s/GHz, 324000 samples/s/core, 135000 samples/s/core/GHz

Core i5 2400S, 2.5 GHz, 4C/4T
Ubuntu 14.04, Blender 2.71, official Blender binary
Time: 3 minutes 15.67 seconds
1059743 samples/s, 423897 samples/s/GHz, 264936 samples/s/core, 105974 samples/s/core/GHz

FX 8350 (stock, 4.1 GHz turbo, 4M/8T)
FreeBSD 10, Blender 2.71, compiled with llvm 3.3
Time: 2 minutes 4 seconds
1672258 samples/s, 407867 samples/s/GHz, 418064 samples/s/module, 101966 samples/s/module/GHz

A dual-threaded AMD module's IPC is lower than even a single-threaded Sandy Bridge core. This may be due to the llvm compiler I used, but I haven't had time to test further.

 

[AtenRa]

People are seriously using Dead Rising 3 as a reliable benchmark for anything?

I dont believe anyone uses DR3 as a benchmark for everything. But some people wanted to say that AMD CPUs doesnt always perform well.

Well they havent seen that FX8350 using GTX780Ti is equal/faster than Haswell Core i5/7 paired with R9 290X. That smells not optimized drivers to me.

Also to note, FX8350 with GTX780Ti scores 50/85fps at 1280x800 when Core i7 3970X Overclocked to 4.9GHz with GTX780ti scores 57/79fps at 1200p.
So having a faster CPU at 1280x800 doesnt help at all at 1200p.

 

[Phynaz]

Also to note, FX8350 with GTX780Ti scores 50/85fps at 1280x800 when Core i7 3970X Overclocked to 4.9GHz with GTX780ti scores 57/79fps at 1200p.
So having a faster CPU at 1280x800 doesnt help at all at 1200p.

57>50

And it's not 1200p, it's 800.

 

[AtenRa]

57>50

And it's not 1200p, it's 800.

FX8350 + GTX780Ti at 1280x800 = 50/85fps

Core i7 3970X @ 4.9GHz + GTX780Ti at 1200p = 57/79fps

So, having a faster CPU at 1280x800 doesnt give you almost anything more at 1200p.

 

[Abwx]

Did you see the link for the tests I ran? Here it is again. I'll even give the relevant parts:

I saw them but you should specify the instructions set level, with such numbers we still dont have a clue.

But that's not always the case. For example in Blender Cycles rendering

A dual-threaded AMD module's IPC is lower than even a single-threaded Sandy Bridge core. This may be due to the llvm compiler I used, but I haven't had time to test further.

Same thing, was it compiled up to SSE3 or higher.?..

I notice in computerbase tests than a 4570 has 38% better MT IPC in Povray than a 2500K but the IPC difference in CB 11.5 is only 11%, so there s something at work that is more than simply a better execution of legacy code, this could be the case in CB given the score but not in Povray.

 

[Phynaz]

So, having a faster CPU at 1280x800 doesnt give you almost anything more at 1200p.

Only a nearly 20% increase in minimum frame rate :whiste:

 

[AtenRa]

Only a nearly 20% increase in minimum frame rate :whiste:

Well its 14% faster using a 4.9GHz 6-Core and we also dont know if that minimum is one time or continuous.

 

[SlowSpyder]

Only a nearly 20% increase in minimum frame rate :whiste:

Is it closer to 10% or 20%? :whiste:


Again, it is a difference on paper. If that is a difference worth making a fuss over, by all means do so. But my guess is that this is only a difference on internet forums and actual playability is very close.

 

[mrmt]

I wonder why we don't see people defending Gulftown or Sandy Bridge as a current gaming platform.

 

[SlowSpyder]

I wonder why we don't see people defending Gulftown or Sandy Bridge as a current gaming platform.

You never see people talking about Sandy Bridge's merits as a gaming CPU? Really?

 

[AtenRa]

Is it closer to 10% or 20%? :whiste:


Again, it is a difference on paper. If that is a difference worth making a fuss over, by all means do so. But my guess is that this is only a difference on internet forums and actual playability is very close.

Not to mention a 10% OC (4.4GHz) on the FX8350 will easily reach higher minimum fps closing the difference to almost zero.

 

[jhu]

I saw them but you should specify the instructions set level, with such numbers we still dont have a clue.

They're compiled for the specific processor.

Same thing, was it compiled up to SSE3 or higher.?..

Decompiling the blender binary shows no FMA instructions, but there is AVX.

I notice in computerbase tests than a 4570 has 38% better MT IPC in Povray than a 2500K but the IPC difference in CB 11.5 is only 11%, so there s something at work that is more than simply a better execution of legacy code, this could be the case in CB given the score but not in Povray.

If you don't believe the Povray numbers, it's easy enough to go download and compile the code yourself.

 

[frozentundra123456]

FX8350 + GTX780Ti at 1280x800 = 50/85fps

Core i7 3970X @ 4.9GHz + GTX780Ti at 1200p = 57/79fps

So, having a faster CPU at 1280x800 doesnt give you almost anything more at 1200p.

If I am reading the charts correctly, there is no test of the 8350 at 1200p, so the results from the low resolution tests might or might not be the same at 1200p. Really though I am disappointed that Game.gpu has gone to low resolution testing for cpus. It is interesting data but they should also include results at the "normal" gaming resolution of 1080 or 1200p. If given the choice, I prefer their previous method of testing the cpus at 1080p.

 

[Phynaz]

Not to mention a 10% OC (4.4GHz) on the FX8350 will easily reach higher minimum fps closing the difference to almost zero.

What would stop somebody from overclocking the Intel CPU and keeping (or increasing) the performance difference? You surely aren't suggesting that only AMD CPUs are overclocked??

Sorry, the goalposts are staying rooted right where they are.

 

[SlowSpyder]

What would stop somebody from overclocking the Intel CPU and keeping (or increasing) the performance difference? You surely aren't suggesting that only AMD CPUs are overclocked??

Sorry, the goalposts are staying rooted right where they are.

See post 210. The FX is at stock clocks, the i7 3970X is already at 4.9GHz. He's saying you could also overclock the FX to pretty well erase the 14% advantage in minimums the already overclocked Intel CPU has in this game.

 

[Phynaz]

Aha. I was looking at the chart Frozen quoted that had it at 3.5

Nevertheless AtenRa is attempting to spin the results.

 

[frozentundra123456]

Like I said in my previous post, comparing the different resolutions may or may not be valid. If one assumes the minimums would be the same at 1200p as at 800p, then the best comparison is between the 4670K and the 8350. There the i5 holds a greater than 50% edge in minimum frame rates, and both cpus are at stock, with the i5 having at least similar overclocking headroom as the 8350.

Really though, this is a poorly designed test, and I dont think we should be trying to extrapolate the data to results that the site did not give.

 

[Abwx]

They're compiled for the specific processor.

It could use up to SSE3 only and be compiled for the specific CPU, this doesnt tell us if all available instructions were used.

Decompiling the blender binary shows no FMA instructions, but there is AVX.

So you re using a case where BD doesnt have better FP perfs because its ISA doesnt use FP (FMA) instructions that improve perfs while the comparison is granted AVX, so the Blender exemple is quite moot since one CPU is handicaped at the compiling and instructions level.

If you don't believe the Povray numbers, it's easy enough to go download and compile the code yourself.

I didnt say that i dont believe your numbers but that it would have been useful to specify the ISA level, without this information comparisons are not very helpfull other than showing Linux in good light compared to W8.1.

 

[jhu]

It could use up to SSE3 only and be compiled for the specific CPU, this doesnt tell us if all available instructions were used.

That doesn't even matter for a renderer. I don't even know what your point is anymore.

So you re using a case where BD doesnt have better FP perfs because its ISA doesnt use FP (FMA) instructions that improve perfs while the comparison is granted AVX, so the Blender exemple is quite moot since one CPU is handicaped at the compiling and instructions level.

What are you even talking about? No FMA instructions are used. Both processors can use AVX. If you have issues with this, go write your own benchmark using FMA instructions.

I didnt say that i dont believe your numbers but that it would have been useful to specify the ISA level, without this information comparisons are not very helpfull other than showing Linux in good light compared to W8.1.

I already gave you what they were. If you don't like the results for these real-world programs, then go find your own and report back. Make sure you also post a link to the source code and compiler arguments.