[AMD] Details on Fiji VR/395x2 surface and launch month

[patrickjp93]

1. i don't have all day.  Just doing a quick comparison, but you will absolutely find the same results if you gathered benchmark data from other websites.

 

2. Yes, of course 256 bit with compression isn't as good.  That's the problem.  Maybe a 256 bit bus without the compression overhead would have less issues, but we're talking about real products not theoretical.

 

3. You see the same effect with the 980.  It's a Maxwell issue that is present on the 970 and 980 but not the TitanX (which retains 40% better performance than a 290x regardless of resolution)  This is obv because of the 384 bit bus on the TitanX removing the bottleneck that we see with the other two cards.

 

The fact that it's not present in the TitanX does support your argument that there is a point where it doesn't matter, but I think it's premature to say HBM won't make a difference going forward when we can clearly see Maxwell 256 bit cards being hamstrung by their memory bus.  

 

I'm sure someone said the same thing about GDDR5 back in the day or SATA 6.  Available bandwidth is always saturated eventually.  It might not matter that AMD is using HBM for the 390x, but you better believe it will be the standard eventually.

1)Make comparisons on valid data and premises. If you choose to waste your time on bad data, that's your issue.

2)According to these performance indices it is, and the numbers Nvidia quotes are best case scenarios when the entire frame buffer consists of highly compressible textures.

3)No, there's no issue with the 980. Has someone stopped following benchmark updates after driver releases? Also, your analysis is still highly flawed. It's much more about the raw core, ROP, and TMU count than the bandwidth because of reduced backpressure.

 

Again, Maxwell is far from hamstrung by the 256-bit bus for the core counts they have. With 50% more cores and resources, you can rebalance the load to eliminate backpressure, and then bandwidth can be a benefit, but there's a limit to the effects of that, which is why core overclocking still has a much higher effect.

 

Oh it'll be the standard for Intel, AMD, and Nvidia before the close of 2016, but that's got far more to do with scientific computing and HPC space competition.

[sgloux3470]

 

3)No, there's no issue with the 980. Has someone stopped following benchmark updates after driver releases? Also, your analysis is still highly flawed. It's much more about the raw core, ROP, and TMU count than the bandwidth because of reduced backpressure.

 

 

 

No, go look at some benchmarks.  The gap between the 290 or 290x and the 980 shrinks with resolution just like it does with the 970.  

 

 

Don't believe me?  Go look at some recent benchmarks containing the two.

http://www.techspot.com/review/991-gta-5-pc-benchmarks/page4.html

http://www.gamersnexus.net/game-bench/1905-gta-v-pc-fps-benchmark-graphics-cards

 

The TitanX works as a control, and it doesn't observe the same performance degradation and maintains the same relative performance of roughly ~35-40% over the 290(x) regardless of resolution. Both the 980 and 970 drop significantly as resolution increases from the 980 being around 20% faster than the 290x to running less than 1% faster at 4K ultra settings.  It's consistent across multiple games as well and across different benchmark websites.  

 

It's also present with the 960.  

http://www.hardwarecanucks.com/forum/hardware-canucks-reviews/68697-nvidia-gtx-960-reference-review-16.html

 

I'm just eyeballing, but it seems Maxwell loses roughly 7% relative performance for each bump up in resolution between 1080p-1440p-4K.  Usually you see the opposite effect where higher powered cards gain in their lead over less powerful ones as the lower ones can't keep up. (Like you do between the 290x and 7970)

[patrickjp93]

No, go look at some benchmarks.  The gap between the 290 or 290x and the 980 shrinks with resolution just like it does with the 970.  

 

 

 

Don't believe me?  Go look at some recent benchmarks containing the two.

http://www.techspot.com/review/991-gta-5-pc-benchmarks/page4.html

http://www.gamersnexus.net/game-bench/1905-gta-v-pc-fps-benchmark-graphics-cards

 

The TitanX works as a control, and it doesn't observe the same performance degradation and maintains the same relative performance of roughly ~35-40% over the 290(x) regardless of resolution. Both the 980 and 970 drop significantly as resolution increases from the 980 being around 20% faster than the 290x to running less than 1% faster at 4K ultra settings.  It's consistent across multiple games as well and across different benchmark websites.  

 

It's also present with the 960.  

 

http://www.hardwarecanucks.com/forum/hardware-canucks-reviews/68697-nvidia-gtx-960-reference-review-16.html

 

I'm just eyeballing, but it seems Maxwell loses roughly 7% relative performance for each bump up in resolution between 1080p-1440p-4K.  Usually you see the opposite effect where higher powered cards gain in their lead over less powerful ones as the lower ones can't keep up. (Like you do between the 290x and 7970)

You think the much lower core count has anything to do with it?  Seriously who the heck taught you people how to do technological analysis? There is back pressure in the pipeline at 1440p and 4K which wasn't present at lower resolutions. Without more raw resources bandwidth won't matter. 2048 cores vs. 2816 and 2880 and you really think this isn't a more pressing issue?

 

The 7970 and 290X have very different pipeline designs. That's not a valid comparison to make, and the intricacies of fixing it go way above what I can write in a reasonably long post. And actually if you look closer NOW instead of against old benchmarks, there's practically no performance loss moving between resolutions for the 980, though for the 970 you start getting cache misses and have to hit that last 0.5GB of memory.

[han_han08]

GPUs are intricate pipelines when it comes to graphics, even though in scientific computing only the cores themselves are used. There is a problem you don't often see discussed for GPUs, even though it's discussed extensively in CPU design: pipeline backpressure. Currently there are so many shaders per TMU and per ROP that even if the shaders are all fed, you can't make progress until the TMU or ROP finishes its task which uses data from 128 shaders, and then all of those ROPS and TMUs produce a final collaborated result before it all gets sent through the pixel shaders to the screen. If the shader count was 96 instead of 128 per ROP and the TMU count was increased in proportion, then you'd get a boost where the cores are staying fed and passing on more results. When that happens increasing bandwidth can be helpful. Until that time, there are some subsets in every scene ever drawn which don't produce back pressure. For those, increasing bandwidth helps, and this can increase fps, but not to the same degree increasing core clocks (which are also hooked to the ROPs and TMUs) will.

 

For the 960 you're well below that 250GB/s threshhold which is a good estimation even if not exact, and there are fewer cores and resources handling the same workloads, so depending on the driver's load balancing per frame, some more bandwidth may be beneficial.

 

The effective clock of 9GHz is a best-case scenario where all information is a highly compressible texture. It's some marketing, but there is truth that you can gain some extra space or more textures by using their color compression tech.

 

If you look at the HBM specs, each stack is a 1024-bit wide interface. There are 4 stacks this time. simply multiply.

 

amazing, can't believe back pressure is still present in cpu/gpus...there's back pressure in forced induction systems, water hammers in pipes, etc...it is damaging to the system if not controlled, is the same for processors?

 

i've only a faint understanding of what you explained  , it seems it is still better to increase core clock (also increasing pixel/clock since you mentioned ROPs?) than bandwidth if it can't be saturated (no benefit?).

 

thanks for the info!

[patrickjp93]

amazing, can't believe back pressure is still present in cpu/gpus...there's back pressure in forced induction systems, water hammers in pipes, etc...it is damaging to the system if not controlled, is the same for processors?

 

i've only a faint understanding of what you explained  , it seems it is still better to increase core clock (also increasing pixel/clock since you mentioned ROPs?) than bandwidth if it can't be saturated (no benefit?).

 

thanks for the info!

For CPUs this iswhy we have buffers and redundant systems to catch these errors and stop the TLBs from accidentally tossing instructions or data from the cache before they're used.

 

Backpressure will always exist for some workloads. No software will ever be so perfect as to avoid it (software of real worth anyway).

[Opcode]

amazing, can't believe back pressure is still present in cpu/gpus...there's back pressure in forced induction systems, water hammers in pipes, etc...it is damaging to the system if not controlled, is the same for processors?

 

i've only a faint understanding of what you explained  , it seems it is still better to increase core clock (also increasing pixel/clock since you mentioned ROPs?) than bandwidth if it can't be saturated (no benefit?).

 

thanks for the info!

Per usual a bunch of made up gibberish that has nothing to do with the post he's trying to counter. GPUs don't have "back pressure" they aren't combustible engines. Nvidia juiced up Maxwell's L2 cache from 1024kb to 2048kb which cuts out a large chunk of the reliance on memory bandwidth which allowed them to cut back on interface width to conserve on cost.

[MEC-777]

390X vs 980Ti (if it's real).

 

Man... I can't wait to see this...

[sgloux3470]

You think the much lower core count has anything to do with it?

 

I wasn't aware that cores could be compared in such a fashion across architectures.  

 

Anyway, I suppose it may not strictly be a memory thing.  I think it is interesting from a real world "which GPU to buy" perspective though.  

[patrickjp93]

I wasn't aware that cores could be compared in such a fashion across architectures.  

 

Anyway, I suppose it may not strictly be a memory thing.  I think it is interesting from a real world "which GPU to buy" perspective though.  

Okay, to be fair I should say SPs instead of cores, because that is the most fundamental unit of data processing. There is back pressure due to the lower processing resources.

[ltguy]

You reminded me, gotta watch out for people dumping their first gen Titans too, see if I can get two of them for a reasonable price vs the 295x2.

Who would do such a thing ???

 

btw

 

wts 2 titans, pst ;P

 

 

*im only kidding, my cards arent for sale. plse dont report me for coc.

[nicehat]

1) Learn some statistics. 3 test rounds is not reliable. 5 is the bare minimum for a 95% confidence interval

 

Drop the pompousness. You really havent contributed any statistical evidence to back up your opinion

 

 

How many times do I have to explain to people games have no bandwidth problem currently.

HBM really won't help gaming much for a very long time

 

 

He provided something (real world observations). However small his sample size, it stands up to your empty opinions any day of the week. Shut up or put up some stats if you want to call him out on his. 

[Ostwind]

395x2 is gonna get rekt by the GTX 2080 Ti

[jato_127]

And In conclusion the AMD R9-395x2 will have a slower clock speed and a higher heat output then the Nvidia GTX 2080 Ti

[beebskadoo]

And In conclusion the AMD R9-395x2 will have a slower clock speed and a higher heat output then the Nvidia GTX 2080 Ti

 4) (256/8)*7*10^9 = 224*10^9 = 224GB/s for GTX 2080Ti

     (512/8)*5*10^9 = 320*10^9 = 320GB/s for R9 395x2

 

 

this means stuff, a smart guy told me.

[Ostwind]

 4) (256/8)*7*10^9 = 224*10^9 = 224GB/s for GTX 2080Ti

     (512/8)*5*10^9 = 320*10^9 = 320GB/s for R9 395x2

 

 

this means stuff, a smart guy told me.

But 2080 is a bigger number so it's way better

[jato_127]

*quote* Beebskadoo
 

4) (256/8)*7*10^9 = 224*10^9 = 224GB/s for GTX 2080Ti

     (512/8)*5*10^9 = 320*10^9 = 320GB/s for R9 395x2

 

 

this means stuff, a smart guy told me.

 

 

 

 

 

Ohhh damn I guess I better learn my statistics xP


 

[Ostwind]

I have made a simple chart here to illustrate how the GTX 2080 Ti and the 395x2 will stack up to each other. I think this will solve the issues that everyone has brought up in the thread.

 

 

[patrickjp93]

 4) (256/8)*7*10^9 = 224*10^9 = 224GB/s for GTX 2080Ti

     (512/8)*5*10^9 = 320*10^9 = 320GB/s for R9 395x2

 

 

this means stuff, a smart guy told me.

It's a bandwidth calculation. 256 bits of width divided by the # of bits per byte multiplied by the effective frequency of the memory circuit which is a quad-pumped DRAM running at 1.5GHz for a 7GHz effective clock, where giga is the prefix for 10^9 in scientific numerics.

 

Bits / bits/byte * 1/seconds-per-cycle = Gigabytes/second = GB/s. 

[Janne]

lol

[ZestyJalapeno]

So guys, graphics cards? They're pretty cool huh?

[CoolaxGaming]

Title updated

[RagnarokDel]

im confused the r9 395x2 has less ram per core than 390x

maybe they mean 8 gb per gpu?

[Norman_the_Owl]

You reminded me, gotta watch out for people dumping their first gen Titans too, see if I can get two of them for a reasonable price vs the 295x2.

You'll never get titans for a reasonable price, first gen is still selling for $550 give or take.

 

Might as well buy a 980 at that point

[IAMSWIFT1]

All over the place.

God damnit, for some reason, I need to press the P key REALLY hard compared to other keys.

Press On The P Real Hard

[That Norwegian Guy]

If it's less than 10.000 NOK (about $1200) I will buy the 395x2.