You might be on to something here, a lot of buffering is being moved around to keep the VRam usage low, if its constantly flushing and uploading the pool it will cause a slowdown, if its moving objects between the GPU buffer and system RAM it will increase the latency massively.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
AMD Working On An Entire Range of HBM GPUs To Follow Fiji And Fury Lineup – Has Priority To HBM2 Cap
- Thread starter Boomstick777
- Start date
You might be on to something here, a lot of buffering is being moved around to keep the VRam usage low, if its constantly flushing and uploading the pool it will cause a slowdown, if its moving objects between the GPU buffer and system RAM it will increase the latency massively.
With the whole is it the HBM or the FIJI architecture or just AMD's memory management, causing not so good 1080p discussion. Unfortunately we are not going to be able to prove things one way or the other until there is another HBM enabled card that is not FIJI to test against. I mean if the first NVidia HBM card also has issues with lower resolution performance then that would point to the HBM being a culprit, if it doesn't then that points to it being down to something AMD has done with FIJI.
Big Pascal already taped out last month.
I expect Nvidia wont be in any rush to release pascal card given that they have no real competition right now, and if a price war did start they would win because their cards are simply much cheaper to produce than HBM/AIO solutions with poor yields.
If they wanted to they could aim at a March release I would think, that is what,some rumours suggested. I think they might just wait for summer 2016 and enjoy the current high sales.
The Pascal architecture is in some sense very similar to Maxwell v2. Maxwell is a very forward looking architecture wit plenty of life in it. Nvidia have talked about Volta (the successor to pascal) even being an evolution of Maxwell. that is why Nvida invested to heavily in it, why we see things like such a huge increase in performance per watt on the same node.
So the new architecture is not such a huge step forward, it is mostly about working with HBM and a the die shrink.
Nvidia have plenty of time to work out supply, manufacturing, testing and driver issues with HBM.
If this was true the performance problem would be @2160p where the memory was moving stuff around all the time, @1080p this is not needed.
The problem is the low clockspeed of the HBM.
It is the same problem you would get if your gaming PC had a massive multi core Xeon chip with a low clockspeed, massively powerful but still bottlenecking the GPUs.
A good way to test this theory Kaap would be to test how much of a performance improvement overclocking HBM at 1080P offers.
I think that's part of the reason I am making the argument in the first place.
In theory with the bandwidth available there should be no improvement from overclocking the memory yet people who have done it have got an improvement.
It suggests that for 1080p the clockspeed is holding HBM back.
I will test properly once mine are waterblocked up.
Not really been reading this thread but I did read Kaap saying that Fiji isn't ideal for 1080P and if you watch my vids, that was quite apparent. I would even say 1440P isn't suited either and it is only UHD that see's any benefit from HBM that I can tell.
Memory IC's are not CPU's, you can have a thousand cores in your Xeon the performance in DX11 is exactly the same as it is if you only have 6 because only one of them is used for DrawCalls.
Run the same thing in DX12 and your DrawCalls are 6 times higher
A GPU Memory architecture does not use one 32Bit lane with the others idling, in the case of Hawaii it will use 16 of them to make up the 512Bit bus.
In the case of HBM its four 1024Bit lanes.
GDDR5: 16x 32Bit @ 1250Mhz = 320GB/s of memory performance
HBM: 4x 1024Bit @ 500Mhz = 512GB/s of memory performance
The end result is how much performance it has, that is measured in GB/s.
Overclocking HBM is not proof that the MHz are a bottleneck, its proof that overclocking something increases performance.
its nothing to do with the clock speed of hbm... the bandwidth is determined by line width AND memory clock together. And your analogy is wrong in a massive way... one is a processing information, the other is a memory system. the way the clocking works is completely different between the two.
So why does overclocking the memory on a Fury X @1080p bring positive gains even though the card already has massive bandwidth compared to say a TitanX ?
Answer because the memory clockspeed is holding HBM back.
Perhaps you should get hold of a Fury card and try it yourself !!!
Here is the Firestrike 1080p standard bench you will notice the people who overclocked the memory on their Fury Xs have got the higher graphics scores.
Last edited:
Soldato
- Joined
- 20 Apr 2004
- Posts
- 4,408
- Location
- Rugby
Think of the "poo"storm if Nvidia "Has Priority To HBM2 is all I've got to say
Soldato
- Joined
- 30 Nov 2011
- Posts
- 11,522
Agree. The bandwidth on the fury x is already huge, overclocking should bring zero gains as overclocking ram on TX and 980ti does nothing, if it was an increase in bandwidth providing the boost it would work on both.
You should also notice that all three people are using different processors... the one at the top having the newest/highest clocked.
And all you have shown as what kaap stated is that overclocking can improve performance. Not that the memory is holding back the system, if anything it shows the cpu choice making the largest difference. but you need to show it on the same system, its not comparable when using different systems.
with the mantle/dx comparison on gregs system its showing the Driver/api being the problem. not the memory.
Last edited:
Soldato
- Joined
- 30 Nov 2011
- Posts
- 11,522
Gfx scores aren't affected by cpu, nice try though.
Overall score and physics score yes, gfx score no, which is what kaap is talking about.
You've got a lower clocked furyx beating one clocked higher, because it also has another +50 on the memory
Overall score and physics score yes, gfx score no, which is what kaap is talking about.
You've got a lower clocked furyx beating one clocked higher, because it also has another +50 on the memory
Last edited:
Because your overclocking it, by overclocking you are increasing the memory bandwidth, 550Mhz is a 10% overclock, your increasing the memory performance from 512GB/s to 563GB/s.
In the same way that if you overclock GDDR5 you are also increasing the memory performance, from 1250Mhz (320GB/s) to 1375Mhz (352GB/s)
Its overclocking, thats what it does, your too hung up on the 500Mhz thing, you think because its a low number it must be slow, its the same mentality that makes people think Nvidia's 1750Mhz memory must be faster than AMD's 1250Mhz, its very simplistic logic which ignores the 512Bit bus on one vs the 256Bit bus on the other.
512Bit @ 1250Mhz = 320GB/s
256Bit @ 1750Mhz = 225GB/s
Its nothing to do with the HBM architecture and the IC's Mhz. it is faster. much faster than GDDR5, doesn't mean its fast enough for what it needs to do on Fiji.
Last edited:
Looks to be scaling with gpu core clock to me.
of course it is affected by CPU... because DX11 is affected by the driver which is affected by the CPU... if it was not affected by the CPU then every card at the same frequency should have the same scores within a small margin of error.
The above is also true when you compare driver versions, newer ones improving performance by lowering driver overhead.
yet there are lower clocked cards with more powerful cups getting higher scores than higher clocked cards with less powerful cpus. you only have to look at the chart to see that.
Last edited:
CPU Physics is very different to DrawCalls, CPU Physics can be run across multiple cores, so the more cores you have the higher performance you will get.
DrawCalls in DX11 are one thread and one thread only, so it makes no difference how many cores you have, only by increasing the CPU clock rate can you increase DrawCall performance.
DrawCalls in DX11 are one thread and one thread only, so it makes no difference how many cores you have, only by increasing the CPU clock rate can you increase DrawCall performance.
Firestrike graphics scores are not effected much if all the people are using top notch i7s which is the case here.
Also if overclocking the memory did not work on the Fury X why have the guys above done it. They are still using the same CPU before and after lol.