delete this because less than half the people around here understand or even want to

[R-T-B]

Going forward, I think the creators of Pci Express need to seriously consider abandoning it instead of updating it, for a faster interlink between GPU and CPU, like Nvidia is trying to do with NVLink. It really needs to happen by someone OTHER than a GPU designer to make it unilateral though.

IMB is the single biggest benefactor of NVlink of the four founders of Pci Express. If that kind of thing is to go mainstream on consumer MBs though, it really needs to be backed by both Intel and AMD, the latter of which is unlikely with Nvidia at the helm.

So right now I'd like to see Intel and IMB get together and come up with a successor to Pci Ex. If that happens I'm sure Dell and HP will jump on board.

Isn't the bandwidth gain from PCIe 2.0 x16 to PCIe 3.0 x16 already shown to be largely useless? Why do we need to toss anything out?

 

[xfia]

i was wondering that rtb but i was really leaning with you on that.. i could also agree with frag that with the introduction of hbm will bring more powerful gpu's and at home enthusiast lvl could be 8k and 4k eyefinity so more bandwidth could be needed. there is also improved latency with dx12 im almost positive will translate into more efficient use of the bandwidth.. or since there is full gpu utilization the bandwidth will saturate with high end gpu's but there is nothing to indicate that. could we see pcie 3.1 in coming years with higher coding efficiency bandwidth and backwards compatible?

 

[Aquinus]

Isn't the bandwidth gain from PCIe 2.0 x16 to PCIe 3.0 x16 already shown to be largely useless? Why do we need to toss anything out?

The problem isn't bandwidth, it's latency. Whenever the GPU needs to stream data from the CPU's memory pool, the GPU has to wait to get that data. So even if it were only to need to send one packet of PCI-E data, it has to wait for that one packet. Bandwidth gives you more of something over time, it does not make it respond faster. This is by no means a result of PCI-E being bad, it's just a result of people forgetting how the actual physical length of PCI-E is long and it takes time for the electrical signal to travel. It's the same argument for moving from GDDR5 to HBM; you're moving memory closer to the GPU therefore latency will be less of a problem (hence why bandwidth was hiked with the super wide bus because it's right next to the GPU, almost like another level of cache (think eDRAM on Iris Pro.)

Also consider this, if a GPU uses its own memory, you have a workflow like this:
GPU Cores -> GPU IMC -> VRAM -> GPU IMC -> GPU Cores
If you have to stream data, you end up doing something like this.
GPU Cores -> GPU IMC -> GPU PCI-E interface -> (Possible PLX chip or PCH,) -> CPU PCI-E interface -> CPU IMC -> Main Memory -> CPU IMC -> CPU PCI-E Interface -> (Possible PLX chip or PCH,) -> GPU PCI-E interface -> GPU IMC -> GPU Cores.

I think you can clearly see why there it latency associated with a GPU needing system memory. Simple fact is no interface is going to change this because latency is determined by circuit distance and the number of devices involved in the entire process.

 

[RejZoR]

Ideal would be CPU and GPU merged into a single die (APU) with 32GB of shared HBM memory around both. It can't get more ideal than this.

 

[Aquinus]

Ideal would be CPU and GPU merged into a single die (APU) with 32GB of shared HBM memory around both. It can't get more ideal than this.

It's ideal if and only if AMD reduces power consumption and if AMD doesn't wait so long that Intel outmaneuvers them. Intel already has the eDRAM on Iris Pro CPUs. I suspect making the jump from traditional DRAM to stacked DRAM might not be that hard in comparison and we might find Intel (like nVidia,) pushing hard on that in the near future. While I would love to be optimistic, I think we can say that AMD has a lot of catching up to do when it comes to the X86 market.

 

[BiggieShady]

your failing to grasp what async shaders do and what they mean.. (1) what is suppose to run on the cpu will be and what is suppose to be run on the gpu will be.. (2) microsoft already said one gcn performance optimization was over 30 percent for lighting effects.
(3) some people have called gpu's stupid throughout history well here it is the gpu brains for everyone that understands the async highway
(4) nvidia may try push harder with gameworks but they will fail to the future.
dont be afraid to watch the animated videos! you get to learn and they make it simple so your kids can learn like its the discovery channel!

Oh boy, this post right here. It simply requires some analysis.

1. When you say shader be it synchronous or asynchronous, you have left the realm of the CPU ... CPU did its job by uploading geometry and textures to VRAM beforehand and while rendering it's only sending draw calls with different shader parameters (which geometry with which textures at what position/rotation/scale + all the other shader 'sliders'). Everything in the shader is run by GPU.
2. Since 2010. all GPUs are general purpose compute machines and saying lighting effects are optimized on a hardware level makes no sense at all. It would make sense saying that per-pixel lighting computations are that kind of parallel computational problem where GCN architecture achieves its peak theoretical throughput.
3. Someone on the internet is calling GPUs stupid?
4. If this is about async shaders, ability for gpu threads to instantiate another gpu threads (yes, async shaders in context when GPGPU is used for graphics) exists with Nvidia since Kepler.

 

[RejZoR]

It's ideal if and only if AMD reduces power consumption and if AMD doesn't wait so long that Intel outmaneuvers them. Intel already has the eDRAM on Iris Pro CPUs. I suspect making the jump from traditional DRAM to stacked DRAM might not be that hard in comparison and we might find Intel (like nVidia,) pushing hard on that in the near future. While I would love to be optimistic, I think we can say that AMD has a lot of catching up to do when it comes to the X86 market.

eDRAM is on a die itself and always comes in ratehr tiny quantities, HBM would still be outside of it, but still closer than RAM or VRAM on graphic card with huge buss width. Meaning you solve bandwidth and latency while having huge capacity.

 

[Aquinus]

eDRAM is on a die itself and always comes in ratehr tiny quantities, HBM would still be outside of it, but still closer than RAM or VRAM on graphic card with huge buss width. Meaning you solve bandwidth and latency while having huge capacity.

Intel's eDRAM is 2 years old, it's not stacked, and was their first shot a putting DRAM next to the CPU. It's small in quantity because it's not exactly new technology. If Intel were to make a brand new eDRAM package now, I suspect that it would be closer to 1GB in size. The simple fact is that GPUs aren't the only device that could benefit from a pool of memory as large as system memory but quicker to respond. Intel's eDRAM much like HBM, is *very* wide. If you haven't read an Intel spec sheet on a CPU with GT3e, then (if you can understand half of it,) I recommend reading it because Intel's eDRAM is very similar to HBM in a lot of ways, only that it's not stacked DRAM. However like HBM, it's clocked lower and the bus is very wide.

I'm more interested in being able to buy a CPU and not need to consider buying DRAM right away. Imagine what that would do for mobile devices. HBM deserves a place in the memory hierarchy between cache and system memory, IMHO.

 

[Mussels]

Intel's eDRAM is 2 years old, it's not stacked, and was their first shot a putting DRAM next to the CPU. It's small in quantity because it's not exactly new technology. If Intel were to make a brand new eDRAM package now, I suspect that it would be closer to 1GB in size. The simple fact is that GPUs aren't the only device that could benefit from a pool of memory as large as system memory but quicker to respond. Intel's eDRAM much like HBM, is *very* wide. If you haven't read an Intel spec sheet on a CPU with GT3e, then (if you can understand half of it,) I recommend reading it because Intel's eDRAM is very similar to HBM in a lot of ways, only that it's not stacked DRAM. However like HBM, it's clocked lower and the bus is very wide.

I'm more interested in being able to buy a CPU and not need to consider buying DRAM right away. Imagine what that would do for mobile devices. HBM deserves a place in the memory hierarchy between cache and system memory, IMHO.

isnt EDRAM included in the Xbox one, with AMD hardware?

 

[Aquinus]

isnt EDRAM included in the Xbox one, with AMD hardware?

Yeah, but I doubt it's Intel's eDRAM (Crystalwell.) I'm sure there is variation in implementation, it's not like DRAM technology is a closely guarded secret.

 

[Frick]

Is this the most hyped thing in computers in quite a while? It's not surpising, but it's stupid.

 

[bpgt64]

your failing to grasp what async shaders do and what they mean.. what is suppose to run on the cpu will be and what is suppose to be run on the gpu will be.. microsoft already said one gcn performance optimization was over 30 percent for lighting effects.
some people have called gpu's stupid throughout history well here it is the gpu brains for everyone that understands the async highway
nvidia may try push harder with gameworks but they will fail to the future.
dont be afraid to watch the animated videos! you get to learn and they make it simple so your kids can learn like its the discovery channel!

You're failing to grasp no one gives a shit untill W1zzard( or any other reasonable 3rd party for that matter) posts a review with an hard proof that it is used and translated into performence in gaming.

Or should we start talking about the merits of eSRAM on Xbox One?

 

[RejZoR]

eDRAM was already used in XBOX 360...

 

[mirakul]

So long story short.
Buy an AMD card now, not much rock with DX11 now but rock with DX12 later.
Buy an nVi card now, rock with DX11 now but not much rock with DX12 later.
Which kind of rock do you want?

 

[rtwjunkie]

Really, bc all the current Nvidia cards are getting full DX12, and even recent ones will get partial.

It's going to be awhile before we get more than one or two DX12 games. The big advantage is going to be the positive effect on DX11 and below games, for which all DX12 cards will benefit.

 

[mirakul]

Really, bc all the current Nvidia cards are getting full DX12, and even recent ones will get partial.

It's going to be awhile before we get more than one or two DX12 games. The big advantage is going to be the positive effect on DX11 and below games, for which all DX12 cards will benefit.

You missed the discussion here. We are talking about how DX12 pumps up AMD's performance compare to DX11. That benefit doesn't exist in nVidia's cards, cause they are well optimized for Dx11 now and have nothing to unlock with DX12.

 

[MxPhenom 216]

because some people just go on ignoring what the people that actually make the stuff say.. highly cynical

Its PR and marketing for people to jump aboard the hype train. Both NVIDIA and AMD do it. Of which should be taken with a grain of salt till great people like Wizzard can actually test it.

You missed the discussion here. We are talking about how DX12 pumps up AMD's performance compare to DX11. That benefit doesn't exist in nVidia's cards, cause they are well optimized for Dx11 now and have nothing to unlock with DX12.

Not exactly right. Amd has been known to have high overhead issues with their drivers in dx11, and I think that's why they developed Mantle for the most part. NVIDIA was able to increase draw calls and reduce overhead since their driver release after GDC when they announced their "magic" driver for dx11. However its hard to say that just AMD will see a boost. NVIDIA will too.

 

[mirakul]

Not exactly right. Amd has been known to have high overhead issues with their drivers in dx11, and I think that's why they developed Mantle for the most part. NVIDIA was able to increase draw calls and reduce overhead since their driver release after GDC when they announced their "magic" driver for dx11. However its hard to say that just AMD will see a boost. NVIDIA will too.

Could you please explain where the boost for nVidia come from, given that their cards was very optimized for this overhead issue?

 

[MxPhenom 216]

Could you please explain where the boost for nVidia come from, given that their cards was very optimized for this overhead issue?

Mainly Tiled Resources. That was pretty much the highlight of the driver release.

 

[BiggieShady]

You missed the discussion here. We are talking about how DX12 pumps up AMD's performance compare to DX11. That benefit doesn't exist in nVidia's cards, cause they are well optimized for Dx11 now and have nothing to unlock with DX12.

Not exactly right. Amd has been known to have high overhead issues with their drivers in dx11, and I think that's why they developed Mantle for the most part. NVIDIA was able to increase draw calls and reduce overhead since their driver release after GDC when they announced their "magic" driver for dx11. However its hard to say that just AMD will see a boost. NVIDIA will too.

Actually any optimizations in DX11 are miniscule in comparison of performance increase DX12 will bring to both teams because draw call overhead is a CPU side optimization. Increase will be so radical that CPU bottleneck in gaming will be thing of a past and that completely shifts balance to GPUs where Nvidia has upper hand in geometry and pixel pushing power while AMD has more shading performance and memory bandwith. That's where the dis-balance is, not in DX12 benefits, they will be win for all.

 

[mirakul]

Actually any optimizations in DX11 are miniscule in comparison of performance increase DX12 will bring to both teams because draw call overhead is a CPU side optimization. Increase will be so radical that CPU bottleneck in gaming will be thing of a past and that completely shifts balance to GPUs where Nvidia has upper hand in geometry and pixel pushing power while AMD has more shading performance and memory bandwith. That's where the dis-balance is, not in DX12 benefits, they will be win for all.

The point is the current line up of AMD suffered more from DX11 and CPU bottleneck. There will be win for all, but the leap on AMD side is more significant.

 

[MxPhenom 216]

The point is the current line up of AMD suffered more from DX11 and CPU bottleneck. There will be win for all, but the leap on AMD side is more significant.

What ever is seen with mantle, will be the same for dx12, if the comparison with 3dmark overhead API benchmarks means anything.

 

[xfia]

Oh boy, this post right here. It simply requires some analysis.

1. When you say shader be it synchronous or asynchronous, you have left the realm of the CPU ... CPU did its job by uploading geometry and textures to VRAM beforehand and while rendering it's only sending draw calls with different shader parameters (which geometry with which textures at what position/rotation/scale + all the other shader 'sliders'). Everything in the shader is run by GPU.
2. Since 2010. all GPUs are general purpose compute machines and saying lighting effects are optimized on a hardware level makes no sense at all. It would make sense saying that per-pixel lighting computations are that kind of parallel computational problem where GCN architecture achieves its peak theoretical throughput.
3. Someone on the internet is calling GPUs stupid?
4. If this is about async shaders, ability for gpu threads to instantiate another gpu threads (yes, async shaders in context when GPGPU is used for graphics) exists with Nvidia since Kepler.

you no doubt have a better understanding of it. to me its just low level optimization at this point.
ok.. nv has had better api overhead in dx11 leaving less untapped performance but with dx12 not only will both be better but amd's gpu's are just capable of more when fully utilized.

 

[Steevo]

Not exactly right. Amd has been known to have high overhead issues with their drivers in dx11, and I think that's why they developed Mantle for the most part. NVIDIA was able to increase draw calls and reduce overhead since their driver release after GDC when they announced their "magic" driver for dx11. However its hard to say that just AMD will see a boost. NVIDIA will too.

High overhead caused by things like, Game works forcing their driver to do work by handing proprietary Cuda bullshit back and forth to the CPU.

Nvidia started hairworks as direct compute, and ran it like that until they realized they could "encourage" users of their prior series to upgrade and make it look like AMD drivers look shitty by instead going overboard with Cuda and PhysX in closed box games.

 

[rtwjunkie]

You missed the discussion here. We are talking about how DX12 pumps up AMD's performance compare to DX11. That benefit doesn't exist in nVidia's cards, cause they are well optimized for Dx11 now and have nothing to unlock with DX12.

Actually no, I missed nothing, as stated by the responses above by @BiggieShady and @MxPhenom 216 before I got back here. My point is the "big win" given by DX12 will apply to all, not just AMD. If anything, it puts both camps on a very level playing field, including pre-DX12 games.