If GDDRX is so great, why is Nvidia using HBM 2 on GP100?
Vega is already being used in massive server farms by Liquid Sky. Desktop sales pale in comparison to server sales.
Nvidia is having massive success with GDDR5X - why on earth haven't AMD jumped on this bandwagon? Do NVIDIA have exclusive access to it?
f GDDRX is so great, why is Nvidia using HBM 2 on GP100?
It's simple. GP104 and GP102 doesn't have IMC with HBM support, it was not designed for use on interposer etc.Why would NV use GDDR5x on 1080Ti (GP102) if HBM2 was cheaper?
It's simple. GP104 and GP102 doesn't have IMC with HBM support, it was not designed for use on interposer etc.
If GDDRX is so great, why is Nvidia using HBM 2 on GP100?
Vega is already being used in massive server farms by Liquid Sky. Desktop sales pale in comparison to server sales.
Because HBM2 is expensive to implement, is expensive itself, makes it prohibitively expensive for large memory configurations, and is just generally a poor solution for consumer products when a GDDR-follow on that can offer similar performance is available.
It has nothing to do with "release rights."
Uh, okay. None of that is true. We don't yet know what the memory controller of Vega can actually do, but AMD used a "simple" HBM controller in the Fury cards - which worked perfectly well - and Nvidia has something comparable in P100. A "simple" memory controller can perfectly well utilize all the bandwidth of the memory connected to it. That's up to a) the bandwidth and speed of the controller (which again in a determinant of what memory ends up being connected to it, not the other way around), and b) the tasks given to it. Or are you saying Nvidia's GDDR5/5X controllers can't utilize the full bandwidth of those either?Sorry, but Nvidia doesn't have a cohesive unified memory controller like Vega does. So, it can not make use of HBM/HBM2 like Vega can. HBM is just a selling point/marketing for Nvidia anyways, their technology can't make full use of it's bandwidth. That is why NVidia cheapened up using inferior GDDR5x.
that explicitly means that things are very, very bad at intel for like 2-3 years nowWhenever there are big delays, you know things dont look good.
That is just by a design choice, they can go (and most certainly will, with professional products first) to 4 stacks any time. With the new hbcc, more bandwidth is simply not needed, and they have this bandwidth on an area much, much smaller than they'd have gotten it with ddr ramTwo stack HBM2 with only a minor bandwidth advantage compared to cut down GP102 (8GB 512GB/s vs 11GB 484GB/s) isn't really impressive - Fury X had 50% more bandwidth than GTX 980 Ti. Meanwhile rumor says the competition is embracing 16GB GDDR6 at >14Gbps for their next generation.
I think NVIDIA hate working with interposers, they just don't have any other choice in the pro marketBecause it's better but also more expensive. Why would NV use GDDR5x on 1080Ti (GP102) if HBM2 was cheaper?
I'm not an NV fan but they clearly have an edge in memory bandwidth and power usage. So they can get away with GDDR5x instead of hbm2. If a technology is better and cheaper, we would have seen way, way faster adoption.
worked 'well' -> it couldn't use more than 2/3 of its theoretical bandwidth. That is changing with hbcc and hbm2 for vega. We must wait to be able to test gp100 against vega, so we can compare their memory solutions, but fury's HMB is not an indicator for this.Uh, okay. None of that is true. We don't yet know what the memory controller of Vega can actually do, but AMD used a "simple" HBM controller in the Fury cards - which worked perfectly well - and Nvidia has something comparable in P100. A "simple" memory controller can perfectly well utilize all the bandwidth of the memory connected to it. That's up to a) the bandwidth and speed of the controller (which again in a determinant of what memory ends up being connected to it, not the other way around), and b) the tasks given to it. Or are you saying Nvidia's GDDR5/5X controllers can't utilize the full bandwidth of those either?
I'm a big supporter of AMDs desicion to go HBM2 with Vega - it has a ton of advantages - but it's very clear that Nvidia's choice of GDDR5X for high-end consumer cards stems from a combination of cost, availability, GDDR5X being "close enough" in performance, and Nvidia's GPU architecture efficiency advantage. Why? Nvidia sells far more GPUs than AMD, so they'd need a much healthier supply of HBM. This doesn't exist, and even if it did, costs for higher volumes would be huge. They can fit the 20+W of a 8GB+ GDDR5X setup within reasonable board TDPs due to their (for now) superior GPU efficiency. And, as the 1070, 1080 and 1080Ti show, GDDR5X is no slouch. Would HBM have been better? Probably, but it would significantly cut into Nvidia's margins (or force prices higher) while giving them a small advantage (over themselves?) at best.
While the Vega memory controller sounds innovative and might be ground-breaking, it certainly isn't the sole reason why AMD went HBM2. After all, they made HBM GPUs way before a controller like that existed.
You do not need to have Unified Memory controller on hardware to get the feature compatibility. Nvidia did with GP100 the same thing, but through CUDA. CUDA Unified Memory has exactly the same level of compatibility with this feature(49 bit register), like Vega has on hardware level. But here are starting the differences.Sorry, but Nvidia doesn't have a cohesive unified memory controller like Vega does. So, it can not make use of HBM/HBM2 like Vega can. HBM is just a selling point/marketing for Nvidia anyways, their technology can't make full use of it's bandwidth. That is why NVidia cheapened up using inferior GDDR5x.
Uh, okay? Numbers/benchmarks to demonstrate that, please? That account for delta color compression and other factors, and that measure VRAM-to-GPU bandwidth only? If you're drawing conclusions from comparisons with Nvidia cards with slower memory, you have to account for their clearly superior compression algorithms.worked 'well' -> it couldn't use more than 2/3 of its theoretical bandwidth. That is changing with hbcc and hbm2 for vega. We must wait to be able to test gp100 against vega, so we can compare their memory solutions, but fury's HMB is not an indicator for this.
Uh, okay? Numbers/benchmarks to demonstrate that, please? That account for delta color compression and other factors, and that measure VRAM-to-GPU bandwidth only? If you're drawing conclusions from comparisons with Nvidia cards with slower memory, you have to account for their clearly superior compression algorithms.
You do not need to have Unified Memory controller on hardware to get the feature compatibility. Nvidia did with GP100 the same thing, but through CUDA. CUDA Unified Memory has exactly the same level of compatibility with this feature(49 bit register), like Vega has on hardware level. But here are starting the differences.
Vega has on hardware level compatibility with Unified Memory, in OpenCL 2.0. CUDA has software level compatibility with Unified Memory in OpenCL 2.0.
Where is the difference? Software layer of abstraction. You will not have the benefit of 49 bit register, if you do not use software, specifically designed for the hardware.
AMD hardware will always adapt itself, to the situation.
It will be interesting from my perspective about Vega. Vega appears to not bring anything to the table in terms of compute capabilities, and I am a part of professional forum, on which professionals make their decisions about hardware purchases, based on gaming benchmarks, and completely ignoring compute benchmarks. I could post time and time again, showing RX 480 being in professional applications just 10% slower than GTX 1070, and nobody there cared about it. That is how strong appeal Nvidia brand has. I will have a very good laugh from them, watching what will happen with Vega.
http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4
780 Ti, Titan X, and 980 Ti all have 336 GB/s. 980 has 224 GB/s.
290X has 320 GB/s. Fury X has 512 GB/s.
Fury X has 1.6x the theoretical bandwidth of the 290X, but does not come close to that in this test. So you don't even have to compare it to Nvidia to see they are not doing so well with HBM. In this one test only, of course. But it's the only one I have seen, so anyone with more link them please.
Nope, they can. HBM will act just like "normal memory" not like just a cache for Data. It is important to understand the difference here, because it changes approach, but does not make one better or worse form the other.I understand about the 49 bit registers..
But what I am talking about is the actual non-volatile memory that interacts with HBM as a cache. Nvidia doesn't have this, therefore they can't use HBM to it's fullest potential. That is why GDDR5x is good enough.
So to avoid compressible textures, you need to look at the "Random texture" numbers.
http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4
780 Ti, Titan X, and 980 Ti all have 336 GB/s. 980 has 224 GB/s.
290X has 320 GB/s. Fury X has 512 GB/s.
Fury X has 1.6x the theoretical bandwidth of the 290X, but does not come close to that in this test. So you don't even have to compare it to Nvidia to see they are not doing so well with HBM. In this one test only, of course. But it's the only one I have seen, so anyone with more link them please.
The 'good news' from this is that AMD have room to work with. Even though Fiji and Vega have the same theoretical bandwidth, if AMD can catch up to Maxwell v2 efficiency then there will be an appreciable bandwidth increase.
I did absolutely not compare fury to any nvidia gpu. I was reading a lot of professional forums back then with tests (not game benches) that check actually usable memory bandwidth. I apologize for not wanting to search for a forum discussion that has happened more than 1,5 years ago.Uh, okay? Numbers/benchmarks to demonstrate that, please? That account for delta color compression and other factors, and that measure VRAM-to-GPU bandwidth only? If you're drawing conclusions from comparisons with Nvidia cards with slower memory, you have to account for their clearly superior compression algorithms.
Thanks for the post.- Radeon RX580/RX570 launching April 18th, right after Ryzen 5
- Same Polaris GPUs, frequency bump (unknown if other differences)
- Polaris 12 positioned against GTX 1050
- Vega expected in May
http://news.mydrivers.com/1/523/523439.htm