This is an Nvidia technology running on dedicated hardware. Consoles won't have DLSS.
Yes, but DLSS 2.0 is Nvidia's exclusive tech. DLSS wont be on AMD consoles
They won't be getting DLSS, but there's nothing preventing Ms and Sony to come up with their own solutions, in fact Ms is already demoing one implementation that can also increase resolution from 540-1080p, and from 1080p to 4k.Ok, they still won't get dlss2.0 since consoles are using amd tech.
It remains to be seen if AMD have an equivalent tech, but they have not announced anything so far.
Actually the tensor cores were not being utilized until now. I also thought that they were used before but that is not the case.Not when you know you're always using said dedicated cores (which is the case here). Dedicated hardware always in use is always superior.
That's precisely why they added hardware instructions, so they are inherently the same. The number of processors makes a difference, and RTX gpus do have more processors to handle that, but the counter-argument is that on consoles the space dedicated for that is also used for other tasks.You're forgetting that the tensor cores are much, much more efficient at doing these kinds of ML workloads than shader cores. The amount of taken up die space is not relevant in this case. And they don't take up as much die space as you think. I believe the number was 7%, could be wrong though.
That's exactly what the hardware they added allows. They added support for dealing with int matrix math, and support to perform multiple of these operations at once since a int is a lot smaller than a full fp32 number.Have you read the full quote yourself? That's referring to the shader cores being able to run 8-bit and 4-bit operations at that level of precision, which is why the TOP figures are what they are.
Again, that's not what the quote is saying. No hardware was added; the shader cores were tweaked to run low-precision workloads more efficiently, not unlike "Rapid Packed Math" (AMD's buzzword for using a single FP32 operation to run two FP16 operations).
Damn, why sony just don't confirm what this hardware has or doesn't XDWe still do not know if PS5 supports increased rate int8.
XSX does.
XSX is like 49 int8 TOPs - RTX 2080 Ti is 220 at ca. 1500mhz (most 2080 Tis run at 1800-1950 mhz in real life though).
Consoles would have to trade ALU for machine learning tasks, but ALU is usually cheap as pretty much anything is always more bandwidth bound than alu bound, especially if you are already targeting a lower resolution, likely just the bandwidth gains from that will more than compensate the ALU you lose.The consoles may have mixed precision support in their shader ALUs, but DLSS here is running on tensor cores. So on the consoles, without further hardware, they'll be trading native rendering performance against reconstruction performance/time. Whether you wind up with a net benefit to performance as in DLSS would be an open question. It certainly wouldn't be as beneficial.
I wonder also how useful lower precision support in shader alus will be for image reconstruction, below fp16 anyway, before you start introducing new problems. Not sure what the answer is there.
I think it's maybe more possible we will see different reconstruction techniques on the consoles that are lighter at processing time than a DLSS-alike would be running on shader ALUs, but better than what we had before on consoles. And maybe even using some neural network processing, but maybe trading generality (as in DLSS) for lower overhead if running on shaders.
But this is machine learning, Nvidia is hardly the only one researching how to upscale images using ML, there are even open source ones already;This is an Nvidia technology running on dedicated hardware. Consoles won't have DLSS.
Is there even any point in buying an AMD GPU until they implement something that performs similiarly well? You'll have the 3060 performing better than the highest end AMD card. With better RT too probably
"Dedicated hw" goes into semantics and grammar.
People are trying to differentiate between discrete processing units (like Tensor cores) and extensions to the SIMD execution units to process 4/8/16 of 32/16/8 bit precision types in parallel.
To be more specific - PS2, XBox(original), PS3, and 360 all had hardware accelerated support for the latter - but it didn't get additional labels other than SIMD acceleration hardware.
Knowing Nvidia they will probably jack the prices of everything up. :/ Not looking forward to the 3080/3080 ti prices.Is there even any point in buying an AMD GPU until they implement something that performs similiarly well? You'll have the 3060 performing better than the highest end AMD card. With better RT too probably
1660 doesn't support dlssI...I've seen the future and it is 540p!
That's crazy! Here I am sitting here glad that I bought a 1660 Super cause I didn't need ray tracing or DLSS. I still don't care for RT, but DLSS would be amazing! Would allow me to use my GPU for far longer than normal. Now I'll have to wait for the 2060 TI or 2060 Super to drop in price.
if a game has amd marketing , there are no nvidia effects to choose from afaik.Someone remind me, how were Nvidia exclusive graphics features (HBAO+ for example) handled in past titles that had AMD marketing on PC? Were they purposefully left out?
Not really, unless DLSS wont be adopted heavily and going by DLSS 2.0 not needed to be trained anymore, and TAA implementation 'replacement', and UE 4 having a beta branch with DLSS implementation, it will be adopted by a lot of games.Is there even any point in buying an AMD GPU until they implement something that performs similiarly well? You'll have the 3060 performing better than the highest end AMD card. With better RT too probably
My point was there's no dedicated hw working in conjunction with compute units on the SXS. It extends SIMD execution units to support multiple data-widths - ie. it's a building block 'of' the CUs if you prefer.the intention of the dedicated hardware that was added by Microsoft and how it works in conjunction with the compute units.
I think the question mark is more on DLSS itself, NVidia doesn't have a history of sharing this type of tech. MS has clearly demonstrated they are interested in investing in the space though so expectation would be they'll provide their own approach to it.This started because someone said "this won't be on consoles." in the context of machine learning upscaling being used on next gen console games.
Assuming linear scaling, yeah this is what we have basically calculated out for ourselves as well.fyi
So if 2080 ti has ~220tops, 2060 rtx has ~100 tops and xsx 49 tops, according to this graph the rtx 2060 will do roughly ~2.83ms @ 4k, hipothetically the xsx with same algorithm should cost about 5.5ms @4k.
Only problem is AMD sponsored games won't use Nvidia tech. It's always a bit of a disappointment when AMD sponsors a game like Resident Evil 2 remake, which desperately needs ray traced reflections & instead AMD offers a logo & nothing more. The one cool thing they ever did was TresFX.Not really, unless DLSS wont be adopted heavily and going by DLSS 2.0 not needed to be trained anymore, and TAA implementation 'replacement', and UE 4 having a beta branch with DLSS implementation, it will be adopted by a lot of games.
AMD's GPU division will be in a really tough spot from now.
Special hardware means special support for these functions on the CUs of the Series X. NVIDIA has DEDICATED units for them, so they don't need to run them on the CUs.OK, now you are being dishonest here. Why are you ignoring where it says "So we added special hardware support for this specific scenario."? If I ask you if the XSX has special hardware to accelerate machine learning code, is your answer no it doesn't?
Assuming linear scaling, yeah this is what we have basically calculated out for ourselves as well.
5.5 ms btw is still A LOT faster than running the game at 4x the resolution.
but at least tress fx runs on nvidia gpus.Only problem is AMD sponsored games won't use Nvidia tech. It's always a bit of a disappointment when AMD sponsors a game like Resident Evil 2 remake, which desperately needs ray traced reflections & instead AMD offers a logo & nothing more. The one cool thing they ever did was TresFX.
My point was there's no dedicated hw working in conjunction with compute units on the SXS. It extends SIMD execution units to support multiple data-widths - ie. it's a building block 'of' the CUs if you prefer.
Special hardware means special support for these functions on the CUs of the Series X. NVIDIA has DEDICATED units for them, so they don't need to run them on the CUs.
I think the question mark is more on DLSS itself, NVidia doesn't have a history of sharing this type of tech. MS has clearly demonstrated they are interested in investing in the space though so expectation would be they'll provide their own approach to it.
The CU will be occupied running the ML code, while on NVIDIA, the CUs will be free to run normal code, while Tensors run the ML code.
The CU will be occupied running the ML code, while on NVIDIA, the CUs will be free to run normal code, while Tensors run the ML code.
No acceleration is done here, they still run on the CUs, which means they consume regular CU TFLOPs available for graphics. While NVIDIA's solution don't use the CUs and don't consume graphics TFLOPs.Again, what is the problem here? The CUs were modified physically to accelerate machine learning code. This is not a software solution. Actual hardware was added to the CU to enable them to accelerate machine learning code.
No acceleration is done here, they still run on the CUs, which means they consume regular CU TFLOPs available for graphics. While NVIDIA's solution don't use the CUs and don't consume graphics TFLOPs.
Lol not the only one. I'm almost positive that XSX will do 4k natively 99.7% of the timeFor a minute there I thought i'd mistakenly entered the XSX deep dive thread.
So much speculation.
You are obviously less informed in this matter, I will refer you to this quote here:Thank you. I will stop talking to you now.
"We knew that many inference algorithms need only 8-bit and 4-bit integer positions for weights and the math operations involving those weights comprise the bulk of the performance overhead for those algorithms," says Andrew Goossen. "So we added special hardware support for this specific scenario. The result is that Series X offers 49 TOPS for 8-bit integer operations and 97 TOPS for 4-bit integer operations. Note that the weights are integers, so those are TOPS and not TFLOPs. The net result is that Series X offers unparalleled intelligence for machine learning."
The difference between having separate hardware (NVIDIA's tensor cores) and having hardware support inside the shader cores (XSX approach) is that in the former scenario, you don't have to compromise on GPU compute power for native rendering in order to support a DLSS-like feature, while in XSX case, you need to share resources between native rendering and the upscaling AI algorithm. For example: the RTX2060 has 100 TOPS of tensor core performance. If we assume that we need 25 TOPS to do upscaling from 1080p to 4K, then it can do that on the tensor cores, and use its full 6 TFLOPS of ALU power to render the native 1080p image. The XSX, on the other hand, needs to apply 25 TOPS of shader compute power out of its 49 TOPS total, or roughly half of its GPU, to the DLSS-like algorithm, leaving only half of its GPU for native 1080p rendering. If it takes half the GPU to perform the DLSS algorithm, you lose quite a bit of the potential benefit.
Having quoted me, you should also read my further posts in which my point of view evolved when faced with new data/information.You are obviously less informed in this matter, I will refer you to this quote here:
This is really getting into semantics - but extending SIMD execution units for different data-widths isn't adding so much as modifying things. There's a reason SIMD implementation have done this since late 90s/early 00s - it has a relatively negligible impact on execution unit size/complexity.
That's exactly what the hardware they added allows. They added support for dealing with int matrix math, and support to perform multiple of these operations at once since a int is a lot smaller than a full fp32 number.
And I'm not basing solely on this quote, I'm talking about the gdc presentation where Ms first said that would be the approach they would take and talked about the performance considerations. With the modifications they made the hardware IS able to use hardware instructions to deal with machine learning workloads.
That's the downside of going with AMD.
Nvidia Raytracing + DLSS is going to be the real game changer here.
If you see dlss 2.0, native 4k feels just stupid. It's insane how much Ressources you save and can put into stuff like raytracing. It's probably also currently the only way to have raytracing, high IQ and good framerates.Lol not the only one. I'm almost positive that XSX will do 4k natively 99.7% of the time
Deliver Us To The Moon, Wolfenstein Youngblood, Control, Mechwarrior, Bright Memory and soon AMID Evil.
Deliver Us To The Moon, Wolfenstein Youngblood, Control, Mechwarrior, Bright Memory and soon AMID Evil.