AMD’s ‘Big Navi’ graphics cards have hit North America for testing

erek

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
10,868
Hook it up !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

"And with those 80 CUs you’re looking at 5,120 RDNA 2.0 cores. Quite how AMD is managing the hardware-accelerated ray tracing we still don’t know. It will surely be designed to support Microsoft’s DirectX Raytracing API, given that is what will likely power the Xbox Series X’s approach to the latest lighting magic, but how that is manifested in the GPU itself is still up in the air."

upload_2020-1-21_19-48-55.png


https://translate.google.com/transl...phell.com/thread-2181303-1-1.html&prev=search
 
There is also Navi 23 which is supposed to be midway between 80 CU Navi 21 & 40 CU Navi 10, i.e. 60 CUs.

It is also expected to be RDNA 2 like the Navi 21.

Most likely that Xbox X is based on Navi 23 with 56 CUs

I read all the above in twitter, don't have a link as of now.
 
NVIDIA and AMD are *roughly* matched on performance for the same shader count right now, so that puts this part at roughly 10% faster than a TU102-400 (Titan RTX) and almost 20% faster than a TU102-300A (RTX 2080 Ti), assuming no architectural improvements, clockspeed reductions, or bottlenecks. Could be an interesting part.
 
Wow.

Performance looks to be absolutely monstrous provided they can provide the necessary bandwidth. With RDNA 2 I'm hopeful they improved compression to make it less of a bandwidth hog. I could see this competing with ampere performance-wise, assuming nvidia hits a 30% performance improvement over turing. My main concern is the doubling of transistor count...what is doing the hardware RT???
 
5150Joker said:
If it's even 15% faster than 2080 Ti for $799 they will have a killer card for sale (once they get some of their driver issues ironed out).
Click to expand...
If a GPU with double everything of the 5700 XT can only beat a 2080 Ti by 15% then AMD has one hell of a broken architecture. RDNA 2 is also supposed to be much more efficient and clock much higher so this should absolutely crap all over the 2080 Ti. And this GPU is not being made to compete with 2080 Ti as Ampere will be on the scene as the competition. Hopefully big Navi has good scaling because the 2080 Ti certainly doesn't. Even at 4k it's only 20% faster than a 2080 Super in the most recent review on Techpowerup. And no it's not because of the clock speed that most people wrongly claim as the actual average clock speed on the FE 2080 Ti is only 95 mhz less than the FE 2080 Super.
 
Last edited:
Ricky T said:
If a GPU with double everything of the 5700 XT can only beat a 2080 Ti by 15% then AMD has one hell of a broken architecture. RDNA 2 is also supposed to be much more efficient and clock much higher so this should absolutely crap all over the 2080 Ti. And this GPU is not being made to compete with 2080 Ti as Ampere will be on the scene as the competition. Hopefully big Navi has good scaling because the 2080 Ti certainly doesn't. Even at 4k it's only 20% faster than a 2080 Super in the most recent review on Techpowerup. And no it's not because of the clock speed that most people wrongly claim as the actual average clock speed on the FE 2080 Ti is only 95 mhz less than the FE 2080 Super.
Click to expand...

Based on the power consumption of RDNA 1 the expectation should be that Navi 21 should be exact doubling of 5600 XT (reference non-overclocked bios) & not doubling of 5700 XT unless AMD managed to work in some RDNA 2 magic ...
 
Marees said:
Based on the power consumption of RDNA 1 the expectation should be that Navi 21 should be exact doubling of 5600 XT (reference non-overclocked bios) & not doubling of 5700 XT unless AMD managed to work in some RDNA 2 magic ...
Click to expand...

I think it'd be silly to expect that RDNA 2 maintains the same degree of efficiency as rdna 1.
 
AMD is really aggressive with their roadmaps right now. I expect they will have something else by the time Ampere is out, or shortly after.

As far as we know, they skipped doing a Zen 2+, looks like they skipped an RDNA1 refresh, and I expect they will go straight from RDNA2 to RDNA3.
 
sabrewolf732 said:
I think it'd be silly to expect that RDNA 2 maintains the same degree of efficiency as rdna 1.
Click to expand...
Exactly. They'll work on compression and other features to improve the capabilities without a doubt, which fits their marketing of incremental improvement strategies they have proclaimed.
 
Ricky T said:
If a GPU with double everything of the 5700 XT can only beat a 2080 Ti by 15% then AMD has one hell of a broken architecture. RDNA 2 is also supposed to be much more efficient and clock much higher so this should absolutely crap all over the 2080 Ti. And this GPU is not being made to compete with 2080 Ti as Ampere will be on the scene as the competition. Hopefully big Navi has good scaling because the 2080 Ti certainly doesn't. Even at 4k it's only 20% faster than a 2080 Super in the most recent review on Techpowerup. And no it's not because of the clock speed that most people wrongly claim as the actual average clock speed on the FE 2080 Ti is only 95 mhz less than the FE 2080 Super.
Click to expand...

You have to consider they will have RT hardware built in and other changes so there could be new bottlenecks that we have yet to see, this is certainly not Navi 1 just scaled up. I'm most curious about it's RT performance more than anything else since the Turing RT performance is pretty shit. With regards to Ampere, nobody knows how it will perform, 7nm could be a huge boost for NVIDIA or a dud, we'll find out soon enough. I don't think we'll see Big Ampere (3080 Ti) or Big Navi till late 2020.
 
N4CR said:
Exactly. They'll work on compression and other features to improve the capabilities without a doubt, which fits their marketing of incremental improvement strategies they have proclaimed.
Click to expand...

Hope for the best, & be prepared for the worst is my motto

AMd's past track record has not been good in GPU power consumption

2 things give me hope however

1. Renoir APU saw a 60% uplift in performance from Vega 12nm to Vega 7nm. AMD said they learnt a lot from it, so I hope the lessons will be baked in RDNA 2

2. Xbox X rumoured to have 56 CUs (based on Navi 23 !? ). If a console can have such a big chip then maybe AMD really cracked the GPU power consumption puzzle !?
 
Marees said:
Hope for the best, & be prepared for the worst is my motto

AMd's past track record has not been good in GPU power consumption

2 things give me hope however

1. Renoir APU saw a 60% uplift in performance from Vega 12nm to Vega 7nm. AMD said they learnt a lot from it, so I hope the lessons will be baked in RDNA 2

2. Xbox X rumoured to have 56 CUs (based on Navi 23 !? ). If a console can have such a big chip then maybe AMD really cracked the GPU power consumption puzzle !?
Click to expand...

Anytime you hear new architecture then you kind of have to throw what previous versions of other architectures did in the garbage can. There is literally no way to know especially since this will carry Ray tracing hardware now. The latest hardware from AMD has been very good on power consumption so that trend might continue, or the card may be great on power consumption in raster performance, but sucks more power with Ray Tracing on. Personally I have never cared about power consumption as long as the card preformed and did not overheat.
 
5150Joker said:
I don't think we'll see Big Ampere (3080 Ti) or Big Navi till late 2020.
Click to expand...

If this is risk production then they are 6 months from being able to go to retail. That of course assumes all is well and no re-spin is required.
 
5150Joker said:
If it's even 15% faster than 2080 Ti for $799 they will have a killer card for sale (once they get some of their driver issues ironed out).
Click to expand...

I'd buy one to tide me over until the 2080 Ti replacement comes out.
 
Fun rumor..

Assuming that it is double the transistors/CU's/ etc as navi1, I think that due to heat and power constraints, it will likely be downclocked from navi1 clocks, and will be 1.7x ish (navi) performance range (factoring in lower clocks + bump from improved IPC). I expect it will get close to or equal the 2080ti in normal raster performance, no way to predict raytracing performance.

By the time this comes out, Ampere will be coming out, and I don't know of any reliable-ish rumors on that part, but I would expect 15% to 30% gains over Turing being realistic, 40 to 50% improvement being optimistic. Going to 7nm has cost clockspeed and seems to be a wall with the power and heat(from AMD's experience), so that move for nVidia may not be the halo that previous node shrinks were. The problems with heat AMD ran into with 7nm, is going to affect new nVidia and Intel parts once they spin up on those processes... I do hold out hope that Intels' delays were due to Intel addressing these clockspeed issues with the 10nm process, and that once they finally go 7nm, clockspeed will not suffer in the way it has with AMD's parts.

Can't wait to find out :)
 
Gideon said:
Anytime you hear new architecture then you kind of have to throw what previous versions of other architectures did in the garbage can.
Click to expand...

Very much this. AMD has to integrate their first hardware RT solution into Navi; they'll need more than a few tricks to pull that off and stay competitive in raster performance too. Nvidia has a die shrink to factor in, so they're potentially better off going into a larger transister budget along with leading raster performance and working on their second RT hardware release, but again, RT is the trick. Nvidia is more ahead on the software side of things than the hardware, I think, as the hardware side of RT is pretty straightforward. The Xbox is one thing, but dealing with all of the intricacies of desktop operating systems is something that AMD has struggled more with.

Gideon said:
Personally I have never cared about power consumption as long as the card preformed and did not overheat.
Click to expand...

Less power draw, less heat, and less noise are always preferable of course. Can't always have it both ways, but this is somewhere that AMD absolutely needs to approach parity if they're trying to push into the high-end market again.

At the very least, give us a three-fan solution at release to go along with the blower solution. It would be nice to not see a repeat of the 5700XT release!
 
KazeoHin said:
I refuse to get excited about a Radeon product.

They're hype stock has been downgraded to "show me results".
Click to expand...

Well, Raja kind of does that or has that effect on folks :) I am set with 3 cards from a 5700 to a 56.
 
I mean, there are bunk rumors for almost every new release, I wouldn't put much faith it in.

BUT if it is true, AMD has a monster on their hands. I could believe they can do it.

Only question is how aggressive the pricing will be. 2020 will definitely be interesting.
 
cybereality said:
I mean, there are bunk rumors for almost every new release, I wouldn't put much faith it in.
Click to expand...

The number of compute units of the die could easily be right -- the actual performance in a shipping card after potential binning, clockspeeds are settled, power is set up and coolers are finalized?

That's a tossup. AMD has had paper monsters fall short before, and they need to overshoot by quite a bit if they're to have a chance at competing with Ampere.
 
A big wild card is in the ray tracing performance. We know it's supported on next-gen consoles, so they must have found a way to do it more efficient than current Nvidia cards.

So, there is a possible situation where Big Navi raster performance is good, within 10% of 2080 Ti, but ray-tracing is much better. That could potentially be a huge selling point.
 
cybereality said:
A big wild card is in the ray tracing performance. We know it's supported on next-gen consoles, so they must have found a way to do it more efficient than current Nvidia cards.

So, there is a possible situation where Big Navi raster performance is good, within 10% of 2080 Ti, but ray-tracing is much better. That could potentially be a huge selling point.
Click to expand...
There is mixed thought on the patent. It may be much more efficient than current Turing execution if the software can back it up. Their patent states it is flexible enough to use what is dedicated RT hardware on NV architectures as both RT or raster for AMD. Then bandwidth/feeding the architecture becomes the limiting point, or maybe another point.. who knows. But also allows good scalability either way..
Considering they probably had a golden shower from microshaft and sony, it's likely to be pretty damn good.
But who knows till they get it out.
 
N4CR said:
There is mixed thought on the patent. It may be much more efficient than current Turing execution if the software can back it up. Their patent states it is flexible enough to use what is dedicated RT hardware on NV architectures as both RT or raster for AMD. Then bandwidth/feeding the architecture becomes the limiting point, or maybe another point.. who knows. But also allows good scalability either way..
Considering they probably had a golden shower from microshaft and sony, it's likely to be pretty damn good.
But who knows till they get it out.
Click to expand...

The dedicated BVH intersection hardware in the AMD patent isn’t reusable for anything else. The reusable bit is that the RT hardware shares the same memory subsystem and caches as the TMUs. With Nvidia the BVH intersection hardware has its own memory pipeline and cache.

The most important part of the AMD patent is that the developer can write custom shader code to control behavior after each BVH node hit. This is way more flexible but will obviously be slower than nVidia’s approach of doing the full traversal in the dedicated RT core and only returning the final hit result to the shader.

It will come down to whether the flexibility enables more efficiency to compensate for the loss in raw speed.
 
erek said:
Hook it up !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

"And with those 80 CUs you’re looking at 5,120 RDNA 2.0 cores. Quite how AMD is managing the hardware-accelerated ray tracing we still don’t know. It will surely be designed to support Microsoft’s DirectX Raytracing API, given that is what will likely power the Xbox Series X’s approach to the latest lighting magic, but how that is manifested in the GPU itself is still up in the air."

View attachment 217429

https://translate.google.com/transl...phell.com/thread-2181303-1-1.html&prev=search
Click to expand...
upload_2020-1-23_9-8-40.png


upload_2020-1-23_9-11-2.png


Source?
Cause I said so... in short.
 
MangoSeed said:
The dedicated BVH intersection hardware in the AMD patent isn’t reusable for anything else. The reusable bit is that the RT hardware shares the same memory subsystem and caches as the TMUs. With Nvidia the BVH intersection hardware has its own memory pipeline and cache.

The most important part of the AMD patent is that the developer can write custom shader code to control behavior after each BVH node hit. This is way more flexible but will obviously be slower than nVidia’s approach of doing the full traversal in the dedicated RT core and only returning the final hit result to the shader.

It will come down to whether the flexibility enables more efficiency to compensate for the loss in raw speed.
Click to expand...

That is an interesting take on the patent tech. When I looked at it, I thought the patent sounded more efficient than Nvidia implementation, not less.
 
EniGmA1987 said:
That is an interesting take on the patent tech. When I looked at it, I thought the patent sounded more efficient than Nvidia implementation, not less.
Click to expand...

Fixed function hardware is almost always faster than a similar amount of general purpose hardware.
 
BrotherMichigan said:
Fixed function hardware is almost always faster than a similar amount of general purpose hardware.
Click to expand...
AMD is putting a dedicated ray tracing unit within the texture unit. The shader core sends a request to the TMU and receives back texture and lighting data. According to the patent, the shader core *can* overwrite results, but typical behavior is the RTU is the primary lighting processing unit. There is nothing general purpose about the ray trace hardware.
 
EniGmA1987 said:
AMD is putting a dedicated ray tracing unit within the texture unit. The shader core sends a request to the TMU and receives back texture and lighting data. According to the patent, the shader core *can* overwrite results, but typical behavior is the RTU is the primary lighting processing unit. There is nothing general purpose about the ray trace hardware.
Click to expand...
This process was already outlined in Vega architecture (just fyi) just didn't mention rt.
 
EniGmA1987 said:
AMD is putting a dedicated ray tracing unit within the texture unit. The shader core sends a request to the TMU and receives back texture and lighting data. According to the patent, the shader core *can* overwrite results, but typical behavior is the RTU is the primary lighting processing unit. There is nothing general purpose about the ray trace hardware.
Click to expand...

The full ray tracing pipeline isn't accelerated though, right? Just part of it.
 
BrotherMichigan said:
The full ray tracing pipeline isn't accelerated though, right? Just part of it.
Click to expand...
You could state it is.


~ This is from Vega 10 arch | which wasn't implemented anywhere as it required for programmers to implement it. (AMD cancelled its support with Vega64 release - as they couldn't get it working - in github there was some talk about problems with older shaders so they decided to not do it - note hardware is already there since vega10 release)
(left traditional vs right which ref what rt will be used as -- was primitive shader)
upload_2020-1-23_15-6-43.png


navi has dual compute unit - so we can expect it to be around 2x as fast per cu as vega is. (i.e. Radeon Vii would get 60fps with 60cu's, and 5700 xt would have around 80fps with 40cu's)
The primitive shaders are supposedly fixed in rdna.

(In short it should offer very little hit on performance, it might be better than nv implementation)
 
Last edited:
You must log in or register to reply here.
Back
Top