• 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.

Radeon VII

I don’t think he’s gimping the performance, indeed he said plus 5%. The good thing about undervolting is you can choose a quiet card with less performance or boost the performance with an increased overclock.

My 56 easy went to stock 64 performance once tweaked. It was poor out the box though. I am a silent pc freak and so undervolting and reducing performance for less dB does appeal to me. Unfortunately in my experiments to get the sapphire pulse 56 completely silent at load (as in no fans or ultra low rpm) you have to cripple the performance.

That is one benefit of the lower tdp Turing cards, couple with a beefy triple slot cooler and you literally get 0db at full load.

That's a shame, I was looking at the Vega Pulse and wondering if it was like the Sapphire Fury Pro Tri-x, I had the Sapphire Fury Tri-x oc edition and that ran at it's claimed clocks out of box & it did so silently, It even did better than cards like the Asus Fury Strix and the Sapphire Fury Nitro that were both redesigns that reduced the open back end or blocked it off completely like the Strix model did. At the time I put it down to it having better airflow through the open back end like the Vega Pulse has so if it hasn't done the Pulse any favours that's a shame, Maybe if they'd given it the longer Tri'x set-up with one third of it's length having an open back end it would have done better.
 
Just a random thought I had while smoking a cig in the freezing cold had to write it down to be first

2080ti is £1000
Just for simplicity the core and ram is 700 the tensor is 100 and the rt core cost 200

My idea

Make a card that is 2 the amount of rt cores only and have it in the second slot for sli

Then buy one of these and a 2060 for amazing raytracing. XD

This has been discussed at length in various RTX threads. Basically it won't work because even NVLink (of which the 2060 doesn't have) does not have anywhere enough bandwidth to pass RT data back and forth between the cards. Plus, it's the RT and Tensor hardware that makes RTX cards expensive to begin with, so if you're talking about a dedicated RTX processor with TWICE the processing power of a full RTX card then it'd be even more stupid money than they are right now.

So no.
 
It is tempting but waiting until we see performance & price is the smarter move as we can get a decent 2080 for under £700 so that might be the better option.

I want to play two of those three games and the game offer wont last more than couple of weeks.
As for reviews we saw them with Vega. Having some muppets reviewers putting the floor voltage to 1200mv and then complaining that the card is overheating having low clocks and performance.
 
This has been discussed at length in various RTX threads. Basically it won't work because even NVLink (of which the 2060 doesn't have) does not have anywhere enough bandwidth to pass RT data back and forth between the cards. Plus, it's the RT and Tensor hardware that makes RTX cards expensive to begin with, so if you're talking about a dedicated RTX processor with TWICE the processing power of a full RTX card then it'd be even more stupid money than they are right now.

So no.

I don't think the RTX and tensor cores have added that much to the overall cost. Maybe 20% of the die area but on a mature process the Turing chips probably don't cost that much more than Pascal equivalents.

Importantly, the RTX hardware depends on the the CUDA cores as well for developing dynamic bounding volume hierarchies, otherwise performance would be hudnred of times slower.

And all of this ignores the point that apart form RTX and tensor cores, trying has been seriously reworked in the back end improving scheduling, scaling, IPC, bandwidth and performance. Turing isnlt just Pascal with RTX bolted on, it is a whole new architecture with significant improvements and new features only for rasterization such as variable rate shading a special programmable geometry subsystem
 
Plus, it's the RT and Tensor hardware that makes RTX cards expensive to begin with, so if you're talking about a dedicated RTX processor with TWICE the processing power of a full RTX card then it'd be even more stupid money than they are right now.

So no.
Tensor & RT cores are not some special entities made by unicorns blood.

AMD normal "cores" in matrix design found in Vega & FuryX operate like tensor cores on computing application. Yet they can work like "normal" cores also shall the application running on it demands it.
Yet the cards aren't expensive given their more complex multi purpose design.
 
I want to play two of those three games and the game offer wont last more than couple of weeks.
As for reviews we saw them with Vega. Having some muppets reviewers putting the floor voltage to 1200mv and then complaining that the card is overheating having low clocks and performance.

The games are a good incentive but I already have RE2 and DMC5 doesn't interest me, nor does The Division 2 if I'm honest, I played the original but never bothered buying it, The 2080 also comes with two games now, Anthem & Battlefield 5.
 
Tensor & RT cores are not some special entities made by unicorns blood.

AMD normal "cores" in matrix design found in Vega & FuryX operate like tensor cores on computing application. Yet they can work like "normal" cores also shall the application running on it demands it.
Yet the cards aren't expensive given their more complex multi purpose design.


Tensor cores are very different to either CUDA cores or AMD's stream processors.

That is why they exist in the first place, dedicated hardware to accelerate specific functions.

You can compute ray-tracing in the CUDA cores/Stream processors, but doing so is on the order of 10-20x slower since it is run in software without specific acceleration of the intersection test, and moreover, all of those CUDA cores/Stream processors get taken up doing ray tracing instead of working on traditional razerization which is a requirement for the next 5-7 years until RTX hardware advances enough that the traditional rasterization can be entirely removed.
 
Tensor cores are very different to either CUDA cores or AMD's stream processors.

That is why they exist in the first place, dedicated hardware to accelerate specific functions.

You can compute ray-tracing in the CUDA cores/Stream processors, but doing so is on the order of 10-20x slower since it is run in software without specific acceleration of the intersection test, and moreover, all of those CUDA cores/Stream processors get taken up doing ray tracing instead of working on traditional razerization which is a requirement for the next 5-7 years until RTX hardware advances enough that the traditional rasterization can be entirely removed.

He mentioned also Tensor cores..... Which are CUDA cores designed for 4x4 FP16 matrices, in essence. Something AMD achieves using its Vega Stream processors that doesn't need those "special" cores.

As for DXR requires the cards to be DX12.1 compliant. No "special" cores. Nothing.
So nobody knows what AMD can do to utilize the computing power of the Vegas. However they aren't doing so because they stick to "all lineup should support ray tracing".
 
Last edited:
D.P. and Panos

My point was a dedicated RTX processor card wouldn't work because you couldn't shuttle the required data between cards fast enough, or at least that's what I understood from discussions that took place on the RTX threads here a while ago.
 
He mentioned also Tensor cores..... Which are CUDA cores designed for 4x4 FP16 matrices, in essence. Something AMD achieves using its Vega Stream processors that doesn't need those "special" cores.

As for DXR requires the cards to be DX12.1 compliant. No "special" cores. Nothing.
So nobody knows what AMD can do to utilize the computing power of the Vegas. However they aren't doing so because they stick to "all lineup should support ray tracing".


No, Tensor cores have nothing to do with CUDA cores and they aren't similar in the slightest. CUDA cores are general purpose arithmetic units and fully programmable. Tensor cores are fixed function accelerators.

You can simulate e what a tensor core does in software with CUDA cores or AMD stream processors, but without the optimization of fixed hardware pipeline performance is aorund 10x slower.


There is no special hardware to be DX12 compliant full stop, you could run the whole thing on a CPU. This is basically what Intel attempted with larrabee and see how far that got them. there is a reason we have GPUs and not just CPUs. Dedicated hardware is much faster at doing the specific use case work.
 
D.P. and Panos

My point was a dedicated RTX processor card wouldn't work because you couldn't shuttle the required data between cards fast enough, or at least that's what I understood from discussions that took place on the RTX threads here a while ago.

You could make a dedicated RTX processor, but the design of RT in Turing makes heavy use of the CUDA cores to do a lot of auxiliary processing. A deeicated RTX processor would ahve to duplicate a lot of the GPU, as well as have its own memory duplicated. The cost would be high and there would be no gains.


What would work is general mutli-GPU RTX using multiple Turing cards because Ray-tracing scales much better the complex hacky rasterization that we now have with heavy inter-frame dependencies. If someone wants double RTX performance then 2 Turing cards in SLI could be made to do that, but it would also be expensive.

In the end, with 7nm slowly maturing a large 7nm die will be possible next year. It is likely Nvidia could at least double the number of RTX cores, along with design improvements and a 25% clock bump from 7nm process you might get 4x the total throughput
 
Back
Top Bottom