AMD Navi 23 ‘NVIDIA Killer’ GPU Rumored to Support Hardware Ray Tracing, Coming Next Year

[Chuk_Chuk]

It was not confirmed no, there was a 505mm2 RDNA2 die leak or rumour, it was assumed 80 CU because 40 CU 251mm2 X2 = 80 CU 502mm2.

In that case, lets say Tom's hardware is correct and we have 56 CUs in 171 mm2

505/171 = 2.95 (Lets call it 2.8)
56*2.8 = 156.8 (round down 156)

Big Navi could have 156 CUs

surely something has to be missing, or has removing all the legacy vega hardware had a huge impact on die size.

 

[humbug]

In that case, lets say Tom's hardware is correct and we have 56 CUs in 171 mm2

505/171 = 2.95 (Lets call it 2.8)
56*2.8 = 156.8 (round down 156)

Big Navi could have 156 CUs

surely something has to be missing, or has removing all the legacy vega hardware had a huge impact on die size.

With respect, to be blunt... not happening.

I think by a couple of week past Nvidia's paper launch we will know.

 

[Chuk_Chuk]

With respect, to be blunt... not happening.

I think by a couple of week past Nvidia's paper launch we will know.

In hindisght i have missed out the memory bus and any ancillary items. However that still puts it well over 80 CUs (Assuming tom's hardware is correct 80 CUs works out to be 244mm2).

 

[Davedree]

It was not confirmed no, there was a 505mm2 RDNA2 die leak or rumour, it was assumed 80 CU because 40 CU 251mm2 X2 = 80 CU 502mm2.

Yeah and that's the mistake with whoever is doubling x2 Navi die's, the actual die size could be 20-30% smaller for a theoretical 80cu Navi as you would only need to scale up the the Cu's and the caches and memory controller geometry engines, rops etc etc. The I/0, display controller, Vce engine and other aspects which take up die space and don't need to be scaled up will reduce the final die size.

It's the same as someone saying Fiji wiould have been 720mm because Tonga 360mm x2 is 720mm, Fiji was 596mm

 

[BigBANGtheory]

In that case, lets say Tom's hardware is correct and we have 56 CUs in 171 mm2

505/171 = 2.95 (Lets call it 2.8)
56*2.8 = 156.8 (round down 156)

Big Navi could have 156 CUs

surely something has to be missing, or has removing all the legacy vega hardware had a huge impact on die size.

something isn’t right there have they accounted for the memory bus/control/caching?

80 CU is imho the largest sensible prediction, probably tested by AMD too but who is to say if it becomes a proper retail product

 

[LoadsaMoney]

Xbox Series X Gets Ray Tracing & Machine Learning Specs, Boasting 3-10X Performance Improvement

https://wccftech.com/xbox-series-x-...specs-boasting-3-10x-performance-improvement/

 

[Chuk_Chuk]

something isn’t right there have they accounted for the memory bus/control/caching?

80 CU is imho the largest sensible prediction, probably tested by AMD too but who is to say if it becomes a proper retail product

No i didn't in my initial calculation. If we assume that 25% of the die space is taken up by that stuff (Is this an over or under estimate?) it comes out as ~120 CUs.

Is there anything else that could be mssing from the calculations?

 

[Davedree]

No i didn't in my initial calculation. If we assume that 25% of the die space is taken up by that stuff (Is this an over or under estimate?) it comes out as ~120 CUs.

Is there anything else that could be mssing from the calculations?

I'm a bit confused by that website and it's probably me, although I re-read it numerous times.
They are saying the whole die cpu+gpu is 360mm, and then the gpu is roughly 47.5%, so 171mm ish?

What's unclear is the L1/L2 cache used in the gpu for the Xbox, theoretically it would be 8mb L2 1mb L1 for 80cu's, so is it using the 4mb/512kb like Navi 10 or is it something like 6m/768kb?

 

[BigBANGtheory]

Is there anything else that could be mssing from the calculations?

https://www.extremetech.com/wp-content/uploads/2019/06/Navi-Slide-2.jpg

 

[Chuk_Chuk]

I'm a bit confused by that website and it's probably me, although I re-read it numerous times.
They are saying the whole die cpu+gpu is 360mm, and then the gpu is roughly 47.5%, so 171mm ish? Yep that is how I am now interpreting it

What's unclear is the L1/L2 cache used in the gpu for the Xbox, theoretically it would be 8mb L2 1mb L1 for 80cu's, so is it using the 4mb/512kb like Navi 10 or is it something like 6m/768kb?

You may get answers to the cache question when the full presentation is done. From the eurogamer tear down from a few months back they've stated that there is 76mb of Sram across the entire SoC.

 

[Grim5]

Xbox Series X deep dive by Microsoft

https://www.tomshardware.com/news/microsoft-xbox-series-x-architecture-deep-dive

Couple good tidpoints

* The Zen 2 CPU in the Xbox only has 8mb L3 cache and is monolithic - what we suspected has been confirmed, next gen consoles are not using a 3700x, they are using a downgraded 4800u/h so they they are slower than the zen 2 desktop CPUs.


* Microsoft confirmed the Series X is significantly more expensive to make than a One X

* The Xbox Series X's Ray Tracing output is rated at "380g/second Peak in Ray Box". Does anyone know what the output of RTX cards are for comparison???

* The Series X's RDNA 2 GPU consumes the same power as the GPU in the One X - 175w TDP

 

[gpuerrilla]

Interesting @Grim5 - that's pretty nice GPU power, im wondering what the dGPU equivelent would be? ~ 10% faster than a 5700XT but on RDNA2. Maybe expect a flavour around the performance of a 5800 for £450

 

[Nutella33]

Couple good tidpoints

Also mention of an ML resolution upscaling like DLSS.

Does anyone know what the output of RTX cards are for comparison??

10Grays/sec apparently.
But we have confusion on whether they are comparable figures (peak, average, use case). It's a similar jump than between the 1000 series and 2000 series apparently.

 

[CAT-THE-FIFTH]

Videocardz has more of the slides:
https://videocardz.com/90788/microsoft-provides-more-details-xbox-series-x-architecture

Looking at Beyond3D it appears RDNA1 can do 5 instructions per clock,and RDNA2 7 instructions per clock. Apparently part of the RT pipeline is shared with the texture units:
https://forum.beyond3d.com/threads/amd-navi-speculation-rumours-and-discussion-2019.61042/page-127

You can see it in one of the slides,4 texture or RT ops at the same time. So some of the peak RT figures might in reality be lower,once you start switching between texture and RT operations. However,this has apparently been done to reduce the number of transistors,and hence the die area. So I suspect Nvidia will be more brute force and AMD is more about efficiency??

 

[Besty]

agreed, just keep your eye set on a 40-50% speed bump, which is realistic and achievable if it’s more then that then happy days - it’s a
bonus.

 

[Grim5]

Videocardz has more of the slides:
https://videocardz.com/90788/microsoft-provides-more-details-xbox-series-x-architecture

Looking at Beyond3D it appears RDNA1 can do 5 instructions per clock,and RDNA2 7 instructions per clock. Apparently part of the RT pipeline is shared with the texture units:
https://forum.beyond3d.com/threads/amd-navi-speculation-rumours-and-discussion-2019.61042/page-127

You can see it in one of the slides,4 texture or RT ops at the same time. So some of the peak RT figures might in reality be lower,once you start switching between texture and RT operations. However,this has apparently been done to reduce the number of transistors,and hence the die area. So I suspect Nvidia will be more brute force and AMD is more about efficiency??

Right, so we can't yet compare the Xbox and Nvidias performance based on that info. Nvidias 11g Ray/sec figure is i
From its own RTX benchmark while the xboxes 320g peak per second is from a DXR benchmarked called Ray Box and they are specifically quoting maximum rather than average. Hard to compare two different methods of ray tracing - what we have to compare is framerate when the time comes but by then we'll have desktop rdna 2 to try anyway

The videocardz details also makes mention that rdna2 does have a dedicated hardware accelerator for ray tracing, but those same accelerator units can be used for rasterisation when ray tracing is not in use. That is definitely more efficient than Nvidia does - it's estimated that a 2080ti would be 10 to 15% faster if it's RT die space was replaced with Cuda cores - so that is a significant benefit to having a unit that can do BVH acceleration and Rasterisation

 

[EastCoastHandle]

It's like watching a poker match between MS, Sony, AMD, Nvidia. Rumor has it that the xsx is a 'template' for RDNA 2.

Highlights:
-XSX uses a Zen 2 Server Class CPU (surprised at this one)
-Machine Learning Acceleration
-Variable Rate Shading (we knew that)
-HDMI 2.1 new monitors for everyone
-380G/sec ray-box peak, 95G/sec ray-tri peak.
-Shader run parallel for BVH traversal, material shading, etc
-N7 Enhanced at 360.4 mm2
-15.4B transistors

----
A best buy store was looted in Chicago (Aug 10) and the new xbox controllers were stolen and posted. So Xbox is already geared up for release. So I have to assume that RDNA 2 is also just as ready if it's suppose to launch before console.

 

[CAT-THE-FIFTH]

Right, so we can't yet compare the Xbox and Nvidias performance based on that info. Nvidias 11g Ray/sec figure is i
From its own RTX benchmark while the xboxes 320g peak per second is from a DXR benchmarked called Ray Box and they are specifically quoting maximum rather than average. Hard to compare two different methods of ray tracing - what we have to compare is framerate when the time comes but by then we'll have desktop rdna 2 to try anyway

The videocardz details also makes mention that rdna2 does have a dedicated hardware accelerator for ray tracing, but those same accelerator units can be used for rasterisation when ray tracing is not in use. That is definitely more efficient than Nvidia does - it's estimated that a 2080ti would be 10 to 15% faster if it's RT die space was replaced with Cuda cores - so that is a significant benefit to having a unit that can do BVH acceleration and Rasterisation

Until we can run a standardised benchmark it will be hard to say how both compare. It appears RDNA2 is trying to use a more area efficient and flexible form of hardware for their RT calculations. This fits very well with what consoles need to do,ie,have a reasonable sized SOC,which can be flexibly used over the lifespan of the console. Apparently the 52CUs in the XBox Series X SOC,uses under 140W at 1.8GHZ!

I do have a feeling Nvidia might be using a more brute force way of doing RT,and possible going for larger dies. So it wouldn't surprise me if Nvidia is still faster in RT with Ampere,but might use more power. Nvidia always pushes for larger dies,than AMD feels comfortable with.

OTH,if RDNA2 is indeed area efficient,it might help on the pricing front for the RDNA2 GPUs,if AMD can have a die size advantage. If power consumption is also better,it should also help in other ways.

OFC,this is all speculation on my part - we will have to wait and see how things pan out in reality.

 

[humbug]

It's like watching a poker match between MS, Sony, AMD, Nvidia. Rumor has it that the xsx is a 'template' for RDNA 2.

Highlights:
-XSX uses a Zen 2 Server Class CPU (surprised at this one)
-Machine Learning Acceleration
-Variable Rate Shading (we knew that)
-HDMI 2.1 new monitors for everyone
-380G/sec ray-box peak, 95G/sec ray-tri peak.
-Shader run parallel for BVH traversal, material shading, etc
-N7 Enhanced at 360.4 mm2
-15.4B transistors

----
A best buy store was looted in Chicago (Aug 10) and the new xbox controllers were stolen and posted. So Xbox is already geared up for release. So I have to assume that RDNA 2 is also just as ready if it's suppose to launch before console.

Can i join in?

 

[EastCoastHandle]

Can i join in?

sure, there's room