Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[Glo.]

It literally cannot be. They already brought out new Navi/RDNA1 products this year. They just brought out the mobile versions. But you guys can keep going in circles debating if they'll overhaul their entire GPU stack with RDNA2 this year. At least the talk about APUs going chiplets and/or jumping to the latest GPU has some reason to consider it.

Dr. Lisa Su have said such quote in earning call on 28th January of this year. Few days AFTER, last Navi/RDNA1 GPU has been released.

So she meant that Navi lineup will get refresh after that product stack. In general I agre with you, it mostly implies that Navi 1 is part of AMD GPU lineup this year, and on top of it come RDNA2 GPUs.

Well I do understand English pretty well, not as good as Native British peopleP), but for me the said quote from Lisa:

In 2019, we launched our new architecture in GPUs, it's the RDNA architecture, and that was the Navi based products. You should expect that those will be refreshed in 2020 - and we'll have a next generation RDNA architecture that will be part of our 2020 lineup. So we're pretty excited about that, and we'll talk more about that at our financial analyst day. On the data centre GPU side, you should also expect that we'll have some new products in the second half of this year.

Especially the bolded part implies specifically that RDNA1 will get refresh in their lineup, and RDNA2 will come on top of it.

 

[Paul98]

I expect RDNA1 won't be around for long, looks like they are doing major architecture changes in a couple phases. Usually it's when they do more incremental changes they re-badge.

 

[DisEnchantment]

I expect RDNA1 won't be around for long, looks like they are doing major architecture changes in a couple phases. Usually it's when they do more incremental changes they re-badge.

Not to mention several HW bugs which need workaround in SW. Looking at the LLVM documentation you can see how many bugs are there that the compiler has to work around.
I won't be surprised, in fact I expect there is a decent "IPC" lost due to the bugs because the compiler has to generate a non optimized code path everytime there is an execution sequence that could lead to these bugs happening.

FeatureVcmpxPermlaneHazard,
FeatureVMEMtoScalarWriteHazard,
FeatureSMEMtoVectorWriteHazard,
FeatureInstFwdPrefetchBug,
FeatureVcmpxExecWARHazard,
FeatureLdsBranchVmemWARHazard,
FeatureNSAtoVMEMBug,
FeatureOffset3fBug,
FeatureFlatSegmentOffsetBug,
FeatureLdsMisalignedBug

Then there are freedesktop bug reports of SDMA buffer corruptions, unnecessary cache invalidation due to inconsitency with HW. I think some of Marek's new commits attempt to make workarounds for these.
Navi12 has some of the bugs solved. But I don't know what product it is.
The moment there is new architecture available, it is safe to expect AMD will drop RDNA1 for new products.

 

[Stuka87]

Not to mention several HW bugs which need workaround in SW. Looking at the LLVM documentation you can see how many bugs are there that the compiler has to work around.
I won't be surprised, in fact I expect there is a decent "IPC" lost due to the bugs because the compiler has to generate a non optimized code path everytime there is an execution sequence that could lead to these bugs happening.

Then there are freedesktop bug reports of SDMA buffer corruptions, unnecessary cache invalidation due to inconsitency with HW. I think some of Marek's new commits attempt to make workarounds for these.
Navi12 has some of the bugs solved. But I don't know what product it is.
The moment there is new architecture available, it is safe to expect AMD will drop RDNA1 for new products.

These types of work arounds are pretty par for the course for any GPU, or really, any sort of hardware. There are tons of compiler fixes for bugs found in CPU's.

 

[DisEnchantment]

These types of work arounds are pretty par for the course for any GPU, or really, any sort of hardware. There are tons of compiler fixes for bugs found in CPU's.

I am not really sure if you are following llvm/mesa/amdgpu/amdkfd development. You can look it up yourself.
In my view, Navi has no precedent in the amounts of workarounds in software for GPUs.

 

[exquisitechar]

I am not really sure if you are following llvm/mesa/amdgpu/amdkfd development. You can look it up yourself.
In my view, Navi has no precedent in the amounts of workarounds in software for GPUs.

Yes, the number of workarounds is unusual from what I've seen. Either way, if the rumored die sizes for the Navi 2x lineup is true, and I believe they probably are at least not far off the truth, RDNA is pretty much dead. Navi10 is made completely useless by Navi23, and we already have a leak that shows that Navi23 will be paired with Cezanne in laptops next year (no wonder, when RDNA laptop offerings are underwhelming). Navi21 and Navi22 offer a whole new level of performance. Only Navi14 may survive for the time being.

 

[Ajay]

Eh, hopefully AMD got out the whips for the Chinese driver dev group. Then RDNA2 products will stand out.

 

[trexfromouterspace]

Navi 1x is abnormally buggy. I would definitely classify it as a shitshow.

 

[Glo.]

I am not really sure if you are following llvm/mesa/amdgpu/amdkfd development. You can look it up yourself.
In my view, Navi has no precedent in the amounts of workarounds in software for GPUs.

Yep. However, funnily enough Navi Linux drivers still appear to be more stable than they were for Windows...

 

[DisEnchantment]

I think this is a known fact in other industries (IoT/Automotive/Mobile et al.) NVIDIA is struggling a lot against Qualcomm/Samsung/MediaTek due to lack of good NAD(Network Access device)/5G integration. NVIDIA has lost significant presence in tablets and smartphones.
CPU wise, Tegra is run of the mill but GPU/Tensor/Accelerated computing etc they have a strong position which is why Tegra is still a strong product overall.

But it seems for a mobile Gaming console, they could be facing stiff competition as well from RDNA. RDNA in Exynos will challenge the remaining strong point of Tegra besides the fact that Exynos has the lead with better NAD/5G/Connectivity features. Software support is key, but if AMD continues to provide strong support for Open Source for their mainstream architectures, taking this to Exynos is fairly easy considering that a lot of the platforms like Android/Automotive/IoT runs some flavor of Linux as the base.

https://twitter.com/x/status/1259011509432094721

 

[Leadbox]

I think this is a known fact in other industries (IoT/Automotive/Mobile et al.) NVIDIA is struggling a lot against Qualcomm/Samsung/MediaTek due to lack of good NAD(Network Access device)/5G integration. NVIDIA has lost significant presence in tablets and smartphones.
CPU wise, Tegra is run of the mill but GPU/Tensor/Accelerated computing etc they have a strong position which is why Tegra is still a strong product overall.

But it seems for a mobile Gaming console, they could be facing stiff competition as well from RDNA. RDNA in Exynos will challenge the remaining strong point of Tegra besides the fact that Exynos has the lead with better NAD/5G/Connectivity features. Software support is key, but if AMD continues to provide strong support for Open Source for their mainstream architectures, taking this to Exynos is fairly easy considering that a lot of the platforms like Android/Automotive/IoT runs some flavor of Linux as the base.

https://twitter.com/x/status/1259011509432094721

Won't nvidia's recent acquisition of Mellanox help with this?

 

[DisEnchantment]

Won't nvidia's recent acquisition of Mellanox help with this?

I will do slight OT and I hope we don't take it too far here. You have to admire NVIDIA for putting up a tough fight in all sectors.

In HPC and supercomputing, they are now fighting Intel besides AMD. Both AMD and Intel have CPU+Accelerator integration advantage.
In Mobile/IoT/Automotive fighting against Qualcomm/Samsung/MTK and many more who are not exactly tiny. AMD is providing Radeon IP to Samsung. Automotive is their last stronghold here.
In Consumer graphics they are leading vs AMD currently but absent in high performance consoles.
In Cloud Gaming they are fighting against behemoths like MS/Google/Sony. Not sure if they will succeed much here. e.g. MS have their own DC, own SW ecosystem and they own a lot of studios.
Got kicked out of Apple ecosystem.

Mellanox acquisition brought NVIDIA new competitors. HPE(Cray Slingshot/Shasta), Intel(wrt Infiniband) just to name a few. And they are really massive competitors.
It is not helping them in their core markets. But they have to diversify because there is only so much to grow.
They have a good chance to make it big in the ARM HPC crowd. Problem there is that there too much diversity and there has to be some consolidation before we can see big gains.

 

[moinmoin]

Nvidia's biggest "problem" is that they like to enforce their own horizontal integration together with their known proprietary vertical integration without being able to actually offer complete packages. Their vertical software stack is what they sell, and the hardware is the entry drug. But the hardware vendors needed to complete Nvidia based systems now are all well on the way to offer more complete horizontal integration with more open vertical stacks.

The interesting part is that this is an industry wide movement of which AMD and Intel are just two participants. I feel Nvidia with its current business model hasn't found its place in that yet, essentially living off its past success (mostly in pushing CUDA).

 

[darkswordsman17]

I think this is a known fact in other industries (IoT/Automotive/Mobile et al.) NVIDIA is struggling a lot against Qualcomm/Samsung/MediaTek due to lack of good NAD(Network Access device)/5G integration. NVIDIA has lost significant presence in tablets and smartphones.
CPU wise, Tegra is run of the mill but GPU/Tensor/Accelerated computing etc they have a strong position which is why Tegra is still a strong product overall.

But it seems for a mobile Gaming console, they could be facing stiff competition as well from RDNA. RDNA in Exynos will challenge the remaining strong point of Tegra besides the fact that Exynos has the lead with better NAD/5G/Connectivity features. Software support is key, but if AMD continues to provide strong support for Open Source for their mainstream architectures, taking this to Exynos is fairly easy considering that a lot of the platforms like Android/Automotive/IoT runs some flavor of Linux as the base.

https://twitter.com/x/status/1259011509432094721

I'm pretty skeptical of that, although I wouldn't rule it out. I don't think the Samsung deal would have much to do with it directly. I think there's two reasons for that. I have a hunch that Nintendo would be a bit averse as I'd think Samsung would want a lot of say in things (and Nintendo likes to have extensive control) or be looking to leverage it for their own stuff as well (which didn't seem to hurt Nvidia, but Nvidia's Shield as a gaming console competitor seemed to have kinda flopped with it being mostly popular for media streaming), and I think Nintendo would just work directly with AMD. It'd allow Nintendo to own the IP more, they have a history of working with AMD (and seem to view it favorably, I just think AMD simply didn't have an ARM chip for Nintendo to sample).

I'm not sure Nintendo cares a ton about 5G, but I could be wrong, especially after the success of Pokemon Go, where it doesn't have to be constant competitive gaming but rather just needs to be able to ping a server now and then coupled with GPS stuff.

I think it'll be tough to get Nvidia out of that position too. I think Nvidia views it as a premium, views its a key to their gaming future (as Nintendo as a brand is remaining and arguably more popular than ever). But I also think they view it as the future of gaming. Portable systems that leverage cloud compute. I'd expect they'd be willing to make a custom chip, and also be a lot more competitive on price in order to keep AMD out.

Won't nvidia's recent acquisition of Mellanox help with this?

Isn't Mellanox more the backend/infrastructure side of things though? The issue Nvidia had is they were lacking in modem IP. I doubt that has changed, and I also doubt Qualcomm is any more receptive to licensing than they were back then, although maybe its easier on 5G (supposedly Apple is gonna develop their own 5G modem).

 

[Ajay]

Mellanox is an infinity band solution. It offers in band compute features that reduced HPC CPU workloads and improve network performance; it is used in allot of current supercomputers. CUDA is Nvidia's golden handcuffs. It's their greatest success, but may turn out, in some ways, to limit their flexibility. Having been left out of some of the most recent high profile supercomputer contracts, it will be interesting to see how they move forward (I don't follow the market, maybe NV are grabbing a lot of small to mid sized supercomputers).

 

[DisEnchantment]

I'm pretty skeptical of that, although I wouldn't rule it out.

I'm skeptical about it too. Especially for Nintendo, but in Automotive, Exynos is a good platform now. Samsung BSP, linux drivers, meta layers are fully open source afaik and allows customisation of everything. In contrast the other guy give you binaries and header files.
If Exynos Auto will launch with Radeon IP, I bet a lot of engineers will be delighted.

I know more here but I will have to refrain myself.

I don't follow the market, maybe NV are grabbing a lot of small to mid sized supercomputers

They have not been winning high profile deals lately except a few smaller ones like NERSC Perlmutter from Cray(uses Cray Slingshot) and SDSC Expanse by Dell (uses Mellanox Infiniband) etc. Both systems use EPYC Rome.
HPE/Cray has been bagging most of the high profile deals these days , almost all of these are Cray Shasta systems with Slingshot interconnect. Intel have their own Infiniband solutions.

 

[moinmoin]

Nintendo isn't going to switch away from Nvidia unless whatever company replaces it offers a toolchain support equal to Nvidia's (so not going to happen, especially with Switch pretty much dependent on NVN).

 

[soresu]

Nintendo isn't going to switch away from Nvidia unless whatever company replaces it offers a toolchain support equal to Nvidia's (so not going to happen, especially with Switch pretty much dependent on NVN).

I'd imagine that there is a lot more invested in the game engines ported to Switch that support NVN than any nVidia specific toolchains related to NVN itself - id Tech, UE4, Unity and RED engine to name a few.

If those engines are built well for abstraction it should not be such a great chore to switch to another API (heheh pun), perhaps even a Vulkan derivative now that there is an open low level API to build a custom API from.

 

[DisEnchantment]

Since @Glo. is causing agitation in other threads, I thought I might add these slides here so he can get busy here instead

AMD explaining why DXR1.1 (DirectX 12 Ultimate) is more suitable on RDNA2 (vs DXR1.0).



If you recollect, RDNA2's RT HW is inside the CU in contrast to Turing which have them external to the shading unit.



AMD's 20.10 driver is internally being tested and is the version to work with Windows 2004 May update and also includes WDDM 2.7 support.



I think RDNA2 performance in DXR1.0 will tank considerably.

 

[uzzi38]

Since @Glo. is causing agitation in other threads, I thought I might add these slides here so he can get busy here instead

AMD explaining why DXR1.1 (DirectX 12 Ultimate) is more suitable on RDNA2 (vs DXR1.0).

View attachment 20840View attachment 20841

If you recollect, RDNA2's RT HW is inside the CU in contrast to Turing which have them external to the shading unit.
View attachment 20843

AMD's 20.10 driver is internally being tested and is the version to work with Windows 2004 May update and also includes WDDM 2.7 support.

View attachment 20842

I think RDNA2 performance in DXR1.0 will tank considerably.

AMD's is in the TMUs to be specific, and so is tied to every 16 shaders (or 4 per CU) you're right.

Just one thing, it's not so much that AMD's implementation is better for DXR1.1 as far as I know - in current DXR RDNA2 still fares rather well compared to Turing. However DXR1.1 is uh, a smarter (and faster) way of going about things that RDNA2 makes possible. I said before that I hope Ampere makes some changes to the RTRT pipeline, and DXR1.1 is why.

EDIT: Editing in a bit more, one sec...

Remember the Minecraft RTRT demo that Microsoft showed off when talking about the Series X? Then remember Minecraft RTX performance across multiple GPUs? Remember, only one dev worked on that demo over a month - it's not like they managed to use DXR1.1 on it - that was a port of the DXR version tweaked to work on the Series X. And one more thing to note would be Digital Foundries comment on that demo.

Loosly quoting: "It feels like a video that's been encoded to a lower framerate". The video provided to the press was a 1080p30 video. Enjoy connecting the dots

 

[DisEnchantment]

I noticed that Demo too a while ago. They removed the possibility to allow peeps to go frame by frame and estimate fps/performance by providing an encoded video instead of a capture.
Regarding the DXR1.1, they mentioned specifically that the data being cached in the CU is useful because it does not need to be fetched externally again.
The combined compute pass of DXR1.1 can allow RDNA2 to gain some performance in this regard whereas it will make no difference to Turing.

Also the number of intersections per second it can do is
Frequency x Number of CU x 4 (4x Ray intersection engine per CU). For XSX = 1825 MHz x 52 x 4 = 380 billion intersections/sec.

 

[Glo.]

Being taken to the board, I will leave you one quote I got some time ago about RDNA2 RT capability.

"Its about 1.5 times the Turing's capability."

Make from it whatever you wish.

 

[DisEnchantment]

"Its about 1.5 times the Turing's capability."

Hard to believe that. XSX can do 380 billion intersection tests per second. Assuming 20 intersections per ray(this is being very generous and it is very scene dependent), that is already 2X the quoted value of Turing's 10 Gigarays/s
And we don't even know how Turing's 10 Gigarays was calculated.
At least the values for XSX are more meaningful.
A proverbial Big Navi could be a different beast compared to the XSX. However, AMD did say the gains are diminishing with the increase of the number of shader resources.

PS:
I am a noob with computer graphics. I am mainly a Linux compute guy... But I have started reading something about this for a bit. In times of dearth of new info and COVID-19, it is just natural progression

 

[Glo.]

I won't go in any direction on the topic here. All I can do is to report what has been told to me. There are factors in which it might be... plausible, like "RDNA2's RT capability is 1.5 Times of Turing, per SM". But im not technical enough to even speculate on the topic, so I will let more educated people, like you are, to speculate what could it mean.

 

[soresu]

I think RDNA2 performance in DXR1.0 will tank considerably.

Hardly bothersome as DXR1.0 will end up as D3D10.1 did - ie nowhere.

All the games supporting it will end up being patched to DXR1.1, for the basic reason of performance if nothing else - both nVidia and MS jumped the gun on real time RT and did it few favours in the long run.