AMD’s new RDNA GPU Architecture has Been Exclusively Designed for Gamers

[FordGT90Concept]

GPGPU physics isn't a memory problem, it's an API problem. There's really only two popular physics API: PhysX (which only GPGPUs on NVIDIA hardware) and Havok (never finished GPGPU implementation). There's no all-inclusive GPGPU API which means you're making costly code paths for specific hardware which is a practice that died out in the 1990s. Even if Microsoft debuts DirectPhysics (only GPGPU agnostic solution on the horizon), it won't see broad adaption because physics code made for Windows likely won't work on Mac OS X, Linux, nor PlayStation so 1990s comes calling again.

GPGPU physics just isn't worth the headaches for 99.98% of developers out there; it's a novelty and not a strong foundation for game design.


Oh, and CPU-GPU bandwidth definitely isn't a sticking point for it. All the CPU communicates is the inputs and the outputs: you're basically just throwing vectors back and forward for objects--trivial.

 

[Vya Domus]

Moving even as little as 100mb takes about 10ms across a PCI-e 3.0 connection, let's say you want 144hz since that's all the rage these days. One frame takes 6.9 ms, but your memory transfer alone for GPGPU purposes takes 10ms , ouch. Now you may not need to move 100mb all the time, but you can easily see how this doesn't quite work very well if you want high framerates, irrespective of how fast the GPU itself is. And let's not even bring in the matter of latency.

This is totally a memory transfer problem and because of that concurrent graphics and GPGPU remains problematic. Take a guess as to why most have given up on GPGPU implementations, it can't all be because of lack of interest, there are technical limitations that prohibit this. Even PhysX implementations usually pertain to small scale effects here and there which can be run on a CPU just fine a lot of the time despite the obvious computational advantage that they have on GPUs.

 

[Vayra86]

absolutely.66 fps vs 60 fps is noticeable.The more intensive the scene (e.g. fighting in assassin's creed odyssey) the more pronounced the diffference is.In regular world exploraton you can see it but it doesn't really have that much impact.
anyway,it doesn't matter if you can see it or not,card A is slower,has no rtx acceleration,runs hotter and louder at the same price as card B = that means you're not getting a product of the same quality.


but only with amd you're getting 16gb vram and pci-e 4.0, up to 69% faster,much wow

boy,for those who mocked the rtx demo,will your face be red now that the most hyped up gpu series since the launch of vega presented you whatever this is

I get why you'd make pci-e 4.0 for x570/ryzen 3 a big thing,the connectivity capabilities of the chipset are just insane.But make your mid-range card all about pci-e 4.0? Hvae they forgotten what actually sells cards completely?

Oh come on. They don't have an RT gimmick to show off, you need to do something to package 2-3 year old performance and make it stick, right.

Nvidia isn't doing a whole lot different since the 1080ti, they just market it better. All GPUs since Pascal are baby steps, let's just call it what it is, and no, I don't consider a 1200 dollar GPU a meaningful step forward, that's way beyond comfort zone for this gaming segment. Think back - Pascal took its sweet time to get replaced, and when it did, the only progress in midrange has been the RTX 2060. The rest might just as well have been gone and nobody would game any different - well, bar that 0.9% that buys the 2080ti.

All things considered I think both companies do a fine job hiding that loooong wait for 7nm. Its just a shame AMD didn't get to do more with Navi than this IMO, they had their sweet time by now as well. They are really both dragging it out because they know there is just a limited gain left on the horizon.

Moving even as little as 100mb takes about 10ms across a PCI-e 3.0 connection, let's say you want 144hz since that's all the rage these days. One frame takes 6.9 ms, but your memory transfer alone for GPGPU purposes takes 10ms , ouch. Now you may not need to move 100mb all the time, but you can easily see how this doesn't quite work very well if you want high framerates, irrespective of how fast the GPU itself is. And let's not even bring in the matter of latency.

This is totally a memory transfer problem and because of that concurrent graphics and GPGPU remains problematic. Take a guess as to why most have given up on GPGPU implementations, it can't all be because of lack of interest, there are technical limitations that prohibit this. Even PhysX implementations usually pertain to small scale effects here and there which can be run on a CPU just fine a lot of the time despite the obvious computational advantage that they have on GPUs.

Interesting, never knew that was such a major limitation for physics.

 

[FordGT90Concept]

Moving even as little as 100mb takes about 10ms across a PCI-e 3.0 connection, let's say you want 144hz since that's all the rage these days. One frame takes 6.9 ms, but your memory transfer alone for GPGPU purposes takes 10ms , ouch. Now you may not need to move 100mb all the time, but you can easily see how this doesn't quite work very well if you want high framerates, irrespective of how fast the GPU itself is. And let's not even bring in the matter of latency.

This is totally a memory transfer problem and because of that concurrent graphics and GPGPU remains problematic. Take a guess as to why most have given up on GPGPU implementations, it can't all be because of lack of interest, there are technical limitations that prohibit this. Even PhysX implementations usually pertain to small scale effects here and there which can be run on a CPU just fine a lot of the time despite the obvious computational advantage that they have on GPUs.

If my math doesn't suck, PCIE 3.0 x16 is 15.75 GB/s which translates to 15.75 MB/ms so you're theoretical example of 100 MB would take 6.35 ms. Only reason why you'd even consider doing that is if you were recording at 144 Hz. In which case, the GPU itself is using its encoding ASIC to cut the bandwidth way down...amount depends on resolution and color depth. The only thing that can get close to saturating that is 4K but...good luck getting 144 fps in the first place. The encoder naturally has high compression rate when duplicate frames are injected so, unless you have a lot of movement and are getting 144+ fps, it's not going to be a problem.

CPUs request very little from GPUs so that upstream used for saving the video stream is mostly unused anyway. CPUs send a lot of data to the GPU.

Again, it's not a bandwidth problem, it's a game design problem. If you have haphazard implementations of physics that are actually more than cosmetic, you're going to get a haphazard game experience from it too. Unless the physics are instrumental to the game design (see Havok and Red Faction: Guerilla) developers always smash the "easy" button.

Interesting, never knew that was such a major limitation for physics.

It's not. Realistic physics just aren't a priority for most developers. Even in cases where it is (Make Sail comes to mind--uses PhysX), they do it on CPU for compatibility sake. There is no 100% hardware and software agnostic physics library. CPU PhysX is attractive because Unreal Engine 4 and Unity both use it and are multi-platform friendly. Any decision to use GPU acceleration for PhysX means the game is broken for all Intel (Make Sail will run on decent integrated graphics), AMD, and ARM-based players.


...this is getting so far off topic.

 

[Vya Domus]

If my math doesn't suck, PCIE 3.0 x16 is 15.75 GB/s which translates to 15.75 MB/ms so you're theoretical example of 100 MB would take 6.35 ms. Only reason why you'd even consider doing that is if you were recording at 144 Hz. In which case, the GPU itself is using its encoding ASIC to cut the bandwidth way down...amount depends on resolution and color depth. The only thing that can get close to saturating that is 4K but...good luck getting 144 fps in the first place. The encoder naturally has high compression rate when duplicate frames are injected so, unless you have a lot of movement and are getting 144+ fps, it's not going to be a problem.

In practice you don't see more than 12-10 GB/s of bandwidth, I am being realistic here. Also I have no idea what recording has to do with any of this. You want compute on the GPU, you need the bandwidth for it, it's a basic fact and when you try to intertwine it with graphics you encounter issues. No matter what you do, you can't avoid this, games need to be interactive systems and data has to be updated with each frame.

You got me curious though, if PCI-e bandwidth is utterly irrelevant for stuff such as GPU accelerated physics how did that benchmark AMD showed worked ? Does it just move worthless data back and forth tie that to framerate and report bogus results ?

 

[FordGT90Concept]

Also I have no idea what recording has to do with any of this.

Because it's the only constant thing that uses a lot of bandwidth that can't be precached.

You want compute on the GPU, you need the bandwidth for it, it's a basic fact and when you try to intertwine it with graphics you encounter issues.

Physics is a high compute, low bandwidth workload as far as the CPU-GPU link is concerned. They were doing heavier physics simulations back on PCIE 1.0 than they're doing now in games.

Find a citation of one developer that said PCIE bandwidth is an issue for their game. Most developers bitch about console bandwidth being so restrictive, not PC, and in those instances, it's because they were doing something wrong (e.g. loading assets into the wrong memory which choked it).

Does it just move worthless data back and forth tie that to framerate and report bogus results ?

Likely. Benchmarks tend to do frivolous things in repeatable ways.

Many people have tested games by blocking PCI lanes and it just doesn't make much difference unless you get under x8 lanes. This has been true for every generation since the first. As graphics cards get faster, so does PCIE. It's always been a non-issue...unless two GPUs are trying to talk to each other.


To be perfectly clear: GPU physics is bad because it robs frames from the game, not because of bandwidth but because compute resources that would be used for graphics are being used for physics. Most players can't tell fake physics from realistic physics but they can tell 30 fps from 15 fps. That's why GPU physics have been relegated to optional cosmetics: turning it off won't break the game.

 

[TheoneandonlyMrK]

In practice you don't see more than 12-10 GB/s of bandwidth, I am being realistic here. Also I have no idea what recording has to do with any of this. You want compute on the GPU, you need the bandwidth for it, it's a basic fact and when you try to intertwine it with graphics you encounter issues. No matter what you do, you can't avoid this, games need to be interactive systems and data has to be updated with each frame.

You got me curious though, if PCI-e bandwidth is utterly irrelevant for stuff such as GPU accelerated physics how did that benchmark AMD showed worked ? Does it just move worthless data back and forth tie that to framerate and report bogus results ?

Imho, both AMD and Intel intend to utilize the GPU and the Cpu coherently together, in order to make ray tracing and other edge cases like AI FASTER, both these use cases can be helped by lower latency, high bandwidth interconnects.
It's easy to get caught up on the way things are, it is important to note where things are going.
Sony's main tech guy said the next generation is about a programmable era? Not a graphical era, despite Rays being in the news.

 

[moproblems99]

Are you making a statement or asking a question for an answer no one outside of AMD could know? Besides that no comment, I'll wait for benches before offering my 2 cents.

Did I leave off the question mark? I thought I saw a slide where they were touting 2070 perf for $499. So I asked. Statements don't usually end in question marks.

 

[Zareek]

Did I leave off the question mark? I thought I saw a slide where they were touting 2070 perf for $499. So I asked. Statements don't usually end in question marks.

Why would they claim claim 2070 performance for $499 when you can already get a 2070 for $499? Maybe 2080 performance for $499 but AMD releasing something that actually has 2080 level performance across the board seems like a pipe dream at this point. It also would make zero sense if they are claiming that Vega VII will still be the top of their stack.

 

[Totally]

Did I leave off the question mark? I thought I saw a slide where they were touting 2070 perf for $499. So I asked. Statements don't usually end in question marks.

I know statements don't end with a question mark. As question it made no sense and there was no way to answer it because you thought you saw something you didn't.

 

[HwGeek]

Is it me or Dr. Lisa su loves to have fun with the new products?
In CES she showed us 9900K vs 8C Ryzen 3000 and in Computex she presented those 3700X/3800X CPU's first that we all expected and then she surprised every one with Ryzen 9 12C.
In similar fashion she showed us the "Look how tiny it is" RX 5700 and showed a demo vs RTX 2070, and then in E3 after presenting the RX 5700 she will surprise us with RX 5900[?] to compete with RTX 2080Ti?

 

[Totally]

absolutely.66 fps vs 60 fps is noticeable.The more intensive the scene (e.g. fighting in assassin's creed odyssey) the more pronounced the diffference is.In regular world exploraton you can see it but it doesn't really have that much impact.

I don't agree with that but don't feel like arguing that point

anyway,it doesn't matter if you can see it or not,card A is slower,has no rtx acceleration,runs hotter and louder at the same price as card B = that means you're not getting a product of the same quality.

I don't how you guys can keep dancing around card A bring faster and has RTX. Card A is faster than card B with RTX off. Card A is slower than Card B with RTX on. So you can't sit there and bring up RTX whenever it is said card A is faster than card B and the response is "well, not by much" because when RTX is on it it is slower.

 

[cucker tarlson]

I don't agree with that but don't feel like arguing that point

of course,cause it's subjective.no point in arguing.

because when RTX is on it it is slower.

totally

 

[moproblems99]

Why would they claim claim 2070 performance for $499 when you can already get a 2070 for $499? Maybe 2080 performance for $499 but AMD releasing something that actually has 2080 level performance across the board seems like a pipe dream at this point. It also would make zero sense if they are claiming that Vega VII will still be the top of their stack.

How about because they already have done this once before? You could get 2080 performance for $699 and they went ahead and launched a gpu with less performance for $699.

I know statements don't end with a question mark. As question it made no sense and there was no way to answer it because you thought you saw something you didn't.

Sounds like this could have been could a good answer: 'No, they didn't say that anywhere. You likely read someone's opinion in an article.' The last half isn't even really needed. Yeah, I think that would have perfectly answered the question.

 

[tfdsaf]

Topic title needs a slight fix

AMDs new RDNA arch has been exclusively renamed for gamers



So far all that console dominance has resulted in zero progress on their GPU side. Just more chips at the same performance and near similar price. And they try to sell you that Strange Brigade performance as if it applies everywhere else while it does not.

Cant say Im feeling the gamer love

OMG the same CRAP all over again. Nvidia has had the SAME architecture for over 10 years now buddy boy. I mean their latest Turing is essentially the same architecture from their GTX 200 days.

The whole old and tiresome and fake and propagandist fake news that AMD's is using the same architecture and somehow Nvidia magically has new architecture every time is the most absurd propaganda I've ever seen on the internet. Fact of the matter is that Turing literally has roots and bottom down is the same architecture as their GTX 200 architecture from over 10 years ago!

IF anything AMD has changed architecture a lot more times and even their GCN architecture has seen major changed with each revision. GCN 1 to GCN 4 is virtually incomparable, even Vega was a massive redesign on GCN 4 which was already massively different from GCN 3.

What's with this fake news propaganda that somehow everything AMD does in the GPU space is GCN and is always the same, while Nvidia the saviors of the universe, the angels of heaven always magically have a new architecture every time brought from the clouds above on unicorns? Complete fake news propaganda! Its worse fake news like the decades old fake news of "AMD drivers are bad".

 

[cucker tarlson]

OMG the same CRAP all over again. Nvidia has had the SAME architecture for over 10 years now buddy boy. I mean their latest Turing is essentially the same architecture from their GTX 200 days.

The whole old and tiresome and fake and propagandist fake news that AMD's is using the same architecture and somehow Nvidia magically has new architecture every time is the most absurd propaganda I've ever seen on the internet. Fact of the matter is that Turing literally has roots and bottom down is the same architecture as their GTX 200 architecture from over 10 years ago!

IF anything AMD has changed architecture a lot more times and even their GCN architecture has seen major changed with each revision. GCN 1 to GCN 4 is virtually incomparable, even Vega was a massive redesign on GCN 4 which was already massively different from GCN 3.

What's with this fake news propaganda that somehow everything AMD does in the GPU space is GCN and is always the same, while Nvidia the saviors of the universe, the angels of heaven always magically have a new architecture every time brought from the clouds above on unicorns? Complete fake news propaganda! Its worse fake news like the decades old fake news of "AMD drivers are bad".

oh dis gon b gud

 

[GoldenX]

nvlink 2.0 on turing can do the same 100gb/s,which helps with gpu-gpu scaling
but cpu-gpu we're not in any danger of running into bottlenecks on pci-e 3.0 and if we do pci-e 4.0 is already here

RAM to VRAM copies still are too slow.

OMG the same CRAP all over again. Nvidia has had the SAME architecture for over 10 years now buddy boy. I mean their latest Turing is essentially the same architecture from their GTX 200 days.

The whole old and tiresome and fake and propagandist fake news that AMD's is using the same architecture and somehow Nvidia magically has new architecture every time is the most absurd propaganda I've ever seen on the internet. Fact of the matter is that Turing literally has roots and bottom down is the same architecture as their GTX 200 architecture from over 10 years ago!

IF anything AMD has changed architecture a lot more times and even their GCN architecture has seen major changed with each revision. GCN 1 to GCN 4 is virtually incomparable, even Vega was a massive redesign on GCN 4 which was already massively different from GCN 3.

What's with this fake news propaganda that somehow everything AMD does in the GPU space is GCN and is always the same, while Nvidia the saviors of the universe, the angels of heaven always magically have a new architecture every time brought from the clouds above on unicorns? Complete fake news propaganda! Its worse fake news like the decades old fake news of "AMD drivers are bad".

If GCN versions are incomparable, then explain the AMD compiler. Backwards compatibility is intact, same as in Nvidia.
"Massive redesigns" and we still have the same crappy performance and power consumption, 7 years later.

 

[cucker tarlson]

RAM to VRAM copies still are too slow.


If GCN versions are incomparable, then explain the AMD compiler. Backwards compatibility is intact, same as in Nvidia.
"Massive redesigns" and we still have the same crappy performance and power consumption, 7 years later.

fact is,all you have to do is look at cases where gcn 1 struggled.polaris and vega get hit in the same way.he's partially right,in saying that there have been improvements.but at its core gcn is still gcn.
nvidia changes incrementally too,but mostly to improve where the prev gen struggled.with turing they got fp32+int32 execution,half precision and finally a proper async support.otherwise,e.g. looking at pascal and maxwell,the architecture did not need any major overhauls.

 

[FordGT90Concept]

Why would they claim claim 2070 performance for $499 when you can already get a 2070 for $499? Maybe 2080 performance for $499 but AMD releasing something that actually has 2080 level performance across the board seems like a pipe dream at this point. It also would make zero sense if they are claiming that Vega VII will still be the top of their stack.

RTX 2070 has ~10.8 billion transistors compared to RX 5700 having ~10.2 billion transistors, PCB costs for Turing are known to be high, and Navi being on 7nm versus Turing on 12nm likely means less wattage so Navi doesn't need as much money spent on the HSF. If AMD can match or slightly exceed RTX 2070 performance at the same price point, AMD is still making more money per unit shipped simply because it's cheaper to produce.

Navi is going to be very competitive in every way against Turing because it doesn't waste transistors on tensor cores and it's a smaller node. The concern I have is what does NVIDIA's 7nm product look like?

 

[NdMk2o1o]

RTX 2070 has ~10.8 billion transistors compared to RX 5700 having ~10.2 billion transistors, PCB costs for Turing are known to be high, and Navi being on 7nm versus Turing on 12nm likely means less wattage so Navi doesn't need as much money spent on the HSF. If AMD can match or slightly exceed RTX 2070 performance at the same price point, AMD is still making more money per unit shipped simply because it's cheaper to produce.

Navi is going to be very competitive in every way against Turing because it doesn't waste transistors on tensor cores and it's a smaller node. The concern I have is what does NVIDIA's 7nm product look like?

Though RVII is 7nm and that didn't set the world alight and depending where you look it fairs anywhere just above a 2070 and in some cases competing with 2080 and everywhere in between, it has been said RVII will remain their flagship (not sure how true or accurate) but if that's the case, what can navi do at 7nm that RVII failed at as it's still questionably loud and hot and uses a lot of power compared to Turing, so I don't see how navi will fair any better. I see it being similar to RVII in that sense in that it might trade blows with the 2070 in some cases but will likely be between that and the 2660 and priced accordingly, which isn't all that bad.

 

[FordGT90Concept]

RTX 2080 = 13.6 billion transistors
Radeon VII = 13.23 billion transistors

Radeon VII in Navi, if the architecture scales well, would have matched or slightly exceeded RTX 2080 in performance. There has to be some serious architectural improvements in Navi to make it more competitive with Turing.

Yeah, Radeon VII is going to be AMD's top product until Arcturus debuts. Thing has 16 GiB HBM2.

AMD wouldn't have matched RTX 2070 pricing if they weren't certain it's a better performance/value proposition because of AMD's lower mindshare. For that reason, I think Radeon VII and RX 5700 are very similar in performance but RX 5700 has half the memory. AMD also doesn't want RX 5700 to cannibalize Radeon VII sales so RX 5700 likely isn't the best Navi can do--it's the best AMD will allow until Arcturus debuts and Radeon VII inventory is cleared. The fact they're not launching an RX 5800 now is proof they're holding something back. I wouldn't be surprised at all if the silicon that would be RX 5800 is what is going into the PlayStation 5.


Eight more days to E3 announcement which should focus on RX 5000 series.

 

[tfdsaf]

Though RVII is 7nm and that didn't set the world alight and depending where you look it fairs anywhere just above a 2070 and in some cases competing with 2080 and everywhere in between, it has been said RVII will remain their flagship (not sure how true or accurate) but if that's the case, what can navi do at 7nm that RVII failed at as it's still questionably loud and hot and uses a lot of power compared to Turing, so I don't see how navi will fair any better. I see it being similar to RVII in that sense in that it might trade blows with the 2070 in some cases but will likely be between that and the 2660 and priced accordingly, which isn't all that bad.

Lisa Su directly addressed this at Computex. Their most expensive GPU is going to remain Radeon 7 for now, literally her words. That leaves room for a RTX 2080 competitor that is cheaper. So again, she worded the answer very carefully. Do I think Navi has a RTX 2080 competitor? Yes. Do I think its going to be released on July 7th? Not sure. I think its more likely they are going to release RX 5070, RX 5060, RX 5060PRO and RX 5050 first.

Later on we are likely to see RX 5080 and RX 5040 to complete their lineup. Their RX 5070 is also really small, and if it performs that good at that size, it could be that AMD has a bigger die that can compete with RX 2080 and RX 2080ti.

According to AMD its 25% faster than Vega and 50% more power efficient. That would mean at 4096 cores it performs 25% faster and has half the power consumption, which would put it at around 150W for an RTX 2070 type performance.

 

[NdMk2o1o]

Lisa Su directly addressed this at Computex. Their most expensive GPU is going to remain Radeon 7 for now, literally her words. That leaves room for a RTX 2080 competitor that is cheaper. So again, she worded the answer very carefully. Do I think Navi has a RTX 2080 competitor? Yes. Do I think its going to be released on July 7th? Not sure. I think its more likely they are going to release RX 5070, RX 5060, RX 5060PRO and RX 5050 first.

Later on we are likely to see RX 5080 and RX 5040 to complete their lineup. Their RX 5070 is also really small, and if it performs that good at that size, it could be that AMD has a bigger die that can compete with RX 2080 and RX 2080ti.

According to AMD its 25% faster than Vega and 50% more power efficient. That would mean at 4096 cores it performs 25% faster and has half the power consumption, which would put it at around 150W for an RTX 2070 type performance.

You're reading way in between the lines here, the most expensive GPU is going to remain the RVII but yet you think they will release a navi product that's cheaper and faster for less money? 25% faster than vega isn't quite good enough unless you're talking IPC or per clock

 

[FordGT90Concept]

According to AMD its 25% faster than Vega and 50% more power efficient. That would mean at 4096 cores it performs 25% faster and has half the power consumption, which would put it at around 150W for an RTX 2070 type performance.

Those numbers work in terms of transistors. If Radeon VII were cut down to 12.75 billion transistors, it would be about where RX 5700 is.

 

[biffzinker]

Oh so you remember well R600? Navi don't have 512-bit memory bus width, nor does it have huge die size. So I don't see what Navi have something in common with 2900XT. Why don't we wait before it comes out

Forgot the ring bus on die.