HalfLife2 RTX Demo Is out!

[oxrufiioxo]

I believe Nvidia when they say they've done XYZ to the RT cores, but where are the benefits? It's like AMD with the chiplet design in RDNA 3. It's highly advanced tech, but did it make the cards faster? Or cheaper? No. It didn't even go into the margins, as we all see they lost some massive cash on Radeon last year. A technology is only as good as the implementation of it. AMD has proved that with FX, then chiplets on RDNA 3, and now Nvidia is proving it with Blackwell, it seems.

I also don't think Blackwell is "forward thinking" in any way. Looking at the diagrams, it's still the same architecture as Turing, only with the INT/FP roles of the shader cores mixed up to be more versatile - which again, shows no gains in real life.


What gain in RT performance? Do you mean gain in performance in general? RT performance alone hasn't improved much since Turing (see my explanation above).

If you just separate how many RT cores they've added and the actual perfomance uplift in RT heavy gaming the uplift is always higher is all I was pointing out if RT perfomance was the same you'd see a regression becuase the amount of RT cores isn't going up significantly.

It's likely why you see much larger gains in RT with the 4090 vs 3090 than raster only.

but even path traced games have a ton of rasterization still so it's impossible to actually know how much better each RT core is.

I'm talking 20-30-40 series 5000 series who knows lol.

It's not 18%. It's less than 10%. If you compared with equal rasterization & shader count & r.o.p's Nvidia has only gained at most 6% increase in RT efficiency every generation. The 3080 ti has much higher rasterization than a 2080 ti. Every one of the comparisons you made here increased rasterization, it's very hard to compare Nvidia's own generations to each other. When nvidia keep changing parts for rasterization while only claiming their RT is better.

You definitely can't isolated them but I've noticed they haven't upped the RT count much each generation likely due to cost they probably want to keep the same level of RT to raster perfomance so people are more reliant on their features.

 

[dgianstefani]

I know it sounds mad, but I don't think Nvidia has done much with the RT cores since Turing. If the 2080 lost let's say 48% performance in X game with RT on vs off, then the 3080 did, too, and the 4080 and the 5080. The only reason we see more FPS is because we see more FPS in general due to having more cores running at higher frequencies, and not because those cores have improved. But the relation between raster and RT performance has stayed the same. This is not the "RT is the future" promise that we've been fed for the last 6 years.

Sure, AMD made a huge leap with RDNA 4 on the RT front because they had more ground to cover, I give you that.


And? I don't care about the implementation, I care about what I see on screen.


Exactly.


That wouldn't be bad. Just like dedicated AI chips wouldn't be bad for those who only need that, while leaving gaming GPUs alone.

Your assumption is made on the idea that RT is just a % workload on top of normal rendering. It isn't. RT is it's own thing, and you cannot just assume normal rendering -50% FPS for example as a rule, if the RT hasn't been calculated for a frame, the frame waits, regardless of whether traditional rendering was completed or not. This means the traditional rendering performance is bottlenecked by RT calculations, and these don't scale simply or to the exact amount % classic raster improvement from the previous gen.

For RT/PT 120 FPS, you need twice or more the power of RT cores, plus memory/cache/CPU etc compared to RT 60 FPS.

Say you have 100 FPS with Gen 1, RT off, 50 FPS with it on.

Gen 2 has 300 FPS with RT off, 100 FPS with it on. You've seen a 3x perf improvement in normal rendering, but only 2x in RT, this means the generational cost of RT went from half FPS to 1/3 FPS, which looks bad right? Yeah the RT cores got worse right? No, you've still doubled the RT/PT FPS.

This is why charts comparing RT on/off performance cost across the generations are misleading. No, the Blackwell RT cores aren't worse than Ada, or not improved, simply because it's -35% perf instead of -30%, they're a completely separate development and can't be compared as a % against classic rendering. If you want a comparison, compare RT on with Ada, and RT on with Blackwell.

Then you end up with this.

Will be nice when classic trick/hack/approximation rendering goes extinct, much easier to compare generations when you aren't limited by an unknown quantity of what percentage the game is bottlenecked by one or the other. It will just all be physically based rendering.

 

[AusWolf]

I'm not adding a quote because I'm covering multiple points here.

So, chip 1 has 1000 shader cores, 500 RT cores, and does 100 FPS in raster and 50 in RT.

Chip 2 released two years later under a new name has 2000 shader cores, 1000 RT cores, and does 200 FPS in raster and 100 in RT.

Has there been an improvement in the RT cores? I would assume there hasn't. It's the same thing just more of it. This is how I see Turing, Ampere, Ada and Blackwell. More cores, higher frequencies, but very similar hardware.

Your assumption is made on the idea that RT is just a % workload on top of normal rendering. It isn't. RT is it's own thing, and you cannot just assume normal rendering -50% FPS for example as a rule, if the RT hasn't been calculated for a frame, the frame waits, regardless of whether traditional rendering was completed or not. This means the traditional rendering performance is bottlenecked by RT calculations, and these don't scale simply or to the exact amount % classic raster improvement from the previous gen.

For RT/PT 120 FPS, you need twice or more the power of RT cores, plus memory/cache/CPU etc compared to RT 60 FPS.

Say you have 100 FPS with Gen 1, RT off, 50 FPS with it on.

Gen 2 has 300 FPS with RT off, 100 FPS with it on. You've seen a 3x perf improvement in normal rendering, but only 2x in RT, this means the generational cost of RT went from half FPS to 1/3 FPS, which looks bad right? Yeah the RT cores got worse right? No, you've still doubled the RT/PT FPS.

This is why charts comparing RT on/off performance cost across the generations are misleading. No, the Blackwell RT cores aren't worse than Ada, or not improved, simply because it's -35% perf instead of -30%, they're a completely separate development and can't be compared as a % against classic rendering. If you want a comparison, compare RT on with Ada, and RT on with Blackwell.

Then you end up with this.

View attachment 390646

That's not the diagram you're looking for. The one you're looking for is about individual games, where W1zz includes a performance loss percentage in brackets when you turn RT on. That number hasn't changed much since Turing.

Edit: The shader vs RT hardware ratio hasn't changed in AMD, either, but their RT performance loss has gone down with RDNA 4.

 

[oxrufiioxo]

That's not the diagram you're looking for. The one you're looking for is about individual games, where W1zz includes a performance loss percentage in brackets when you turn RT on. That number hasn't changed much since Turing.

That likely has more to do with Nvidias performance targets than how much the RT cores are improved or not improved and until we have a benchmark that only stresses the RT cores no one except Nvidia can know how much better they actually are.

 

[AusWolf]

That likely has more to do with Nvidias performance targets than how much the RT cores are improved or not improved and untill we have a benchmark that only stresses the RT cores no one except Nvidia can know how much better they actually are.

Oh we do know. See my edit on my post.

How much performance you lose when you enable RT is indicative of how improved your RT hardware is. Also, the ratio of extra raster+RT cores vs performance gained. If it's 1:1, there has been no advancement.

 

[dgianstefani]

That's not the diagram you're looking for. The one you're looking for is about individual games, where W1zz includes a performance loss percentage in brackets when you turn RT on. That number hasn't changed much since Turing.

What this means is that classic render hardware and RT hardware have improved at a similar rate, not that RT hardware has stagnated. This is the point, and it is the diagram I'm looking for.

How much performance you lose when you enable RT is indicative of how improved your RT hardware is. Also, the ratio of extra RT cores vs performance gained. If it's 1:1, there has been no advancement.

No. You misunderstand how these things work.

 

[AusWolf]

What this means is that classic render hardware and RT hardware have improved at a similar rate, not that RT hardware has stagnated. This is the point, and it is the diagram I'm looking for.

But if it's in 1:1 ratio with the extra number of cores, then where's the advancement?

No.

Yes.

 

[dgianstefani]

But if it's in 1:1 ratio with the extra number of cores, then where's the advancement?

You really should know that performance doesn't scale with hardware 1:1. This is why the 5090 is not twice as fast as the 5080. The fact that performance has scaled 1:1 with cores means that the usual amount lost to scaling issues was made up with architectural improvements.

Yes.

I can't really be bothered to explain it again, but I suggest you reread what I wrote and try again to understand it.

 

[AusWolf]

You really should know that performance doesn't scale with hardware 1:1. This is why the 5090 is not twice as fast as the 5080.

You're missing my point.

5080 vs 4080 Super. Similar number of cores and frequency = similar performance. Even in RT. No advancement.

This is the first time stagnation shows, because this is the first time we're getting the same number of cores as last gen.

 

[wolf]

I also don't think Blackwell is "forward thinking" in any way. Looking at the diagrams, it's still the same architecture as Turing, only with the INT/FP roles of the shader cores mixed up to be more versatile - which again, shows no gains in real life.

Yes at a high level the architecture itself is exceedingly similar, but the block diagram alone doesn't tell the entire story though, and parts that are represented the same on a block diagram can and have had improvements across subsequent generations. TPU covered it pretty well here. To just look at the block diagram and compare it to Turing isn't considering all available information, and it absolutely is possible that future workloads can and will take advantage of the differences.

 

[dgianstefani]

Yes at a high level the architecture itself is exceedingly similar, but the block diagram alone doesn't tell the entire story though, and parts that are represented the same on a block diagram can and have had improvements across subsequent generations. TPU covered it pretty well here. To just look at the block diagram and compare it to Turing isn't considering all available information, and it absolutely is possible that future workloads can and will take advantage of the differences.

Neural Shaders, Mega Geometry and support for other currently unimplemented technologies would be a prime example of this.

 

[AusWolf]

Yes at a high level the architecture itself is exceedingly similar, but the block diagram alone doesn't tell the entire story though, and parts that are represented the same on a block diagram can and have had improvements across subsequent generations. TPU covered it pretty well here. To just look at the block diagram and compare it to Turing isn't considering all available information, and it absolutely is possible that future workloads can and will take advantage of the differences.

Sure, but if I can't see those differences in action, then it's like they weren't even there. This is why people didn't care about AMD FX, this is why not many cared about RDNA 3, but now you're telling me that with Blackwell, it's absolutely fine?

Neural Shaders and support for other currently unimplemented technologies would be a prime example of this.

We'll see. I stopped believing in the future with AMD FX.

By the time neutral rendering becomes a thing, Blackwell may very well be long obsolete.

 

[dgianstefani]

Sure, but if I can't see those differences in action, then it's like they weren't even there. This is why people didn't care about AMD FX, this is why not many cared about RDNA 3, but now you're telling me that with Blackwell, it's absolutely fine?

The difference is that NVIDIA has both a fantastic education investment in universities and developer support, combined with a sizeable enough marketshare that advancements they make that require developer input matter, whereas when AMD, with ~5% marketshare does something differently to tradition, not many bother.

 

[AusWolf]

The difference is that NVIDIA has both a fantastic education investment in universities and developer support, combined with a sizeable enough marketshare that advancements they make that require developer input matter, whereas when AMD, with ~5% marketshare does something differently to tradition, not many bother.

What 5%? You're also forgetting consoles.

 

[dgianstefani]

What 5%? You're also forgetting consoles.

The most popular console in the world, the Nintendo Switch, uses NVIDIA hardware, and so does its replacement. The Switch sold more than the Xbox and PS5 combined.

 

[wolf]

Sure, but if I can't see those differences in action, then it's like they weren't even there.

I already said it's disappointing, like RDNA3 chiplets in your example where there was no realised benefit and different workloads won't change that. And unlike your example, Blackwell architecture has capability within that can be leveraged, it just isn't yet. I'm not trying to excuse it so I'm confused as to your interpretation here, I'm just stating how it works and how those examples differ.

but now you're telling me that with Blackwell, it's absolutely fine?

I don't recall telling you that at all. I recall saying it's a more forward looking architecture where the forward looking features aren't being made use of yet.

You actually put it best earlier in this very thread - "Not what I said. Don't distort my point, please."

 

[dgianstefani]

Lets stick to the topic, HL2 RTX.

 

[AusWolf]

The most popular console in the world, the Nintendo Switch, uses NVIDIA hardware, and so does its replacement. The Switch sold more than the Xbox and PS5 combined.

Look, I don't want to start another Nvidia vs AMD flame war. Do you?

I said what I see and think and that's that.

I already said it's disappointing, like RDNA3 chiplets in your example where there was no realised benefit and different workloads won't change that. And unlike your example, Blackwell architecture has capability within that can be leveraged, it just isn't yet. I'm not trying to excuse it so I'm confused as to your interpretation here, I'm just stating how it works and how those examples differ.

AMD proved that the future is pointless with FX. By the time applications started to really stretch its multitasking capabilities, time has long gone past the poor thing (which was really forward thinking at the time of its release).

I don't recall telling you that at all. I recall saying it's a more forward looking architecture where the forward looking features aren't being made use of yet.

Fair enough. See above. Blackwell may long be obsolete by the time we see any of that "future" coming towards us.

Lets stick to the topic, HL2 RTX.

Exactly, thank you.

 

[phanbuey]

i'm a little sad at the performance to be honest... i was hoping for something like quake 2 or portal rtx where you could tweak the settings a bit and get it to even run pretty well on a 2080 (back then).

HL2 was a very performant game - i was able to even run it really well on my Radeon 9550 (w/ tpu pencil mod).

It doesn't bode well for morrowind / other rtx remakes.

 

[oxrufiioxo]

i'm a little sad at the performance to be honest... i was hoping for something like quake 2 or portal rtx where you could tweak the settings a bit and get it to even run pretty well on a 2080 (back then).

HL2 was a very performant game - i was able to even run it really well on my Radeon 9550 (w/ tpu pencil mod).

It doesn't bode well for morrowind / other rtx remakes.

Haven't tried it on my 3080ti but given I'd only classify the 4090 at 1440p uw as adequate it probably is pretty rough.

It's fine on a 4070 at 1080p though.

 

[lexluthermiester]

i'm a little sad at the performance to be honest..

Yeah, it is a bit on the heavy side. Remember though this is a beta demo. It will be refined.

i was hoping for something like quake 2

That would have been nice. If you look closely you will see they've done a lot more than just add ray tracing and bump-mapping into the HL2 engine.

 

[Visible Noise]

Looks like crap, runs like crap, waste of time. NEXT!

I must be getting senile, because I’m surprised I didn’t already have you on ignore. Oh well, better late than never as they say.

 

[Apocalypsee]

Funny how back then Half Life 2 run awful with GeForce FX in DirectX 9 mode. How many people actually remembers it? Despite its technical advancement in graphics (then) it still offers DirectX 8 mode AND DirectX 7 for people with older cards to enjoy the game.

DirectX 9 Performance Impact - Half Life 2 GPU Roundup Part 2 - Mainstream DX8/DX9 Battle

 

[wolf]

Funny how back then Half Life 2 run awful with GeForce FX in DirectX 9 mode. How many people actually remembers it? Despite its technical advancement in graphics (then) it still offers DirectX mode AND DirectX 7 for people with older cards to enjoy the game.

DirectX 9 Performance Impact - Half Life 2 GPU Roundup Part 2 - Mainstream DX8/DX9 Battle

Your post got a Love from me because it had two ingredients I love, links to Anandtech (goodnight sweet prince ) and reference to the FX series. Such an awful series that has such a special place in my heart as one of the runts of the litter so to speak.

 

[AusWolf]

Your post got a Love from me because it had two ingredients I love, links to Anandtech (goodnight sweet prince ) and reference to the FX series. Such an awful series that has such a special place in my heart as one of the runts of the litter so to speak.

My brother had an FX 5200. Oh. My. God! That card was really something else.

Edit: Now that I think about it, something seems to be haunting the FX name. Also the 5000 series model number (only in GPUs though)