Nvidia Blackwell Architecture Speculation

[Rys]

I know product launches are exciting for all concerned and the hot takes almost write themselves, but the amount of thread cleanup and post deletion at the moment to remove incredibly low effort posts, especially those that violate the spirit of the forum guidelines, is getting out of hand. It's not just in this thread either.

The guidelines are clear: "Zero...overt negativity, shallow dismissals or low-effort shitposting of any kind. Signal must be maintained or boosted at all times."

In this particular case, "I hope someone can underclock the memory on the 5080 and compare it vs the 4080super. The IPC is looking like almost 0%.", would have worked just fine. You didn't need the rest.

 

[Kaotik]

oh man, a possible future RTX 5080 Ti is going to be the 4080 Super Super for real-I swear to you version

RTX 5080 Ti/Super/whatever can only happen if NVIDIA has tons of salvage GB202 dies that won't be fed to pro cards. RTX 5080 is already full GB203 and just updating memory to 3 Gb chips won't cut it

 

[Boss]

I know product launches are exciting for all concerned and the hot takes almost write themselves, but the amount of thread cleanup and post deletion at the moment to remove incredibly low effort posts, especially those that violate the spirit of the forum guidelines, is getting out of hand. It's not just in this thread either.

The guidelines are clear: "Zero...overt negativity, shallow dismissals or low-effort shitposting of any kind. Signal must be maintained or boosted at all times."

In this particular case, "I hope someone can underclock the memory on the 5080 and compare it vs the 4080super. The IPC is looking like almost 0%.", would have worked just fine. You didn't need the rest.

That's fair, I missed that part..... However, I wasn't shit posting at all, I was just being honest. If you go back a decade plus, we haven't see uplifts this bad. To release a generation with almost 0 ipc improvement? I can't recall the last time Nvidia released such a product that didn't come along with a huge price cut. I don't even want to delve into the price discussions as it's a massive distraction. Personally speaking, for me, there's no justifiable reason for a new generation if you can't deliver a meaningful increase in IPC. The only exception to this may be node shrinks. As node shrinks get more difficult, more emphasis will be placed on the architecture to seek out performance gains. If this is the type of performance we can expect from Nvidia, then I'm not too hopeful at all. How many real node shrinks do we have left? 2-4 and then its cooked? Each of those nodes will be extremely expensive.

 

[Albuquerque]

What about instructions per watt? Or howabout performance per dollar?

LevelOneTechs or DigitalFoundry (I need to go back and figure out which one) review did a frames per dollar calculation across the last three gens of NVIDIA x080- and x090-series and AMD x900 and x800 series. The 5800 provides, by a notable margin, the very best frames per dollar of every prior card in the list, and we're not talking about FG and DLSS here. They didn't even try to bring in the inflation devaluation of prior cards, because then it would've been even more terrible for all prior gens.

In terms of actual, literal bang for your buck, it's a win. That doesn't mean everyone has to like it, but to pretend it's a nothingburger misses a few key and IMO pertinent points.

Edit: It was the DF review. Here's their Youtube review timestamped to their discussion of the dollars per frame metric:

 

[homerdog]

RTX 5080 Ti/Super/whatever can only happen if NVIDIA has tons of salvage GB202 dies that won't be fed to pro cards. RTX 5080 is already full GB203 and just updating memory to 3 Gb chips won't cut it

Sticking with GB203, they could call it 5080 24GB and price it the same as the original 5080 which could then be phased out. Kind of like they did with the 3060 8GB but in reverse

 

[Kaotik]

Sticking with GB203, they could call it 5080 24GB and price it the same as the original 5080 which could then be phased out. Kind of like they did with the 3060 8GB but in reverse

Yes, but it wouldn't be "Super" or "Ti", just memory upgrade

 

[Kaotik]

What about instructions per watt? Or howabout performance per dollar?

LevelOneTechs or DigitalFoundry (I need to go back and figure out which one) review did a frames per dollar calculation across the last three gens of NVIDIA x080- and x090-series and AMD x900 and x800 series. The 5800 provides, by a notable margin, the very best frames per dollar of every prior card in the list, and we're not talking about FG and DLSS here. They didn't even try to bring in the inflation devaluation of prior cards, because then it would've been even more terrible for all prior gens.

In terms of actual, literal bang for your buck, it's a win. That doesn't mean everyone has to like it, but to pretend it's a nothingburger misses a few key and IMO pertinent points.

Edit: It was the DF review. Here's their Youtube review timestamped to their discussion of the dollars per frame metric:

Of high end cards perhaps, but still nothing to write home about

 

[RobertR1]

It’s a poor card because if someone wanted to skip the 4000 gen and jump to the 5000 gen at the 1k mark, they could have been there a year or two ago instead of waiting for minimal gains.

Btw I have a 4090 so this is just an observation. I don’t see the enthusiasm for this card and past initial window, let’s see how sales play out.

Perhaps, they’re hoping 3080 owners are willing to take a price bump? The 5000 series is just not aimed at 4000 owners. Perhaps that’s the perspective.

All the above is assuming msrp btw.

 

[Albuquerque]

It's always been true that the lowest-end cards have the highest "value" in terms of dollars per frame. The unfortunate part is, at some point, you actually need more frames for the games to be playable. A Geo Metro was always a highly efficient car, until you needed to comfortably carry five people and their luggage. Then it didn't matter how efficient it was, because you couldn't pull out of the driveway

In terms of the high end, it's more "affordable" now than it's been in almost a decade. I get that it seems counter-intuitive, I know a few people will be waiting to pounce on their keyboard with BUT BUT BUT BUT THE GENERATIONAL UPLIFT, but that has nothing to do with how much money you're plunking down to get framerate out. This is what DF was getting at with their statement; the 5080 is a better value (again, dollars per frame generated) than every other top-end card produced in the last three generations.

That also doesn't mean it's time for anyone to upgrade.

 

[arandomguy]

I just want to examine something which is the concept of architectural improvements. I think there might be expecations issues here on what can actually be achieved in terms of better/more advanced design efficiency over just more transistors spent and node improvements as they were typically done in conjunction.

Take Kepler to Maxwell. Same node and people remember that as a massive architectural improvement. But how was that actually achieved? GM204 has about 33% more transistors than GK104 at about 50% better performance. If you were to normalize for transistor count than the performance improvement was roughly 10%. Now in terms of Blackwell, GB203 and AD103 have roughly the same transistor count on the same node. Well it's not quite apples to apples as GB203 does benefit from a newer memory technology but it also carries some more feature set improvements as well.

Setting aside Nvidia what if we look at Navi 33 vs Navi 23? Navi 33 has about 20% more transistors (and also on a slightly better node iteration). The 7600XT isn't even 20% faster than either the 6600XT or 6650XT Navi 23 configurations.

Ultimately raw performance improvements are going to cost transistors and if that transistor cost is not going down at the rate of the past well we can either accept declining user experience changes or try to be creative elsewhere in terms of impacting the user experience.

Let's even just set aside the PC GPU hardware wars for the moment and look elsewhere. Is the console industry also "in" on everything as well? The PS5 to Pro did not deliver the same raw improvements much less at the same cost levels relative the PS4 generations. Microsoft's stance seems to be based on what they know they aren't even going to bother. And I think we are going to see with Nintendo and the Switch 2 it's going to end up on the conservative side as well of expecations in terms of raw performance.

 

[Boss]

What about instructions per watt? Or howabout performance per dollar?

LevelOneTechs or DigitalFoundry (I need to go back and figure out which one) review did a frames per dollar calculation across the last three gens of NVIDIA x080- and x090-series and AMD x900 and x800 series. The 5800 provides, by a notable margin, the very best frames per dollar of every prior card in the list, and we're not talking about FG and DLSS here. They didn't even try to bring in the inflation devaluation of prior cards, because then it would've been even more terrible for all prior gens.

In terms of actual, literal bang for your buck, it's a win. That doesn't mean everyone has to like it, but to pretend it's a nothingburger misses a few key and IMO pertinent points.

In terms of performance per dollar, it's actually worse than the 4080 super in my region. People are trying to ignore the 4080 super as if it doesn't exist. It exists and it's the only card the 5080 should be compared to as the 4080 got discontinued. In terms of performance per watt, most people don't care about that at all and the difference between the 5080 and 4080 super is marginal. It's really in the same psu class. People are out here praising the 5090 for blasting 575w of power for a 33% increase in performance. Coincidentally, there's a strong correlation between the power/cuda core increase in the 5090 vs 4090 and performance increase.

Edit: It was the DF review. Here's their Youtube review timestamped to their discussion of the dollars per frame metric:

3 things... Firstly, at least where I live, the 5080 will not be sold for msrp anytime soon so that comparison is irrelevant. Secondly, the idea of comparing the msrp of the new product vs the msrp of the old product makes absolutely no sense. Using the xtx as example, it can be purchased at $829.99 usd right now so the choice the consumer faces is can I buy an xtx at 829.99 or a 5080 at $999. The 5080 provides worse value per frame than the xtx in today's prices. Finally, every gen at minimum offers better performance per dollar when compared to the msrp. If you can't even do that, you shouldn't even bother just like Intel shouldn't have bothered with the core ultra.

Edit: What's even more disappointing about this architecture is the lack of improvement in performance per RT core. In many ways, RT performance improvement is almost worse than the raster improvement.

 

[Kaotik]

It's always been true that the lowest-end cards have the highest "value" in terms of dollars per frame. The unfortunate part is, at some point, you actually need more frames for the games to be playable. A Geo Metro was always a highly efficient car, until you needed to comfortably carry five people and their luggage. Then it didn't matter how efficient it was, because you couldn't pull out of the driveway

On the rasterization side 7900 XTX is the next highest after 5080 and offers better perf/dollar, though

 

[trinibwoy]

It’s a poor card because if someone wanted to skip the 4000 gen and jump to the 5000 gen at the 1k mark, they could have been there a year or two ago instead of waiting for minimal gains.

This is exactly where I’m at. Had many opportunities to purchase a 4090 but didn’t pay attention as I’m on a 4 year cycle to maximize impact of each upgrade. I guess it’s my fault for expecting more from the same process node. If you told me Blackwell would essentially stand still on both node and architecture I would’ve lost that bet.

Regrets aside anyone who just wants 5080 performance today shouldn’t be too upset.

 

[Rys]

That's fair, I missed that part..... However, I wasn't shit posting at all, I was just being honest.

You weren't shitposting, but it was a shallow dismissal and overtly negative. Hopefully you can continue to explain what you mean by "IPC" in the context of the Blackwell architecture and why it's the main thing you're hanging your hat on to make your argument.

 

[arandomguy]

This is exactly where I’m at. Had many opportunities to purchase a 4090 but didn’t pay attention as I’m on a 4 year cycle to maximize impact of each upgrade. I guess it’s my fault for expecting more from the same process node. If you told me Blackwell would essentially stand still on both node and architecture I would’ve lost that bet.

Regrets aside anyone who just wants 5080 performance today shouldn’t be too upset.

But would you been happy with higher prices but more performance on the same node?

I went into this on a previous post but I think people have the wrong historical impression of improvements on the same node. If we look at the previous 2 times in Turing and Maxwell in which Nvidia stayed on the same node they also increased transistor counts significantly in both cases. While with Blackwell the only die that looks like it will increase tranistor count to any meaningful degree is GB202. The rest of the stack seems to be same and being able to extract 10% (and some other improvements) does seem fairly in line with what happended with Kepler->Maxwell.

It's worth mentioning that better performance per transistor spent isn't a given either. There are examples of regressions.

 

[DavidGraham]

This gen has disappointed in terms of performance, it's a straight up power hike to gain performance. The only thing we gained are new features, which can be summarized as follows:

DLSS4 TNN: works on all RTX hardware, gives an image quality and performance boost, because the new performance preset is now comparable to the old quality preset or to native, works well on RTX 40 GPUs and best on RTX 50.

Mega Geometry: again works on all RTX hardware, this will provide a big fps boost in path traced and ray traced scenarios. RTX 50 GPUs may provide bigger boost than older GPUs, and may have better VRAM utilization while doing so.

RTX Hair: new hardware primitive support exclusive to RTX 50, probably means RTX 50 GPUs will have increased performance with this technique vs older gen, but by how much? It's up in the air.

Neural Rendering: again works on all RTX cards, since they will be interfacing through DirectX Cooperative Vectors, which will access the tensor cores, RTX 50 GPUs are designed specifically to accelerate this workload, so they will have better performance than older gen, but by how much? again it's up in the air.

So, apart from DLSS Multi FG, all the new features are available on all RTX generations, RTX 50 will just have slightly better performance in DLSS4, maybe moderately better performance with Mega Geometry and RTX Hair, and maybe a very large performance lead with Neural Rendering, we simply don't know yet as NVIDIA didn't provide us with numbers or benchmarks for these features.

 

[Boss]

You weren't shitposting, but it was a shallow dismissal and overtly negative. Hopefully you can continue to explain what you mean by "IPC" in the context of the Blackwell architecture and why it's the main thing you're hanging your hat on to make your argument.

We all know what IPC is in the cpu space and I use the term loosely with GPUs but, when you control all other variables as much as possible but the SMs, how much performance improvement are we really seeing with the architecture. The 5080 vs 4080 super are great cards to make this comparison. The SMs counts are relatively similar and you can limit power, clocks and bandwidth(under clock memory). I expect to find a near 0 improvement between both GPUs in gaming workloads. In fact, there are games right now where the 5080 performs worse than the 4080 super such as delta force...

 

[Dangerman]

This gen has disappointed in terms of performance, it's a straight up power hike to gain performance. The only thing we gained are new features, which can be summarized as follows:

DLSS4 TNN: works on all RTX hardware, gives an image quality and performance boost, because the new performance preset is now comparable to the old quality preset or to native, works well on RTX 40 GPUs and best on RTX 50.

Mega Geometry: again works on all RTX hardware, this will provide a big fps boost in path traced and ray traced scenarios. RTX 50 GPUs may provide bigger boost than older GPUs, and may have better VRAM utilization while doing so.

RTX Hair: new hardware primitive support exclusive to RTX 50, probably means RTX 50 GPUs will have increased performance with this technique vs older gen, but by how much? It's up in the air.

Neural Rendering: again works on all RTX cards, since they will be interfacing through DirectX Cooperative Vectors, which will access the tensor cores, RTX 50 GPUs are designed specifically to accelerate this workload, so they will have better performance than older gen, but by how much? again it's up in the air.

So, apart from DLSS Multi FG, all the new features are available on all RTX generations, RTX 50 will just have slightly better performance in DLSS4, maybe moderately better performance with Mega Geometry and RTX Hair, and maybe a very large performance lead with Neural Rendering, we simply don't know yet as NVIDIA didn't provide us with numbers or benchmarks for these features.

Yeah, maybe RTX 50 will take advantage of future games but as it stands it doesn't even feel like Turing which was more impressive than this. Just wonder despite bragging about new architecture we don't see the gains of Turing or Blackwell levels.

 

[arandomguy]

Yeah, maybe RTX 50 will take advantage of future games but as it stands it doesn't even feel like Turing which was more impressive than this. Just wonder despite bragging about new architecture we don't see the gains of Turing or Blackwell levels.

Pascal to Turing was only a slightly improved node but more importantly Nvidia basically spent 50% more transistors on each comparable chip to acheive the gains and features over Pascal.

For example TU106 (RTX 2070) has 50% more transistors than GP104 (GTX 1080). RTX 2070 is about 20% faster over GTX 1080 on aggregate according to TPUs launch review. GB203 (RTX 5080) from AD103 (RTX 4080 Super) has roughly the same amount of tranistors, it's slightly over 10% faster going by TPU on aggregate.

You can take that from what you will, which will ultimately depend on what you believe in terms of the costs per transistor of modern chips.

 

[Boss]

This gen has disappointed in terms of performance, it's a straight up power hike to gain performance. The only thing we gained are new features, which can be summarized as follows:

DLSS4 TNN: works on all RTX hardware, gives an image quality and performance boost, because the new performance preset is now comparable to the old quality preset or to native, works well on RTX 40 GPUs and best on RTX 50.

While improved, it still has noticeable artifacts, ghosting, etc. I tested it in cyberpunk and it took me less than 10 seconds to spot some issues.

Mega Geometry: again works on all RTX hardware, this will provide a big fps boost in path traced and ray traced scenarios. RTX 50 GPUs may provide bigger boost than older GPUs, and may have better VRAM utilization while doing so.


RTX Hair: new hardware primitive support exclusive to RTX 50, probably means RTX 50 GPUs will have increased performance with this technique vs older gen, but by how much? It's up in the air.

By the time this is relevant and has achieved mass adoption, RTX 7000 series or 8000 series will be out.

Neural Rendering: again works on all RTX cards, since they will be interfacing through DirectX Cooperative Vectors, which will access the tensor cores, RTX 50 GPUs are designed specifically to accelerate this workload, so they will have better performance than older gen, but by how much? again it's up in the air.

DirectX means long lead times/uptake. Until all GPU vendors can support this, it's going nowhere fast.

So, apart from DLSS Multi FG, all the new features are available on all RTX generations, RTX 50 will just have slightly better performance in DLSS4, maybe moderately better performance with Mega Geometry and RTX Hair, and maybe a very large performance lead with Neural Rendering, we simply don't know yet as NVIDIA didn't provide us with numbers or benchmarks for these features.

It's a very disappointing generation and I say this as someone who was initially hyped for to potentially upgrade to the 5090. After I saw the combination of $2000usd, 575 Watts and 33% performance improvement over the 4090, I was out. I'll just hold on to my 4080 super and wait till the next node shrink.