GPU IPC Showdown: NVIDIA Blackwell vs Ada Lovelace; AMD RDNA 4 vs RDNA 3

[LittleBro]

Nvidia hits 92% gpu market share and AMD for some strange reason are overpricing it's RX 9060 XT series gpus value just isn't there as for example RX 9070 XT. AMD hits his lowest point ever in history! Let's wait until 95% maybe then AMD wakes up. AMD are dumb to miss such an opportunity because RTX 50 Series are one of the worst gpu generations ever released by nvidia.

This is strange. I'd like to know what is the statistical sample they are doing this research on.
What happened in Q4/2024 that AMD GPUs market share surged from 10% to 17% and Nvidia lost 8%?

Lot of people around me decided to go with team red this year. Sales of RX 9070 (XT) vastly outperformed sales of RTX 5080 at Mindfactory in March 2025.
I'd expect AMD to gain 2-3% market share, not lose 2%. Well, the big problem was unobtainium of RX 9070 (XT) at MSRP prices, but so were RTX 5090-80-70Ti.

 

[igormp]

They do not have the A.I. chip demand that nGreedia has, so production allocation away from high profit SKU's is not such a big deal to AMD.

Even thought Instinct doesn't have as much demand as Nvidia's lineup, there's still quite some demand for it, to the point even their partner developers have a hard time acquiring one of those.
And something that's even more relevant is that AMD has a shit ton of demand for their Zen dies, which go into Epyc and Ryzen, which are far smaller than a GPU die and sell way more. Taking fab allocation out of those to make gamer GPUs would be moving away from high profit/high volume SKUs.

 

[Plexiglas]

Even thought instant doesn't have as much demand as Nvidia's lineup, there's still quite some demand for it, to the point even their partner developers have a hard time acquiring one of those.
And something that's even more relevant is that AMD has a shit ton of demand for their Zen dies, which go into Epyc and Ryzen, which are far smaller than a GPU die and sell way more. Taking fab allocation out of those to make gamer GPUs would be moving away from high profit/high volume SKUs.

Yeah, in 2024 Meta bought around 170,000 MI300X and 220,000 NVidia H100s. Microsoft bought 500,000 NVidia and 100,000 AMD cards.

FT Article on it

 

[Tomorrow]

At this point I would say what AMD is doing is deliberate. They simply do not want to increase their market share, if they wanted to, they would reduce the price of their cards, and would have manufactured more. They do not have the A.I. chip demand that nGreedia has, so production allocation away from high profit SKU's is not such a big deal to AMD.

There have been many whispers of collusion between the two family members of these two companies for years. But another, less conspiratorial take would be AMD management is simply incompetent...

Yes i have to agree. As crazy as it sounds it's the only conclusion that makes sense. No one can be that incompetent for 10+ years in a row by accident or negligence. Which only leaves deliberate action. Im not sure i believe in outright collusion between green and red (the infamous meme about both shaking hands, but holding bombs behind their backs).

AMD as the runner-up has so many levers it could pull to grab market share. Low hanging fruit first.
Just like Ryzen team did with steady gains each year. When a company is at 90%+, it's much easier to lose market share even when doing everything right and Nvidia has not been doing everything right, far from it.

Based on circumstantial evidence AMD was initially planning on releasing the 9070 XT at 699. Another -50 vs 5070 Ti this time.
Only public pressure eventually made them cave and lower it to 599, but that was short lived as very few people got their cards at 599 or anywhere close to it.

I paid ~13% over MSRP two months after the release and i consider myself lucky with that. (EU).
The only thing i can not fault is availability here in the EU. Availability of AMD cards has been very good and much better from what is see from their (and Nvidia's) US numbers.

This is strange. I'd like to know what is the statistical sample they are doing this research on.
What happened in Q4/2024 that AMD GPUs market share surged from 10% to 17% and Nvidia lost 8%?

Not sure, but their latest data is from Q1 2025 and RDNA4 launched in March of this year. Just at the tail end of Q1 so i dont expect it to influence the results until the Q2 numbers come out next month as Q2 ends in June. As for why Nvidia lost 8% in Q4 2024? I suspect this was because they stopped Lovelace production in anticipation of Blackwell launch and the channel started to run dry on 40 series cards. Perhaps this contributed to the loss.
Informed DYI builders naturally knew Blackwell was already coming so i suspect DYI sales of Lovelace crawled to a halt too.

Lot of people around me decided to go with team red this year. Sales of RX 9070 (XT) vastly outperformed sales of RTX 5080 at Mindfactory in March 2025.
I'd expect AMD to gain 2-3% market share, not lose 2%. Well, the big problem was unobtainium of RX 9070 (XT) at MSRP prices, but so were RTX 5090-80-70Ti.

Same here. I was on Nvidia since 2016. Total of past 9 years. Now im on AMD because this time their card made more sense for my needs and i got sick of Nvidia screwing me over. Im fairly sure in not in the minority this time. I've seem many more people try AMD this time. Either as their first or returning after long time like me.

 

[jnv11]

This is strange. I'd like to know what is the statistical sample they are doing this research on.
What happened in Q4/2024 that AMD GPUs market share surged from 10% to 17% and Nvidia lost 8%?

Lot of people around me decided to go with team red this year. Sales of RX 9070 (XT) vastly outperformed sales of RTX 5080 at Mindfactory in March 2025.
I'd expect AMD to gain 2-3% market share, not lose 2%. Well, the big problem was unobtainium of RX 9070 (XT) at MSRP prices, but so were RTX 5090-80-70Ti.

I think that Nvidia was intentionally draining the market to ensure that the RTX 40 series was sold out in Q4/2024 before Nvidia introduced the RTX 50 series cards. Nvidia did not want to compete against a supply of RTX 40 series cards like it did when it introduced the RTX 40 series which had to compete with the RTX 30 series cards that remained on shelves due to a massive oversupply and which proved to be great bargains versus lower end RTX 40 series cards.

 

[MegurineMiku]

Not only there is no performance progress in Ada vs. Blackwell, there is also no energy efficiency progress as well. Blackwell is just bigger 40xx with messed up drivers.

no it is not, blackwell doesnt support 32bit Physx and pulls more power while offering the same performance (5070 vs 4070s), it is a regression in every aspect apart from AI. People can argue that blackwell achieved same performance with less cores (same as 7800XT vs 6800XT) but gamers dont care about technical stuff.

 

[LittleBro]

Same here. I was on Nvidia since 2016. Total of past 9 years. Now im on AMD because this time their card made more sense for my needs and i got sick of Nvidia screwing me over. Im fairly sure in not in the minority this time. I've seem many more people try AMD this time. Either as their first or returning after long time like me.

And then if you look at reports from notable e-shops ... for instance sales of RX 9070s outperformed RTX 5080s sales in Mindfactory.de by ten times.
Especially RX 9070 XT's stock were a pain in the ass even in April 2025. Most of people pay for goods when they are delivered or picked up from shop.
Maybe we'll see impact of these late deliveries in Q2/2025 results.

 

[Daven]

+20% means to be on pair with the 5080 2 years later

That would mean the 070XT class would be on par with the 5080. That's pretty good actually. The 070XT class would even beat the 5080 if CUs are added or clocks are increased.

Oh and I just looked at the latest benchies. The 9070XT is only 16% behind the 5080 so the next generation being 20% faster would totally kill that $1500 (real price) part from a perf/$ perspective.

 

[Rightness_1]

no it is not, blackwell doesnt support 32bit Physx and pulls more power while offering the same performance (5070 vs 4070s), it is a regression in every aspect apart from AI. People can argue that blackwell achieved same performance with less cores (same as 7800XT vs 6800XT) but gamers dont care about technical stuff.

Blackwell (the chip) does support 32Bit Physx. It was a business decision by nGreedia to stop support for 32Bit Physx, and is imposed via the driver, not the chip.

As far as I can tell, Blackwell is the same chip with a few bolted on extras, with some changes to what is disabled or not, such as FP8 support, which was fused off on the previous architecture on their consumer chips. It's just overclocked more because they had to give some performance uplift over the previous gen.

I wonder if the 60x0 series is going to be a simple die-shrink of Blackwell with yet again extra cores added. I don't think nGreedia has made a new architecture for quite some time now. They are just bolting on more and more cores, but the cores are not new. Just minor changes to the PCiE interface, display engine and media blocks, the rest is mostly the same.

 

[R-T-B]

messed up drivers.

They are fixed now but otherwise, yeah. You are on target.

 

[igormp]

Blackwell (the chip) does support 32Bit Physx.

What's your source on that?

As far as I can tell, Blackwell is the same chip with a few bolted on extras, with some changes to what is disabled or not, such as FP8 support, which was fused off on the previous architecture on their consumer chips. It's just overclocked more because they had to give some performance uplift over the previous gen.

Again, what's your source on that? FP8 was already supported in Ada.

 

[Rightness_1]

What's your source on that?

Again, what's your source on that? FP8 was already supported in Ada.

NVIDIA Support

nvidia.custhelp.com

Have you never read GPU specifications charts before? There has been differentiation between the exact same chips from nVidia for many years, just because one is sold as a Quadro, and the other a GeForce card... But they are the same chip, just with different options fused, such as allowed max rate of FP etc and other limitations placed in the drivers.

 

[Sol_Badguy]

I am seeing ~zero IPC gains when comparing 30 series (Ampere) to 50 series (Blackwell).
The 5080 has the same general stats as the 3090 Ti.
The 3090 Ti has marginally (5%) higher bandwidth (which certainly won't make a difference) but the 5080 has massively increased L2 Cache (which certainly could make a difference). So there is this extra variable which might matter or not.

The 5080 however has much higher clock speed.
The official Boost clocks are 1860 MHz and 2617 MHz respectively. That means a 40.7% increase.
But let's use the real life gaming frequency: 1995 MHz and 2640 MHz.
Yes, the drivers used are different and the game suite used is different, however I don't expect significant changes using the same conditions.
So I'll take these frequencies as universally valid.
That would give a 32.3% increase for the 5080.

In the following (very recent) review, I see a 30.8% difference in raster at 1440p and 29.2% in raytracing at 1440p. This is using the same testing conditions, but we don't know the actual averaged frequencies. No matter, they're surely close to the ones above.

So basically the 5080 has about 30% better performance (raster and raytracing) and about 30% higher clock speed.
Sure it has that huge L2 Cache which might contribute a bit to this percentage, I'm assuming scaling isn't perfect, meaning 30% higher frequency doesn't translate to exactly 30% higher performance.

Thus for me the IPC gains are pretty much zero, no point in splitting hairs over 0.5...1% or whatever.

Now if this is progress or not, well... should we take a look at power draw to help us decide?

It would be great if AMD manages (sooner rather than later) to turn the tables on nVidia, not with this gen though, this was to help them get their foot in the door. If they make the right moves from now on it can happen, imagine that: Radeon having its Ryzen moment.

 

[igormp]

NVIDIA Support

nvidia.custhelp.com

Have you never read GPU specifications charts before? There has been differentiation between the exact same chips from nVidia for many years, just because one is sold as a Quadro, and the other a GeForce card... But they are the same chip, just with different options fused, such as allowed max rate of FP etc and other limitations placed in the drivers.

That link says nothing about the underlying 32-bit support in Blackwell, which you claim is supported. If it were the case, one would be able to run 32-bit CUDA on it through other means, which so far hasn't happened.

And you seem to be confusing yourself with some ideas. The RTX Pro 6000 and the 5090 share the same underlying chip, with some lasered units just as you said.
But the 4090 and the RTX 6000 Ada use completely different chips from the above.
The ISA between Ada and Blackwell is different as well.

 

[Rightness_1]

That link says nothing about the underlying 32-bit support in Blackwell, which you claim is supported. If it were the case, one would be able to run 32-bit CUDA on it through other means, which so far hasn't happened.

And you seem to be confusing yourself with some ideas. The RTX Pro 6000 and the 5090 share the same underlying chip, with some lasered units just as you said.
But the 4090 and the RTX 6000 Ada use completely different chips from the above.
The ISA between Ada and Blackwell is different as well.

I must be the only one getting confused by this statement... "CUDA Driver will not support 32-bit CUDA applications on GeForce RTX 50 series (Blackwell) and newer architectures".

 

[igormp]

I must be the only one getting confused by this statement... "CUDA Driver will not support 32-bit CUDA applications on GeForce RTX 50 series (Blackwell) and newer architectures".

That doesn't imply by any means that the underlying hardware still supports it.
I recommend you to take a look at how PTX and SASS work on Nvidia GPUs.

 

[jnv11]

I found several articles from Chips and Cheese that could help explain this apparent increase in IPC in RDNA 4:

RDNA 4's "Out-of-Order" Memory Accesses

Examining RDNA 4's out-of-order memory accesses in detail, and investigating with testing

chipsandcheese.com

This article shows that AMD's hardware scheduler prior to RDNA 4 wasted time stalling the pipeline due to false dependencies (a.k.a. WAW or write after write situations) where two instructions need to write, but the later write has to wait for the first write to finish before executing. RDNA 4 fixed that problem. Nvidia's and Intel's software scheduler in the driver either minimized or eliminated such stalls wherever possible by using the CPU to schedule those instructions at the cost of higher driver overhead. AMD now supports more out of order operations to boost IPC. It appears that one of the motivations to add more out of order operations is to improve ray tracing performance, but this optimization is probably helping a lot more than just ray tracing. This makes me wonder how much performance that AMD has been leaving on the table with its hardware scheduler due to design errors. Software schedulers can use CPU time to optimize the instructions that are sent to the GPU and can be fixed with driver updates if there is a design error. However, software schedulers add CPU overhead, and can't react to real-time condition changes on the GPU. Hardware schedulers can cut CPU driver overhead and could react to real-time condition changes on the GPU, but a bad hardware scheduler might require a new chip design to fix.

RDNA 4’s Raytracing Improvements

Raytraced effects have gained increasing adoption in AAA titles, adding an extra graphics quality tier beyond traditional “ultra” settings.

chipsandcheese.com

AMD introduced several new ray tracing optimizations like BVH instructions that do double the work of RDNA 3 BVH instructions and oriented bounding boxes which can cut the number of BVH instructions that need to be generated. This can raise the apparent IPC but actually lower the number of actual instructions that were executed, meaning that real work per instruction increased.

Dynamic Register Allocation on AMD's RDNA 4 GPU Architecture

Modern GPUs often make a difficult tradeoff between occupancy (active thread count) and register count available to each thread.

chipsandcheese.com

This shows another possible optimization that aims to increase GPU utilization by reducing register pressure. However, the graphs of the Port Royal benchmark GPU utilization in that article show that forcing the driver to launch ray tracing threads to force dynamic register allocation instead of running ray tracing within the threads that need ray tracing slows things down. Maybe thread calling and launching overhead overcomes the increased GPU utilization that dynamic register allocation allows. If AMD's driver team can figure out how to use dynamic register allocation if it is useful, this could be some real fine wine in the future. It is also possible that this implementation of dynamic register allocation could also have serious design errors that make it useless in RDNA 4 much like Nvidia's first attempt at asynchronous compute in Maxwell had design errors making asynchronous compute worse than useless in Maxwell, requiring Nvidia to beef up the scheduling hardware to make asynchronous compute work in Pascal.

EDIT: Forgot about hardware being able to react to real-time conditions on the GPU.