AMD Reveals RDNA 4 GPU Architecture Powering Next Gen Radeon RX 9070 Cards

Long-time Slashdot reader MojoKid writes: AMD took the wraps of its next gen RDNA 4 consumer graphics architecture Friday, which was designed to enhance efficiency over the previous generation, while also optimizing performance for today's more taxing ray-traced gaming and AI workloads. RDNA 4 features next generation Ray Tracing engines, dedicated hardware for AI and ML workloads, better bandwidth utilization, and multimedia improvements for both gaming and content creation. AMD's 3rd generation Ray Accelerators in RDNA offer 2x the peak throughput of RDNA 3 and add support for a new feature called Oriented Bounding Boxes, that results in more efficient GPU utilization. 3rd Generation Matrix Accelerators are also present, which offer improved performance, along with support for 8-bit float data types, with structured sparsity.

The first cards featuring RDNA 4, the Radeon RX 9070 and 9070 XT go on sale next week, with very competitive MSRPs below $600, and are expected to do battle with NVIDIA's GeForce RTX 5070-class GPUs
The article calls it "a significant step forward" for AMD, adding that next week is "going to be very busy around here. NVIDIA is launching the final, previously announced member of the RTX 50 series and AMD will unleash the 9070 and 9070 XT."

Re:

[Valgrus Thunderaxe]

No, that would be the blood of Christian babies.

Re: RDNA

[zawarski]

That accounts for the efficacy.

Re:

[Growlley]

thought that was the blood of christian virgins but going by the divorce rate amongst them - there aren't any.

Assuming this doesn't turn into a paper launch

[rsilvergun]

And gamers can actually get the cards then it'll be a huge leg up for AMD. The cards are going to be $550 to $600. The equivalent in video card is currently selling for between $900 and 1,500. That's not scalpers that's what the board partners are selling it for.

no

[Anonymous Coward]

A lot of the cards are going to be premium cards, they will go for well over $600. And the non-XT version is being produced in far smaller numbers (basically, only as many as they have bad dice that don't make the cut for XT). But hopefully there's enough $600 ones to satisfy the hungriest customers.

MDNA would have been better name

[zawarski]

College kids would have bought them by the dozen.

Re:It's more than hardware...

[bjoast]

As a Linux user, I have the exact opposite experience.

Re:

[Narcocide]

AMD really cleaned up their act with regards to Linux support since then. Since about the RX 560's, AMD's new generation (post-radeon "amdgpu") drivers have been markedly more stable than NVidia's offering and at least arguably parallel in performance, with some outright wins. Nvidia, on the other hand, has stagnated, and only looks high-performing and stable compared to other offerings from way back then.

Re:

[drinkypoo]

If only they would clean up their act WRT ROCm, they have already announced that it won't officially support "all" RDNA 4 cards out of the box so some users will have to rely on third parties for that.

You're right about Nvidia drivers, they are pretty bad right now. And Nvidia designs are also bad. The ongoing power connector problem is pathetic and thermal imaging testing shows that it's Nvidia's fault again, they are drawing the power mostly from two wires. And hacks which put resistors back into the desi

I had good luck with my RX 580

[rsilvergun]

... Until I didn't. It actually works really well running off my old i5 that I have connected to My TV in the living room. But for some reason as soon as I put it in my at the time brand spanking new Ryzen 5600 based computer it kept tanking and I reboot to find the driver replaced with the default Microsoft one. I didn't solve it because I picked up a GTX 1080 off eBay because they had gotten ridiculously cheap but I gathered the problem is Microsoft is, or probably was I mean it's been a while, replacing the updated AMD driver I had installed with whatever they have on Windows update and not asking me whether I want to do that or not.

This is a classic case of AMD doing the right thing and releasing a new driver to Microsoft every quarter like they're supposed to and it biting them in the ass. Nvidia by comparison hasn't updated their driver since the OS shipped.

Also board partners have a nasty habit of cutting corners on AMD with the electronics. Gigabyte and ASUS or particularly bad about this. So unless you know that and go with a sapphire or XFX you are probably in a world of hurt. And finally you can't plug an AMD card into an unstable system. Nvidia will use tricks like the old via four-in-1 driver did to detect unstable hardware and avoid using the cards features whereas an AMD card assumes you know what you're doing and so if your motherboard manufacturer cheeped out on your PCI slot or your power supply isn't up the snuff welcome to crash City. Oh and you may very well have to undervolt your core because the crummy board partner you bought a card from over volted it so that you would get better numbers in benchmarks and they can charge you an extra $50 or a hundred bucks for a custom overclock...

That's a lot of information to have to be aware of when you're buying a video card and so for most users they just haul off and buy the Nvidia solution. AMD needs to ideally police their board partners better and they need to add more functionality to their software to detect skeezy hardware and automatically down clock and slow down the video card

Re:

[DrMrLordX]

Fuck NV and 12VHPWR/12V-2x6, space invaders, bumpgate, overpriced 8GB cards, and all their other bullshit.

Re:

[thegarbz]

12VHPWR isn't an NVIDIA thing it's a PCI-SIG standard, and there are AMD cards such as the RX 7900XT which also feature the connector. As for space invaders, AMD had problems with their cards too, especially when you want to extend your range back 20 years all the way to bumpgate.

On the other hand your complaints about overpriced 8GB cards is completely relevant, as is other bullshit (paper launches comes to mind).

Re:

[DrMrLordX]

AMD hasn't used 12VHPWR or 12V-2x6 on any reference design. NV meanwhile does stuff like this:

https://m.youtube.com/watch?v=... [youtube.com]

Re:

[bloodhawk]

Exact opposite experience, currently AMD user as Nvidia drivers funked me over so many times.

Vastly undercuts Nvidia

[locater16]

Let's watch autistic fanboys undercut the entire industry they're entertained with by screaming, stomping, and buying Nvidia products anyway.

8-bit floats?

[Waccoon]

Can anyone elaborate on what uses there are for 8-bit floats in a GPU? That's not much precision, so what can you do with that?

Re: 8-bit floats?

[madsh]

Large Language Models - so GPU might not be the best name for this co-processor, that outperform the main CPU.

Re:

[thegarbz]

That's not much precision, so what can you do with that?

Computer graphics doesn't need precision. Errors can propagate without much visual impact in many cases. The point of keeping the complexity low is to increase speed / efficiency. Matrix multiplication at low precision is what GPUs are fundamentally good at and it has been introduced a few years back to speed up AI work loads too (Nvidia H100 has an 8bit float processing unit).

Re:

[VaccinesCauseAdults]

8-bit floats are useful for large language models where the very low precision is acceptable and the reduced memory size and higher throughput leads to faster/cheaper training and inference. https://docs.nvidia.com/deeple... [nvidia.com]

Re:

[bn-7bc]

I must be missing something, but hvat usecare a bit floats, unless it is gor values between 0 and 1, i mean you have only 256 values to use. Or is an 8 bit float significantly longer than 8 bit, and in that case what etectheva bits we are talking about here?

Re:

[VaccinesCauseAdults]

A FIXED-point format has a fixed absolute error and the representable values are distributed linearly. For example, 0/255, 1/255, 2/255 and so on, up to 255/255 would be an 8-but fixed point format with range [0,1]. Large numbers gave more relative precision. Whereas a FLOATING-point format has (approximately) fixed relative error and the representable values are distributed (approximately) logarithmically. It has the same number of values available between each power of two. Large and small numbers have t