AMD confirms big power-efficiency gains with Radeon 7000 GPUs coming November 3

[Daniel Sims]

Something to look forward to: Hot on the heels of Nvidia's unveiling of the RTX 4000 graphics cards, AMD announced a release date for its latest GPU generation. The company expects to launch RDNA 3 graphics cards on a schedule similar to Nvidia's Ada Lovelace rollout this fall.

Around the same time Nvidia announced the price and launch dates of its new RTX 4000 series graphics cards this week, AMD confirmed it plans to reveal and launch its RDNA 3 GPUs on November 3. The company promises more details as the event approaches.

Recent rumors suggest AMD will first release RDNA 3's flagship card -- the Radeon RX 7900 -- based on the Navi 31 GPU. Navi 33 and Navi 32 GPUs are to follow. That cadence echoes the plan Nvidia just confirmed to launch the GeForce RTX 4090 on October 12, followed by the 16GB and 12GB variants of the RTX 4080 in November.

There aren't many concrete details on RDNA 3, but we know it uses TSMC's 5nm node process. The series will also support AV1 encoding and decoding. Intel's freshman Arc GPUs were the first to include AV1 encode, with shocking performance results. Support for DisplayPort 2.0 and its 80 Gbps bandwidth will enable RDNA 3 cards to push video feeds up to 240Hz at 4K.

Join us on November 3rd as we launch RDNA 3 to the world! More details to come soon! #RDNA3 #AMD pic.twitter.com/oftq1Fjrgt

— Scott Herkelman (@sherkelman) September 20, 2022

Power efficiency is one of AMD's primary focus points. On Monday, the company's blog highlighted the performance-per-watt gains of its last two generations of GPUs, claiming RDNA 3 will bring similar improvements.

The first generation of RDNA cards in 2019 increased performance-per-watt by 50 percent over GCN, and RDNA 2 improved by 65 percent over RDNA. The company says RDNA 3 should gain another 50 percent performance-per-watt over RDNA 2 through a new generation of Infinity Cache, a switch to a chiplet design inspired by the company's recent Ryzen CPUs, an optimized graphics pipeline, and rearchitected compute units.

Lower power consumption could make RDNA 3 GPUs attractive to users living in countries like the UK, where energy bills have seen worrying increases. In June, AMD expressed fears that by 2025, GPU TDPs could reach 600-700W. Concerning rumors suggested the RTX 4090 might draw 600W, but the card's official spec sheet says it will require 450W, the same as last generation's 3090 Ti.

Permalink to story.

https://www.techspot.com/news/96043-amd-confirms-big-power-efficiency-gains-rdna3-gpus.html

 

[Iceblazer]

"The first generation of RDNA cards in 2019 increased performance-per-watt by 50 percent over GCN, and RDNA3 improved by 65 percent over RDNA."

Sixth paragraph was probably meant to read that RDNA2 improved 65% over RDNA instead of RDNA3. (Ed. note: thanks, now corrected)

 

[Eldritch]

Their performance, RT tech, answer to DLSS 3 and price will be deal maker/breaker. With stratospheric prices of Nvidia, the field is wide open. Let's see if AMD can score a homerun.

 

[Mr Majestyk]

Can't wait to see what the 7800XT brings with FSR 3.0 and how much they have improved RT this time around. Rasterisation won't be a problem vs Lovelace, it's RT that needs to at least be fair bit stronger than Ampere, even if it doesn't beat Lovelace. With FSR it won't really matter that much about RT, as no one will run it sans FSR anyway.

 

[cheapcomputers]

¨The company says RDNA 3 should gain another 50 percent performance-per-watt¨

That´s good news. Polaris and Vega didn´t compare very well to Nvidia in efficiency. It would be nice to see RDNA 3 in a Ryzen 7*** desktop apu.

 

[Last1Standing]

If anything we've ALL learned from sneaky AMD's CEO Su from Zen3/RDNA2 releases, is that she's the perfect snake to strike at consumers when they could cease the moment with high prices.

Don't look for any relieve from AMD if they could match Lovelace, especially if they manage to outperform Lovelace even if its only around ~3%!

 

[Kam7r]

Where are those nvidia drones who where making fun of the R9 290 gen' when amd was bad in efficiency ? we don't ear them anymore ? I don't know why ? even if my electricity is still cheap , I'll never buy a 500w Gpu... hope AMD deliver... even tho I'm currently satisfied with my 6700xt

 

[poltevo]

Don’t expect prices too be much lower then Nvidia’s. I do expect they will be slightly cheaper, do well in raster and be more efficient. However AMD will continue to lag behind in ray tracing and well behind in tensor performance.

 

[NumberSix]

Power efficiency sounds more appealing to me and could sway me away from Nvidia, last time I had an AMD card it said ATI on it so its been a while.

 

[EnesK]

If they can match 4080 12gb with 7700xt, that would be awesome since they can claim they beat 80 level card with 70 level card. There were rumors about 7700 xt was faster than 6900 xt which makes it around 3090/ti levels which is better than 4080 12gb, that sounds about right.

 

[takaozo]

https://news.ycombinator.com/item?id=17804494#:~:text=of%20AMD's%20CEO-,%22Technically%2C%20Lisa%20Su's%20grandfather%20is%20Jen%2DHsun%20Huang's%20uncle,uncle%2C%20but%20close%20relatives.%22

Is Lisa Su related to Jensen?


"Technically, Lisa Su's grandfather is Jen-Hsun Huang's uncle. They are not exactly niece and uncle, but close relatives."

Both competing in large corporations and Lisa looks to be a better CEO, let's wait for cards to be reviewed on both sides and maybe for the market to cool down a bit.

 

[R00sT3R]

If AMD want to win the next GPU cycle all they have to do is launch their cards at $200 less than Nvidias across the stack.

$1200 for a (real) 4080 is a mistake that will cost Nvidia dear.

 

[MarcusNumb]

Well I can see the 7xxx series will be about 100 to 150usd more expensive than the last gen, given the wide gap that NVIDIA left open. In a year of high inflation everywhere, it would be a miracle if they can keep the prices the same or even only 50usd more.
Personally, I'm not gonna upgrade my GPU cause I paid too much for a GPU to use in only 2 years (3070ti - 850euros).

 

[Irata]

If AMD want to win the next GPU cycle all they have to do is launch their cards at $200 less than Nvidias across the stack.

$1200 for a (real) 4080 is a mistake that will cost Nvidia dear.

Nah, it will make them sell lots of 4070 12 GB based on 4080 16 GB reviews.

 

[Geralt]

My next card. I don't want the Nvidia sun in my rig.

 

[Puiu]

Their performance, RT tech, answer to DLSS 3 and price will be deal maker/breaker. With stratospheric prices of Nvidia, the field is wide open. Let's see if AMD can score a homerun.

Theoretically there should be really big RT improvements with this generation because they are adding WMMA instructions. It can also used for FSR. This means proper dedicated hardware acceleration for this type of workloads.

The addition of DP 2.0 is in my opinion really important. Nvidia for some reason decided to stick with DP 1.4.

 

[takaozo]

Looks to me that GPU market is keeping the trend to a luxury product.
Fewer and fewer will afford new products.
Some will buy used from the first ones.
Some will keep their current ones for longer time.
Some will go to console or mobile.

 

[Eldritch]

This means proper dedicated hardware acceleration for this type of workloads.

It should as just having the instruction set without dedicated fixed function hardware will be a huge mistake.

 

[WhiteLeaff]

The slide says >50%, so it's greater than 50%

I'm glad someone will take the path of efficiency, and not just push 600w+ TBP lol

 

[Theinsanegamer]

Looks to me that GPU market is keeping the trend to a luxury product.
Fewer and fewer will afford new products.
Some will buy used from the first ones.
Some will keep their current ones for longer time.
Some will go to console or mobile.

Some will just buy the GPUs in their price range like they always have. You can still buy $150-$200 GPUs.

 

[neeyik]

Theoretically there should be really big RT improvements with this generation because they are adding WMMA instructions. It can also used for FSR. This means proper dedicated hardware acceleration for this type of workloads.

The WMMA instructions are only in ROCm - which isn't designed for gaming use. Besides, the inclusion of the instructions doesn't necessarily mean RDNA 3 will be sporting separate matrix cores, as matrix operations are already supported in RDNA 2. It's CDNA that has them as a separate block.

 

[Uncle Al]

Goodbye NVIDIA ..... you won't be missed.

 

[Puiu]

The WMMA instructions are only in ROCm - which isn't designed for gaming use. Besides, the inclusion of the instructions doesn't necessarily mean RDNA 3 will be sporting separate matrix cores, as matrix operations are already supported in RDNA 2. It's CDNA that has them as a separate block.

The rocWMMA update specifies support for GFX9 but the rumour mentions GFX11 which is RDNA3.

What we do know for certain is that AMD does have some sort of rearchitected compute unit for ray tracing.

 

[tkabou]

If anything we've ALL learned from sneaky AMD's CEO Su from Zen3/RDNA2 releases, is that she's the perfect snake to strike at consumers when they could cease the moment with high prices.

Don't look for any relieve from AMD if they could match Lovelace, especially if they manage to outperform Lovelace even if its only around ~3%!

True, but that’s business and standard practice honestly. On the other hand, FREE MARKETS decide the winner (or at least in practice, should). There’s always enough suckers to be found to buy at launch, at the most expensive price, and unfortunately for the rest of us, prices follow sales patterns.

 

[neeyik]

What we do know for certain is that AMD does have some sort of rearchitected compute unit for ray tracing.

That much is a given. What AMD had in RDNA 2 is very much equivalent to what Nvidia had with Turing: first release tech is always fairly basic. It will be interesting to see if they continue with the current structure, where the TMUs perform the ray-triangle intersection calculations and the BVH traversal is done via the shader cores, or whether they go with something completely new.