AMD's Frank Azor Expects Upcoming Presentation to Fully Detail RDNA 4 GPUs

[AusWolf]

Gotta love these "say something without saying anything" social media posts.

 

[alwayssts]

Considering the 5080 chip: 370mm²
Approximate dies per wafer (ignoring defects) = 181 dies x 0.85 (85% yield rate) = 153 chips

Original N4P = $17,000, $111/chip.
N4C = $15,555 = $101/chip

AMD should use 20Gbps (cheap, and in good supply), Samsung never put 24Gbps chips into mass production... AMD only sells chips, so its profit margin must be calculated per unit. While the percentage margin may be high, the absolute profit per chip is relatively small. For example, a 100% margin on a $100 chip still results in just $100 of profit. However, AMD has hundreds of millions to billions in R&D, manufacturing, and operational costs that must be recouped through these sales. Then, $600 is possible, but AMD needs to convince everyone in the chain to tighten margins for this to work.

On Nvidia’s side, they weren’t satisfied with just selling chips. Now, they’ve secured massive quantities of GDDR7 to bundle with their GPUs, ensuring AIBs buy the complete set—further increasing their profits.

I am figuring ~70% yield. It is likely higher and therefore chips cheaper. I am being conservative and showing how these products are figured out (roughly).
Yes, 20gbps cheap and why using it. 24gbps is a BIG question mark. It was made, 27gbps was not (bc it couldn't yield). Don't trust the 'sample' listed by Samsung; they do that to protect their customers.
Just like GDDR7 pricing, which isn't as high as you may think. They want you to think it is, but it isn't, esp for nvidia. They ofc use cheap Micron as leverage against Samsung/Hynix which make better chips.
It doesn't even matter (GDDR7) anyway bc it doesn't effect the performance tiers. It's all a gimmick. 5080 literally can't take advantage of more than ~18GB/~26.3gbps on avg OCed...but has 16GB.
They totally didn't clock it (and limit it's potential) to around where 16GB 24gbps GDDR6 could OC. Totally not gonna refresh it above the actual limitation of 9070xtx with 24GB and actually use 18-20GB.
Just like GDDR6x (made cheaply by Micron using PAM4 versus more-expensive normal GDDR6 from Sam/Hynix). It's super weird how GDDR6x clocks less than actual 20gbps GDDR6. Oh wait. Micron.

nVIDIA's margins are insane. Freaking. Not. Okay. DO NOT BUY BC BAD VALUE AND THEY STILL CUT RAM/ABSOLUTE PERF BELOW EACH TIER. That is what I keep trying to explain.
It's genius marketing on 5D-chess levels, but awful for consumers whom fall for these gimmicks.
Many don't understand the limitations/planned obsolesence (until later). It's not shown correctly in reviews.
They sell the hype (32GB with 5090, GDDR7 as whole with 5080, high clocks with 5070), but in doing that miss the mark of actually matching bw/buffer/compute perf (which AMD actually does).

I'm truly hoping one of these days people actually test all these cards (minimum frameate to show limitations, absolute perf from OCing) in ways to show my point.

AMD does so many things that are smart to keep prices down for consumers (while still allowing the same absolute performance as a tier or more up from nVIDIA) and people just don't get it.
This includes bin at threshold clocks/high voltage, but leave that voltage/power limit for the capability of clocks on the table so you can OC it while matching absolute perf for buffer/bw/compute/process.
nVIDIA does the literal exact opposite which is make units that are a very good match at stock clocks, and efficient, but maxed out for RAM capabilties so overclocking doesn't make sense. Limted by buffer.
Typically...5070 will probably clock pretty high to get close to maxing out the capability of 12GB. It is yet another gimmick (high clocks) so they could (probably) cut the unit/engine count down to 4/6144sp.
They sell these cards that are on the wrong side of the general performance expectations (look at RAM/minimum framerates) on snake oil software...sometimes marketed as hardware.
Hey, cool. They can make 4/8-bit ops that take less space on the die than 16/32. They can make them do simple things...cool! They're also not FP32, and can't do FP32.
IOW, they are selling you less for more. They have trained people that because normally a FP32 unit would have to do that calc, but they can do it with lesser, it's a feature.
It's actually a way to save money (for them) AND sell it for more.

I know so many people just don't understand, and that's fine...it's just sad.
I'm not arguing the uses they've created and/or their software ecosystem they've built. While a lot of it is built on innovation from elsewhere, they did take the time/money to get it off the ground.
I am, however, explaining to you how their business (and products) is built on marketing/selling perpetual units.
Not actually smart engineering wrt to what's best for the customer.

On top of that their margins are literally out of control. It is apparently AMD's job to show people this.

 

[JustBenching]

599$-649$ for the 9070xt, raster performance +15% over 5080, no clue about RT.

 

[mb194dc]

Some people are wilfully deluding themselves. Like the 5080, you can't escape from the cards leaked hardware specifications. 64 compute units, only 256bit memory bus.

The 9070xt will be slower than the 7900xt, probably by a margin at 4k. Which you can get for 620 now.

No reason to wait or be excited for this launch, just like 5x nvidia cards. If you want an AMD card in the same price ball park, just drop the 620.

 

[dir_d]

Some people are wilfully deluding themselves. Like the 5080, you can't escape from the cards leaked hardware specifications. 64 compute units, only 256bit memory bus.

The 9070xt will be slower than the 7900xt, probably by a margin at 4k. Which you can get for 620 now.

No reason to wait or be excited for this launch, just like 5x nvidia cards. If you want an AMD card in the same price ball park, just drop the 620.

I agree with the performance but if the price is right they will sell out everywhere.

 

[LastDudeALive]

If I was an AIB partner I would be furious with AMD right now. Preemptively conceding this generation to Nvidia by leaving GPUs out of your CES presentation and killing the hype was bad enough, but the fact that they pushed the launch back by a month (for prices or software, it doesn't matter) means that inventory is sitting in warehouses, burning cash. And AIBs will make even less money with lowered MSRPs. Even though there's no RTX 50 stock, very few people will decide to get an RX 9000 instead, there's just no compelling marketing or features. Plus, like last gen, there's plenty of unsold RX 7000 cards, so AMD is really just competing against itself.

 

[AusWolf]

If I was an AIB partner I would be furious with AMD right now. Preemptively conceding this generation to Nvidia by leaving GPUs out of your CES presentation and killing the hype was bad enough, but the fact that they pushed the launch back by a month (for prices or software, it doesn't matter) means that inventory is sitting in warehouses, burning cash. And AIBs will make even less money with lowered MSRPs. Even though there's no RTX 50 stock, very few people will decide to get an RX 9000 instead, there's just no compelling marketing or features. Plus, like last gen, there's plenty of unsold RX 7000 cards, so AMD is really just competing against itself.

So what would you get instead of a 9070? A 5070? Oh wait, it still isn't out, just like the 9070. You can wait for one, but can't wait for the other? So how did AMD "concede" this generation? Remember, AMD is not competing with the 5090 or 5080, so whether those cards are out or not is inconsequential.

 

[alwayssts]

Some people are wilfully deluding themselves. Like the 5080, you can't escape from the cards leaked hardware specifications. 64 compute units, only 256bit memory bus.

The 9070xt will be slower than the 7900xt, probably by a margin at 4k. Which you can get for 620 now.

No reason to wait or be excited for this launch, just like 5x nvidia cards. If you want an AMD card in the same price ball park, just drop the 620.

Yeah....I really don't think they want to sell 7900xt for less than $600. That's why they made it (also limited clock potential of 7800xt).
If you figure out the die size, it's essentially a native 7900xt (combined on 4nm) but also beefed up N32 bc clocks/tensor (combined on 4nm).
You could go at it from either direction and come to the same conclusion.
That 4GB of ram probably makes <$600 feasible, and allowed them to add 'tensor' ops while making a well-balanced/cheap (if not high-clocking and close to maxing the 16GB buffer) 256-bit/16GB chip.
As I've said before, 7900xt is pretty much limited to the capabilities of 16GB (just like 4080s/5080), but this will do it by using less units and potentially the high-clocking capability of the process.
IOW, the absolute (oc) performance of these cards should lay between 7800xt/7900xt, with both of those cards encompassing the spectrum of usefulness wrt 16GB. Hope that makes sense.

So what would you get instead of a 9070? A 5070? Oh wait, it still isn't out, just like the 9070. You can wait for one, but can't wait for the other? So how did AMD "concede" this generation? Remember, AMD is not competing with the 5090 or 5080, so whether those cards are out or not is inconsequential.

AMD is (probably) competing with 5060ti/5070....but also the 4070 series/5070, 4080-5070ti, and 5070ti-5080. It depends on which way you look at it. As an average Joe, or a nerd.

They tend to sell to average Joe's, and typically keep us nerds happy.

 

[Firedrops]

Gotta love these "say something without saying anything" social media posts.

Out of infinite generosity I have helped TPU write their next article the next time some random guy mentions AMD on social media:



In a highly anticipated and momentous occasion that set the stage for what could only be described as a tantalising glimpse into the future of graphical processing power, Advanced Micro Devices (AMD), a venerable titan in the realm of semiconductor innovation, embarked upon the grand unveiling of its inaugural wave of cutting-edge RDNA 4 graphics cards. This historic debut, which took place at the dawn of the calendar year, ushered forth the arrival of the much-speculated-upon Radeon RX 9070 XT and its ostensibly more modest sibling, the RX 9070 (non-XT), into the ever-evolving technological landscape that is the PC gaming hardware sector.


However, rather than a smooth and seamless transition into the collective consciousness of an eager audience comprised of hardware aficionados, analysts, and industry watchers alike, AMD’s enigmatic and somewhat perplexing approach to presenting these much-anticipated components resulted in an aura of bewilderment. The press corps, along with fervent gaming hardware enthusiasts, found themselves caught in a veritable quagmire of ambiguity, as Team Red opted for a strategy that, at best, could be described as unconventional and, at worst, as downright confounding.


At CES 2025, an event renowned for showcasing the cutting edge of technological prowess, AMD extended to select attendees the rare and privileged opportunity to physically interact with demonstration samples of these latest RDNA 4-based offerings. And yet, despite this hands-on experience, a curiously stringent shroud of secrecy was maintained around the technical intricacies and performance benchmarks that typically accompany such a landmark unveiling. Board partners, typically instrumental in disseminating key details to the wider market, remained curiously silent—perhaps bound by unseen and unspoken obligations to refrain from shedding light on what many desperately sought to comprehend.


In the aftermath of this enigmatic debut, a steady trickle of leaks, whispers, and unverified murmurs began to permeate the ether of the online tech community. These snippets of speculative information, purportedly emanating from well-placed industry insiders, suggested that AMD had originally laid out an ambitious timeline targeting a late-January commercial launch of the new Radeon lineup. However, unforeseen circumstances appear to have necessitated a re-evaluation of these plans, with a revised schedule now seemingly aligning with an as-yet-unconfirmed timeframe in the not-too-distant future. Speculation runs rife that a rescheduled rollout looms on the horizon, with next month emerging as the most probable juncture for an official release. Lending credence to this notion, a report originating from Benchlife, relayed via the ever-watchful eyes at VideoCardz, postulates that a pre-launch showcase extravaganza has been tentatively earmarked for the latter days of February—a revelation that has sent ripples of anticipation throughout the enthusiast community.


Against this backdrop of uncertainty, a lone voice emerged from the digital expanse, daring to engage in direct dialogue with one of AMD’s key figures. Enter the intrepid and unwavering Dee Batch, a steadfast supporter of Radeon-branded gaming hardware, who took it upon himself to reach out to none other than AMD’s Chief Architect of Gaming Solutions, Frank Azor. In a missive imbued with both urgency and sincerity, Batch implored Team Red’s executive leadership to seize what he perceived to be a golden opportunity to sway the allegiances of prospective GPU purchasers. He posed a pressing question: “Can we see the RDNA 4 full presentation? I honestly feel you can prevent many gamers from getting a GeForce RTX 5070 or RTX 5070 Ti GPU... Please, do not miss this opportunity to gain gamer mind share.”


The response from Azor was succinct, almost tantalisingly so. With an air of cryptic certainty, he delivered a reply consisting of a mere handful of words: “Yes, full details are coming soon.” And with that, the floodgates of speculation were once again flung wide open. This brief yet impactful exchange acted as a lightning rod for further discourse, as additional members of the community waded into the discussion. VideoCardz noted that the AMD executive was demonstrably absorbing the deluge of consumer feedback, particularly in relation to the much-debated and hotly anticipated manufacturer’s suggested retail prices (MSRPs) for the forthcoming RDNA 4 series.


Adding further intrigue to the ongoing pricing conundrum, Azor himself took to the digital sphere in late January to emphatically quash a rumour that had been circulating regarding the Radeon RX 9070 XT potentially carrying an eye-watering $899 price tag at launch. In stark contrast, rival chipmaker NVIDIA has already laid its cards on the table, formally announcing that its GeForce RTX 5070 will debut at a relatively more palatable $549, while the slightly more powerful RTX 5070 Ti is set to command a price of $749. As the gaming and hardware communities collectively hold their breath in anticipation, devoted AMD enthusiasts remain hopeful that Team Red will rise to the occasion, presenting a pricing strategy that is not only competitive but also capable of galvanising the masses into choosing Radeon over the ever-present allure of GeForce.


And so, the saga of RDNA 4 continues, a tale woven with intrigue, speculation, and the ever-present promise of groundbreaking revelations yet to come.

 

[AusWolf]

Out of infinite generosity I have helped TPU write their next article the next time some random guy mentions AMD on social media:



In a highly anticipated and momentous occasion that set the stage for what could only be described as a tantalising glimpse into the future of graphical processing power, Advanced Micro Devices (AMD), a venerable titan in the realm of semiconductor innovation, embarked upon the grand unveiling of its inaugural wave of cutting-edge RDNA 4 graphics cards. This historic debut, which took place at the dawn of the calendar year, ushered forth the arrival of the much-speculated-upon Radeon RX 9070 XT and its ostensibly more modest sibling, the RX 9070 (non-XT), into the ever-evolving technological landscape that is the PC gaming hardware sector.


However, rather than a smooth and seamless transition into the collective consciousness of an eager audience comprised of hardware aficionados, analysts, and industry watchers alike, AMD’s enigmatic and somewhat perplexing approach to presenting these much-anticipated components resulted in an aura of bewilderment. The press corps, along with fervent gaming hardware enthusiasts, found themselves caught in a veritable quagmire of ambiguity, as Team Red opted for a strategy that, at best, could be described as unconventional and, at worst, as downright confounding.


At CES 2025, an event renowned for showcasing the cutting edge of technological prowess, AMD extended to select attendees the rare and privileged opportunity to physically interact with demonstration samples of these latest RDNA 4-based offerings. And yet, despite this hands-on experience, a curiously stringent shroud of secrecy was maintained around the technical intricacies and performance benchmarks that typically accompany such a landmark unveiling. Board partners, typically instrumental in disseminating key details to the wider market, remained curiously silent—perhaps bound by unseen and unspoken obligations to refrain from shedding light on what many desperately sought to comprehend.


In the aftermath of this enigmatic debut, a steady trickle of leaks, whispers, and unverified murmurs began to permeate the ether of the online tech community. These snippets of speculative information, purportedly emanating from well-placed industry insiders, suggested that AMD had originally laid out an ambitious timeline targeting a late-January commercial launch of the new Radeon lineup. However, unforeseen circumstances appear to have necessitated a re-evaluation of these plans, with a revised schedule now seemingly aligning with an as-yet-unconfirmed timeframe in the not-too-distant future. Speculation runs rife that a rescheduled rollout looms on the horizon, with next month emerging as the most probable juncture for an official release. Lending credence to this notion, a report originating from Benchlife, relayed via the ever-watchful eyes at VideoCardz, postulates that a pre-launch showcase extravaganza has been tentatively earmarked for the latter days of February—a revelation that has sent ripples of anticipation throughout the enthusiast community.


Against this backdrop of uncertainty, a lone voice emerged from the digital expanse, daring to engage in direct dialogue with one of AMD’s key figures. Enter the intrepid and unwavering Dee Batch, a steadfast supporter of Radeon-branded gaming hardware, who took it upon himself to reach out to none other than AMD’s Chief Architect of Gaming Solutions, Frank Azor. In a missive imbued with both urgency and sincerity, Batch implored Team Red’s executive leadership to seize what he perceived to be a golden opportunity to sway the allegiances of prospective GPU purchasers. He posed a pressing question: “Can we see the RDNA 4 full presentation? I honestly feel you can prevent many gamers from getting a GeForce RTX 5070 or RTX 5070 Ti GPU... Please, do not miss this opportunity to gain gamer mind share.”


The response from Azor was succinct, almost tantalisingly so. With an air of cryptic certainty, he delivered a reply consisting of a mere handful of words: “Yes, full details are coming soon.” And with that, the floodgates of speculation were once again flung wide open. This brief yet impactful exchange acted as a lightning rod for further discourse, as additional members of the community waded into the discussion. VideoCardz noted that the AMD executive was demonstrably absorbing the deluge of consumer feedback, particularly in relation to the much-debated and hotly anticipated manufacturer’s suggested retail prices (MSRPs) for the forthcoming RDNA 4 series.


Adding further intrigue to the ongoing pricing conundrum, Azor himself took to the digital sphere in late January to emphatically quash a rumour that had been circulating regarding the Radeon RX 9070 XT potentially carrying an eye-watering $899 price tag at launch. In stark contrast, rival chipmaker NVIDIA has already laid its cards on the table, formally announcing that its GeForce RTX 5070 will debut at a relatively more palatable $549, while the slightly more powerful RTX 5070 Ti is set to command a price of $749. As the gaming and hardware communities collectively hold their breath in anticipation, devoted AMD enthusiasts remain hopeful that Team Red will rise to the occasion, presenting a pricing strategy that is not only competitive but also capable of galvanising the masses into choosing Radeon over the ever-present allure of GeForce.


And so, the saga of RDNA 4 continues, a tale woven with intrigue, speculation, and the ever-present promise of groundbreaking revelations yet to come.

And this proves that their strategy works. When people spend their precious time writing pieces like this, they're not spending that time talking about the 5070, which we all know where will land in both performance and price.

 

[Jtuck9]

They sell these cards that are on the wrong side of the general performance expectations (look at RAM/minimum framerates) on snake oil software...sometimes marketed as hardware.
Hey, cool. They can make 4/8-bit ops that take less space on the die than 16/32. They can make them do simple things...cool! They're also not FP32, and can't do FP32.
IOW, they are selling you less for more. They have trained people that because normally a FP32 unit would have to do that calc, but they can do it with lesser, it's a feature.
It's actually a way to save money (for them) AND sell it for more.

I'm wondering where these margins go

NVIDIA Reveals Secret Weapon Behind DLSS Evolution: Dedicated Supercomputer Running for Six Years

At the RTX "Blackwell" Editor's Day during CES 2025, NVIDIA pulled back the curtain on one of its most powerful tools: a dedicated supercomputer that has been continuously improving DLSS (Deep Learning Super Sampling) for the past six years. Brian Catanzaro, NVIDIA's VP of applied deep learning...

www.techpowerup.com