AMD Radeon RX 9070 Series Technical Deep Dive

[BigMack70]

simple gamers buy what is best when it comes to their purchasing power

It's a small sample size, but all the people I know IRL who play games on PC are "simple gamers" who don't get involved in tech forums / etc. And all of them - for years - buy whatever Nvidia GPU is within their budget. None of them have ever purchased a GPU more expensive than around $400. They don't even think about AMD. I'm only beginning to hear them even think about AMD CPUs and AMD has been better than Intel for enthusiasts for years.

I don't think people on these forums recognize what a niche bubble the online tech space is. The 9070 XT is between fine and great for this bubble, depending on your expectations and what measuring stick you use.

For the "simple gamer"? This isn't going to do anything.

 

[mechtech]

Hopefully circumstances will never arise when a GPU like the 6400/6500 series ever have to exist as a discrete GPU ever again.

IMO they're fantastic GPUs for their intended purpose (providing extra shader power to weak integrated Radeon laptop graphics that already have a full media engine and all the expected display outputs).

They were never designed to be discrete GPUs, because they're not a complete GPU by modern expectations. I bet the engineers who worked on them cry themselves to sleep at night knowing how their great idea was misused in the most inappropriate and disappointing way imagineable :\

ya, but with todays market, a 4GB card with a chip like that for $100 would be a nice option......for htpc or a no igpu cpu replacement if needed...........especially if your card dies and you don't have $600 to get a new gpu.............

I remember my moms old pc gpu dying during covid, I had to get GT1030 2GB for $130 cause the next card up was $900 :|

 

[geniekid]

This was a very informative article summarizing everything we know about the 9070 cards for now. Eagerly awaiting the reviews next week!

 

[RandallFlagg]

We're comparing price/performance using die sizes for dies that were stable and performant, right? Just figured we'd take it to it's logical conclusion. It's 875 adjusted for inflation, so a 45% increase in price for a 68% increase in die size means it's a better value. N4C is also mature, it's the third revision of TSMC's 5nm process and is cheaper to produce than their high-end nodes.

While we're comparing die sizes, a more relevant look would be at last gen's 7800 XT - the 9070s are a value regression compared to that card.

Die size is a frankly stupid way to look at it.

What does just the wafer at an N4 cost vs N5 vs N7? What was the initial capital investment for the Fab?

Cost per transistor is a better way to measure it, and that hasn't changed much in a decade. In fact, it has been going up.

And these new GPUs have a *lot* more transistors in them than GPUs from a decade ago, or even just a few years ago.

For example, the 7700XT has ~28B transistors @ N5 node, while the 6700XT had ~18B @ N7 node. And the cost per transistor went up, so.. you get the picture.

 

[Denver]

Die size is a frankly stupid way to look at it.

What does just the wafer at an N4 cost vs N5 vs N7? What was the initial capital investment for the Fab?

Cost per transistor is a better way to measure it, and that hasn't changed much in a decade. In fact, it has been going up.

And these new GPUs have a *lot* more transistors in them than GPUs from a decade ago, or even just a few years ago.

For example, the 7700XT has ~28B transistors @ N5 node, while the 6700XT had ~18B @ N7 node. And the cost per transistor went up, so.. you get the picture.

View attachment 387350

SRAM can no longer keep pace with the growing density of logic and has become the most challenging component to scale in modern designs demanding ever-larger caches. While, at times, achieving a 50% performance gain requires nearly doubling the transistor count, this only further distorts your perspective.

Cumulative Scaling (28nm to 3nm):

• Logic: ~9.8x density increase (0.10x area).
• SRAM (N3B): ~3.42x density increase (0.29x area).
• SRAM (N3E): ~3.26x density increase (0.31x area).

Next, we must consider the development cost per chip. These costs are far from negligible. How many units must be sold to break even?

- Estimated Design Costs
Industry reports (IBS, Gartner) provide ballpark figures for total chip design costs (including EDA tools, IP licensing, engineering, and tape-out) at advanced nodes. These are rough averages and vary by chip complexity (like CPU vs. simple ASIC):

 

[N/A]

How do you explain that GCD absorbs x4 MCD 2.050 million with 16MB infinite cache 50MTr / mm² and still remains 150MTr / mm². Oh it's N7.

 

[RandallFlagg]

SRAM can no longer keep pace with the growing density of logic and has become the most challenging component to scale in modern designs demanding ever-larger caches. While, at times, achieving a 50% performance gain requires nearly doubling the transistor count, this only further distorts your perspective.

Cumulative Scaling (28nm to 3nm):

• Logic: ~9.8x density increase (0.10x area).
• SRAM (N3B): ~3.42x density increase (0.29x area).
• SRAM (N3E): ~3.26x density increase (0.31x area).

View attachment 387355

Next, we must consider the development cost per chip. These costs are far from negligible. How many units must be sold to break even?

- Estimated Design Costs
Industry reports (IBS, Gartner) provide ballpark figures for total chip design costs (including EDA tools, IP licensing, engineering, and tape-out) at advanced nodes. These are rough averages and vary by chip complexity (like CPU vs. simple ASIC):

View attachment 387356

Right, but in simpler terms there is no relationship between die size and cost when comparing different nodes.

Another way of looking at this is, if Nvidia were to make a 1080 Ti GPU today on N5 it would probably cost exactly the same to make as it did originally. However, the die would likely be ~1/7th the original size.

Originally it was 470mm^2 @ 16nm so it would probably be ~70-100mm^2 (67mm^2 literally, but as you say some parts haven't shrunk as much as the ideal transistor) on the N5 node, or roughly the size of a modern A18 Apple SoC that is on TSMC N3E.

 

[Rover4444]

Die size is a frankly stupid way to look at it.

What does just the wafer at an N4 cost vs N5 vs N7? What was the initial capital investment for the Fab?

Cost per transistor is a better way to measure it, and that hasn't changed much in a decade. In fact, it has been going up.

And these new GPUs have a *lot* more transistors in them than GPUs from a decade ago, or even just a few years ago.

For example, the 7700XT has ~28B transistors @ N5 node, while the 6700XT had ~18B @ N7 node. And the cost per transistor went up, so.. you get the picture.

View attachment 387350

If that's the case and the picture is true, I don't know what's going on with AMD. Production cost would be wildly over MSRP at 53,900 million transistors.

 

[RandallFlagg]

If that's the case and the picture is true, I don't know what's going on with AMD. Production cost would be wildly over MSRP at 53,900 million transistors.

No. Both Nvidia and AMD have huge margins on these chips, always have.

 

[Denver]

Right, but in simpler terms there is no relationship between die size and cost when comparing different nodes.

Another way of looking at this is, if Nvidia were to make a 1080 Ti GPU today on N5 it would probably cost exactly the same to make as it did originally. However, the die would likely be ~1/7th the original size.

Originally it was 470mm^2 @ 16nm so it would probably be ~70-100mm^2 (67mm^2 literally, but as you say some parts haven't shrunk as much as the ideal transistor) on the N5 node, or roughly the size of a modern A18 Apple SoC that is on TSMC N3E.

It would indeed be more expensive in 5nm, as you're essentially overlooking the development costs behind the chip and only factoring in the cost per isolated wafer, which doesn't align with reality
Companies use the density budget offered at each node to increase performance in the form of products that hook the customer, having RT and AI in the equation has made this a little worse, but at the end of the day that's the logic: basically more things to divide your area budget into.


"In the first half of the year the industry shipped 18.2 million graphics cards for desktop systems, up 46% from the same period in 2023, according to JPR. However, shipments of graphics boards in Q3 2024 totaled 8.1 million units — down from 8.9 million units in Q3 2023. This is believed to be a result of inventory correction at AMD as well as the end of both Ada Lovelace and RDNA 3 lifecycles."

More than 251 million GPUs shipped in 2024, according to new research

But what about discrete GPUs?

www.tomshardware.com

Nvidia’s dominant hold on 80-90% of the GPU market grants them the unique advantage of both healthy margins and competitive pricing. However, instead of fair prices, they have consistently opted to prioritize higher profit margins at any cost. In contrast, AMD, which holds only 10-20% of the market share, struggles to cover its development expenses with this limited revenue.

If AMD didn't have the consoles and other parallel businesses using Radeon IP they'd be out of business. Their margin in the gaming sector is too thin to pay for the R&D needed, I suppose that's also the reason for merging CDNA and RDNA: to simplify hardware and software development.

 

[RandallFlagg]

It would indeed be more expensive in 5nm, as you're essentially overlooking the development costs behind the chip and only factoring in the cost per isolated wafer, which doesn't align with reality
Companies use the density budget offered at each node to increase performance in the form of products that hook the customer, having RT and AI in the equation has made this a little worse, but at the end of the day that's the logic: basically more things to divide your area budget into.


"In the first half of the year the industry shipped 18.2 million graphics cards for desktop systems, up 46% from the same period in 2023, according to JPR. However, shipments of graphics boards in Q3 2024 totaled 8.1 million units — down from 8.9 million units in Q3 2023. This is believed to be a result of inventory correction at AMD as well as the end of both Ada Lovelace and RDNA 3 lifecycles."

More than 251 million GPUs shipped in 2024, according to new research

But what about discrete GPUs?

www.tomshardware.com

Nvidia’s dominant hold on 80-90% of the GPU market grants them the unique advantage of both healthy margins and competitive pricing. However, instead of fair prices, they have consistently opted to prioritize higher profit margins at any cost. In contrast, AMD, which holds only 10-20% of the market share, struggles to cover its development expenses with this limited revenue.

If AMD didn't have the consoles and other parallel businesses using Radeon IP they'd be out of business. Their margin in the gaming sector is too thin to pay for the R&D needed, I suppose that's also the reason for merging CDNA and RDNA: to simplify hardware and software development.

That chart attempts to aggregate all of that into one display. We're not arguing here, I'm just saying that you can present the information in a simple easy way without digging into the morass of a million details.

Edit: What's going on is no different than what I see in other places, like cars. Cars are extremely expensive today compared to 10 years ago. But 10 years ago, I didn't have Apple CarPlay, backup sensors, lane keeping assist cameras, blind spot monitors, auto stop-start, hotspot WiFi with cellular WAN, satellite tracking my car, an app to see what my tire pressure is and decode any CEL codes, an LCD infotainment system, and so on. Some of that existed sure, but in higher trims and mostly higher end manufacturers.

So today everybody wants the features that used to only exist in high end cars, and guess what? They get to pay high end car prices.

Same in the GPU space. GPUs are not getting much more advanced, they're just getting more transistors in a smaller area hence higher performance. And the cost per transistor has not gone down, so they simply cost more.

Chips aren't getting cheaper to make — the cost per transistor stopped dropping a decade ago at 28nm

Google says chips aren't getting cheaper.

www.tomshardware.com

TSMC Will Reportedly Charge $20,000 Per 3nm Wafer

GPUs and SoCs to get more expensive

www.tomshardware.com

TSMC's 2nm process to be unprecedentedly costly as $30,000+ per wafer figure floated

As TSMC prepares the 2nm process, customers should brace themselves for a significant price hike. Recent reports indicate that costs may nearly double compared to previous technologies.

www.notebookcheck.net

 

[Rover4444]

No. Both Nvidia and AMD have huge margins on these chips, always have.

Oh. In that case I'm just gonna ignore that, then... AMD still should have priced their cards at $450/$550.

 

[RandallFlagg]

Oh. In that case I'm just gonna ignore that, then... AMD still should have priced their cards at $450/$550.

Neither AMD nor Nvidia set the prices for their cards, the market does that. Just like you can't get a 5070 Ti for $750, you won't be able to get one of these cards for $600.

I do think they both stumbled into a new strategy during COVID though.

They post a complete BS MSRP, knowing they won't supply enough cards to market. Then they use all their capacity to make server CPUs (AMD) and AI accelerators (AMD) that keep corporate margins and profits high, and only produce enough GPUs to show that they do exist and maintain presence in those markets.

Then the market has its way with prices.

The traditional way to do this is to do market research and set the price based on a supply / demand curve, but I don't think either company cares enough about the GPU market to try to actually find the sweet spot in the curve. They clearly are not doing that now though.

What they are doing now is padding the pockets of retailers, to stay in good graces in case they need retail cooperation later on (if say, crypto and AI both collapse). MSRPs are just a trick right now. If they could get away with it, I bet neither AMD nor Nvidia would make a single GPU.

 

[Hecate91]

Neither AMD nor Nvidia set the prices for their cards, the market does that. Just like you can't get a 5070 Ti for $750, you won't be able to get one of these cards for $600.

I do think they both stumbled into a new strategy during COVID though.

They post a complete BS MSRP, knowing they won't supply enough cards to market. Then they use all their capacity to make server CPUs (AMD) and AI accelerators (AMD) that keep corporate margins and profits high, and only produce enough GPUs to show that they do exist and maintain presence in those markets.

Then the market has its way with prices.

The traditional way to do this is to do market research and set the price based on a supply / demand curve, but I don't think either company cares enough about the GPU market to try to actually find the sweet spot in the curve. They clearly are not doing that now though.

What they are doing now is padding the pockets of retailers, to stay in good graces in case they need retail cooperation later on (if say, crypto and AI both collapse). MSRPs are just a trick right now. If they could get away with it, I bet neither AMD nor Nvidia would make a single GPU.

View attachment 387400

The market responded to the 5070Ti being $1000 by buying it anyway, the same happened during the crypto hype, people complained yet used bots or paid scalpers to buy a card at 2-3X over MSRP.
I think Nvidia could get away with it, especially since they're making loads of money with AI, gaming is just pocket change to them. But the same doesn't apply to AMD since they've only lost market share and don't have the mindshare buying a new GPU no matter how high the prices are. AMD claimed good availability with the 9070 and the rumor is cards have been in the hands of retailers since January, hopefully there is enough stock to meet demand.

 

[RandallFlagg]

The market responded to the 5070Ti being $1000 by buying it anyway, the same happened during the crypto hype, people complained yet used bots or paid scalpers to buy a card at 2-3X over MSRP.
I think Nvidia could get away with it, especially since they're making loads of money with AI, gaming is just pocket change to them. But the same doesn't apply to AMD since they've only lost market share and don't have the mindshare buying a new GPU no matter how high the prices are. AMD claimed good availability with the 9070 and the rumor is cards have been in the hands of retailers since January, hopefully there is enough stock to meet demand.

When MSRP is set at a level that is not correctly gauging demand, the market starts to correct for that - it's actually called 'price discovery'.

The market is looking for the correct price, and as you say it elevated it from $750 to what I see as around $1200 for the 5070 Ti. It may settle a little lower, it's still in price discovery. But I can guarantee you that the correct market price is not $750.

The reason AMDs market share shrank is that they did not make enough cards, plain and simple. If they made more, prices (of both AMD and Nvidia cards) would drop and move down on the curve.

This curve is a real thing. And I guarantee you that both companies know what the curve looks like for their products, and know they did not set their MSRP correctly for the supply they were going to generate. They just aren't interested in maximizing profits here, it's more of a hedge against disruptions in AI and Data Center spaces. Their profits come from there, not here.

Supply and Demand Curves Explained

In economics, supply and demand curves govern the allocation of resources and the determination of prices in free markets.

www.economicsonline.co.uk

 

[Rover4444]

The reason AMDs market share shrank is that they did not make enough cards, plain and simple. If they made more, prices (of both AMD and Nvidia cards) would drop and move down on the curve.

I don't think that's true. The 7000 series was well stocked throughout it's lifespan and many of their cards were priced poorly upon release, so they didn't sell enough volume and lost market share.

Neither AMD nor Nvidia set the prices for their cards, the market does that. Just like you can't get a 5070 Ti for $750, you won't be able to get one of these cards for $600.

I do think they both stumbled into a new strategy during COVID though.

They post a complete BS MSRP, knowing they won't supply enough cards to market. Then they use all their capacity to make server CPUs (AMD) and AI accelerators (AMD) that keep corporate margins and profits high, and only produce enough GPUs to show that they do exist and maintain presence in those markets.

Then the market has its way with prices.

The traditional way to do this is to do market research and set the price based on a supply / demand curve, but I don't think either company cares enough about the GPU market to try to actually find the sweet spot in the curve. They clearly are not doing that now though.

What they are doing now is padding the pockets of retailers, to stay in good graces in case they need retail cooperation later on (if say, crypto and AI both collapse). MSRPs are just a trick right now. If they could get away with it, I bet neither AMD nor Nvidia would make a single GPU.

View attachment 387400

If they want to increase market share this generation and push volume, which is their stated goal this generation, they must lower their prices. This pans out on the supply/demand graph by moving the supply line to the right.

 

[randomTPUreader]

Looks interesting. Even though AMD isn't targeting the high end they may have a real winner on their hands with RNDA4 if the performance is really there at a price that undercuts similar offerings from Nvidia. Even so I'll wait to see independently verified benchmarks and if they're even available before getting excited.

Personally, I'm not at all interested in video cards that cost anywhere close to $1000 or more. In my 33 years of computing the single most expensive GPU I've ever purchased in actual dollars is my current RX 6600 XT which was $320 in Aug 2022.

Even adjusting for inflation the single most expensive GPU I ever purchased was a Radeon DDR for $255 in Dec 2000 or about $465 in early 2025 dollars.

 

[Vayra86]

Right, but in simpler terms there is no relationship between die size and cost when comparing different nodes.

Another way of looking at this is, if Nvidia were to make a 1080 Ti GPU today on N5 it would probably cost exactly the same to make as it did originally. However, the die would likely be ~1/7th the original size.

Originally it was 470mm^2 @ 16nm so it would probably be ~70-100mm^2 (67mm^2 literally, but as you say some parts haven't shrunk as much as the ideal transistor) on the N5 node, or roughly the size of a modern A18 Apple SoC that is on TSMC N3E.

Die size + node = # transistors. Its the same thing, ballpark, at least, as die size and counting for inflation. After all, shrinks, like inflation, happens over time.

 

[mkppo]

Impressive they managed to cram that many transistors into 356mm2. Doesn't that surpass TSMC's quoted N4C peak density?

Also seems like it scales beyond 304w TDP. If power limits are fully uncapped, I think it's time to get one of the 3x8 pin variants and crank it all the way up.

9070 vanilla is priced like a turd

 

[RandallFlagg]

Die size + node = # transistors. Its the same thing, ballpark, at least, as die size and counting for inflation. After all, shrinks, like inflation, happens over time.

Nope, doesn't work like that anymore. It costs more $$ to make 100M transistors at 3nm than it did at 28nm.

TAnd we are putting a lot more transistors in a new GPU.

So as example, a 980 was on 28nm and had 5.2B transistors with a die of 398mm2. The 4080 was on N5 with 45.9M transistors and 378mm2. I would expect based on the chart, that the 4080 die cost about 8.8X more to make than the 980 die. There's no inflation adjustment there, and this is just for the GPU itself - not the other components board heat sink etc.

So it starts to make sense why the 4080 would have an MSRP of $1200, vs the 980's $550 in 2014.

Now for inflation, $550 in 2014 is $738 today. The rest is mostly going to be that GPU die.

Ref:

Chips aren't getting cheaper to make — the cost per transistor stopped dropping a decade ago at 28nm

Google says chips aren't getting cheaper.

www.tomshardware.com

 

[N/A]

Yet a 9070 with 54B Xtors costs 599 in 2025 lol. Nvidia has some explaining to do.

Impressive they managed to cram that many transistors into 356mm2. Doesn't that surpass TSMC's quoted N4C peak density?

For 50% logic. Rest SRAM and phy, when it has more logic that number shoots up.

But 7900XtX had the same 150 density with external cache something doesn't add up.

 

[Vayra86]

Nope, doesn't work like that anymore. It costs more $$ to make 100M transistors at 3nm than it did at 28nm.

TAnd we are putting a lot more transistors in a new GPU.

So as example, a 980 was on 28nm and had 5.2B transistors with a die of 398mm2. The 4080 was on N5 with 45.9M transistors and 378mm2. I would expect based on the chart, that the 4080 die cost about 8.8X more to make than the 980 die. There's no inflation adjustment there, and this is just for the GPU itself - not the other components board heat sink etc.

So it starts to make sense why the 4080 would have an MSRP of $1200, vs the 980's $550 in 2014.

Now for inflation, $550 in 2014 is $738 today. The rest is mostly going to be that GPU die.


View attachment 387416

Ref:

Chips aren't getting cheaper to make — the cost per transistor stopped dropping a decade ago at 28nm

Google says chips aren't getting cheaper.

www.tomshardware.com

I dont disagree on that either. But when looking at the subject and statement I was discussing, about 9070XT value compared to a 1070ti at a similar MSRP corrected for inflation and similar die size; this does mean the 9070XT has a reasonable price point given its die size. More than reasonable even.

 

[Bigshrimp]

I wonder if this will also be a paper launch. <scratches head>

 

[RandallFlagg]

I dont disagree on that either. But when looking at the subject and statement I was discussing, about 9070XT value compared to a 1070ti at a similar MSRP corrected for inflation and similar die size; this does mean the 9070XT has a reasonable price point given its die size. More than reasonable even.

I would agree the MSRP is reasonable. Keep in mind AMD and Nvidia have huge margins. AMDs latest non-GAAP gross margin is 53%, meaning that they make something for $0.47 and sell it for $1.00 This is all inclusive of all operating, R&D, marketing, transport, etc.

But that margin is almost certainly not on GPUs.

Their margins are extremely high on server CPUs. The Epyc 9755 for example, has 16 chiplets for a total combined die size of 1130mm^2. This is about 3X the size of a single 9070 die. It costs almost $13,000 though. An entire card 2/9070 XT MSRPs at $600, so 3 of them (same combined die size) are $1800. vs $13,000 for just a single 9755 Epyc server chip. There are some other costs with the server chip since it has chiplets, but no way is it that much.

So I don't think the market price for their GPUs will be reasonable, I don't think they'll make enough supply to keep the prices at MSRP. Not when they can make truckloads more money making server CPUs.

I have no reason to hope for that, just my cold pragmatic view. I'd like to get a 5070 or 9070 / 9070 XT class card, so I hope I'm wrong on the market pricing.

 

[Guwapo77]

Will buy to support. Keen to support Team Red

I'm doing the same thing. I was dead set on buying a 5090, but at this level of greed, I simply can not support. I know I won't get anything remotely close to the performance I want, but to hell with it team red for another 2 years or so.