But i still don't quite understand what's new about rdna 4 "out of order memory". Could you go into more detail about this rdna 4 feature in the rx 9070 article?
I hope they are fixing the front end bottlenecks (you can see it in GPA analysis of some workloads). Without fixing that it doesn’t make sense to make bigger GPU configs. Also I wonder what clocking improvements come with 18A.
Yeap, please do some workload based analysis of RDNA4. Even though AMD says 4096ALUs …. It’s actually 8192ALUs because half of them can do both INT and FP32. They are very much like NV previous generation here.
It would be helpful if you could explain what happens in games like Indiana Jones, where the 9070 XT drops to the level of the 4060 Ti. Architecturally, there’s nothing that justifies this. Nvidia's snake oil in action?
Excellent! Now let's hope for some nice Celestial desktop products coming out, putting on more credible pressure on the other remaining desktop GPU vendors to get to actually reasonable innovation and pricing again.
wonder if Intel is really selling their B580 and 570 at a loss? Their first dGPU generation (Alchemist) certainly lost them money, but is it also true for Battlemage? That generation has good uptake and is selling well, unlike how Alchemist did. I have no problem believing that Intel is not making Nvidia-like margins on their current dGPUs, but not having profit margins above 50% per unit (like Nvidia) isn't the same as selling at a loss.
Intel is not being very efficient in this market and appears to be operating at a loss by positioning a mid-range 5nm chip with a die size of 270mm²(5070's 263mm²) in a bid to compete in the low-end market segment, 4060 is only 159mm², 7600XT 200mm² on cheap 6nm. The development of the Battlemage on TSMC's 5nm node incurred a significant cost of nearly half a billion dollars, a factor that must be considered when evaluating the profitability of this venture. Despite these efforts, Intel has only managed to sell a few hundred thousand units so far.
Each chip is estimated to cost TSMC approximately $90–$100 for production. When factoring in Intel's gross margin (Let's say, approximately $60), the chips are sold to AIBs for around $150. Additional costs, such as memory, thermal solutions, PCB, assembly, and more, further increase the price. Once AIBs add a 10–15% profit margin, followed by retailers adding another 10–20%, it becomes evident that there is little room for profitability in this price-sensitive low-end market.
With only 500,000 chips sold(They sold way less), Intel has generated an estimated revenue of $30 million, then approximately 6% of its initial development cost. For Intel to break even, it would need to sell an estimated 7–8 million units, which equates to nearly a third of the current discrete GPU market. Achieving such a volume is highly unlikely, making profitability virtually unattainable under the current circumstances.
Yeah, I think most people only think about the cost/wafer, as if the entire development of the architecture and specific chip were free. The mask and validation alone can cost tens of millions. lol
While I agree on TSMC needing competition, which is partially up to Intel too, both AMD and especially Nvidia have high margins on their GPU products, happily stagnating on progress and up-pricing/-naming performance tiers - us consumers would absolutely benefit from more competition.
I'm too tired to explain in detail, but AMD's net profit margins on consumer GPUs is very thin, hovering around 10%
Making money with GPUs is much more complicated than with CPUs. There are a lot of parties trying to make a profit along the way, and these companies need to pay massive support teams over the long term.
Great article!
But i still don't quite understand what's new about rdna 4 "out of order memory". Could you go into more detail about this rdna 4 feature in the rx 9070 article?
Thank you very much!
Great stuff, wonder whether we'll get a dGPU equivalent of Xe³.
I hope they are fixing the front end bottlenecks (you can see it in GPA analysis of some workloads). Without fixing that it doesn’t make sense to make bigger GPU configs. Also I wonder what clocking improvements come with 18A.
Yeap, please do some workload based analysis of RDNA4. Even though AMD says 4096ALUs …. It’s actually 8192ALUs because half of them can do both INT and FP32. They are very much like NV previous generation here.
More threads, finer grained register blocks, more scoreboard tokens, and the scalar register all designed to address front-end problems, right?
NVIDIA had "STOC" a.k.a. "opacity micromaps" since Ada. Maybe worth a mention.
Interesting. Recently it got added to the DXR 1.2 standard, so I guess it's picking up. Fun to see raytracing techniques spread
It would be helpful if you could explain what happens in games like Indiana Jones, where the 9070 XT drops to the level of the 4060 Ti. Architecturally, there’s nothing that justifies this. Nvidia's snake oil in action?
Excellent! Now let's hope for some nice Celestial desktop products coming out, putting on more credible pressure on the other remaining desktop GPU vendors to get to actually reasonable innovation and pricing again.
Intel is not going to move the price pointer anywhere. Battlemage was just being sold at a loss.
It's TSMC that needs competition
While I have seen this stated quite frequently, I
wonder if Intel is really selling their B580 and 570 at a loss? Their first dGPU generation (Alchemist) certainly lost them money, but is it also true for Battlemage? That generation has good uptake and is selling well, unlike how Alchemist did. I have no problem believing that Intel is not making Nvidia-like margins on their current dGPUs, but not having profit margins above 50% per unit (like Nvidia) isn't the same as selling at a loss.
Intel is not being very efficient in this market and appears to be operating at a loss by positioning a mid-range 5nm chip with a die size of 270mm²(5070's 263mm²) in a bid to compete in the low-end market segment, 4060 is only 159mm², 7600XT 200mm² on cheap 6nm. The development of the Battlemage on TSMC's 5nm node incurred a significant cost of nearly half a billion dollars, a factor that must be considered when evaluating the profitability of this venture. Despite these efforts, Intel has only managed to sell a few hundred thousand units so far.
Each chip is estimated to cost TSMC approximately $90–$100 for production. When factoring in Intel's gross margin (Let's say, approximately $60), the chips are sold to AIBs for around $150. Additional costs, such as memory, thermal solutions, PCB, assembly, and more, further increase the price. Once AIBs add a 10–15% profit margin, followed by retailers adding another 10–20%, it becomes evident that there is little room for profitability in this price-sensitive low-end market.
With only 500,000 chips sold(They sold way less), Intel has generated an estimated revenue of $30 million, then approximately 6% of its initial development cost. For Intel to break even, it would need to sell an estimated 7–8 million units, which equates to nearly a third of the current discrete GPU market. Achieving such a volume is highly unlikely, making profitability virtually unattainable under the current circumstances.
Yeah, I think most people only think about the cost/wafer, as if the entire development of the architecture and specific chip were free. The mask and validation alone can cost tens of millions. lol
While I agree on TSMC needing competition, which is partially up to Intel too, both AMD and especially Nvidia have high margins on their GPU products, happily stagnating on progress and up-pricing/-naming performance tiers - us consumers would absolutely benefit from more competition.
I'm too tired to explain in detail, but AMD's net profit margins on consumer GPUs is very thin, hovering around 10%
Making money with GPUs is much more complicated than with CPUs. There are a lot of parties trying to make a profit along the way, and these companies need to pay massive support teams over the long term.