Comment by brudgers
3 days ago
The market segmentation you describe makes a lot of sense to me. But I don't think the situation is a matter of under-investment and is instead just fundamental market economics.
Nvidia can afford to develop a comprehensive software platform for the compute market segment because it has a comprehensive share of that segment. AMD cannot afford it because it does not have the market share.
Or to put it another way, I assume that AMD's efforts are motivated rational economic behavior and it has not been economically rational to compete heavily with Nvidia in the compute segment.
AMD was able to buy ATI because ATI could not compete with Nvidia. So AMD's graphics business started out trailing Nvidia. AMD has had a viable graphics strategy without trying to beat Nvidia...which makes sense since the traditional opponent is Intel and the ATI purchase has allowed AMD to compete with them pretty well.
Finally, most of the call for AMD to develop a CUDA alternative is based on a desire for cheaper compute. That's not a good business venture to invest in against a dominate player because price sensitive customers are poor customers.
> Finally, most of the call for AMD to develop a CUDA alternative is based on a desire for cheaper compute. That's not a good business venture to invest in against a dominate player because price sensitive customers are poor customers.
Nvidia’s gross margins are 80% on compute GPUs, that is excessive and likely higher than what cocaine and heroin dealers have for gross margins. Real competition would be a good thing for everyone except Nvidia.
80% isn't a ridiculous margin if nobody else is selling the same compute. Software margins famously go much higher, to as much as 95% or high-nines for cloud products. People pay the price hand over fist, because there simply isn't equivalent hardware to compete with. It's practically a steal for certain HPC customers that want the latest and greatest out of TSMC.
I agree with both your comment and the parent comment - serious competition could spell the end for CUDA's dominance. But there will never be serious competition, CUDA has the head-start and their competitors threw in the towel with OpenCL. Khronos can't get Apple to sign onto a spec and they can't get AMD to change their architecture - open GPGPU compute is stuck in neutral while Nvidia is shifting into 6th gear. Reality is that Nvidia could charge cloud-level margins and get away with it, because Apple is the only other TSMC customer with equivalent leverage and they pretend the server market doesn't exist.
Competing on price when the entrenched incumbent has 80% margins does not sell me on the idea.
Or to put it another way, it would not be good for AMD.
Finally, most of the call for AMD to develop a CUDA alternative is based on a desire for cheaper compute. That's not a good business venture to invest in against a dominate player because price sensitive customers are poor customers.
This is such a key point. Everyone wants cheaper and cheaper compute - I want cheaper and cheaper compute. But not large-ish company wants to simply facilitate cheapness - they would a significant return on their investment and just making a commodity is generally not what they want. Back in the days of the PC clone, the clone makers were relatively tiny and so didn't have to worry about just serving the commodity market.
The demand for clones was also diffuse…the potential market included SMB and consumers and the market was exponentially expanding. The compute market is scaling linearly among a relative few established players who buy in bulk and have long B2B relationships with nVidia.