3 ms·

A better solution would be for AMD to invest in bringing their ML stacks up to date to work with PyTorch and such.

modeless 5y ago
That would be nice. It really boggles the mind how thoroughly AMD has missed the boat on ML. And it really seems like they aren't on any trajectory to catch up even today. I've given up on them. Instead I'm hoping that Intel's imminent entry into the discrete GPU market does better. Nvidia is in desperate need of some competent competition in ML.
- qayxc 5y ago
  You seem to overlook the fact that NVIDIA poured immense resources into their ML software stack. This was back in AMD's Bulldozer days (2011), when the company struggled both financially and technologically. Meanwhile NVIDIA sponsored universities with graphics cards and had already developed their CUDA ecosystem (in 2007) when AMD was still busy with the ATI acquisition. In 2011 NVIDIA had an annual net income of about $500 million, while AMD had a net loss of $600 million at the same time and kept struggling for the following 5 years. In other words, NVIDIA already had an existing ecosystem of professional grade H/W accelerators and S/W infrastructure, when AMD was still a CPU manufacturer without a dedicated GPU division. When AMD acquired ATI, NVIDIA was already in the process of transitioning their GPGPU stack from data centre-only products to consumer hardware. AMD has powerful ML hardware today, but that's data centre and supercomputer only. They didn't miss the boat on ML - they were busy not drowning while NVIDIA was handing out goodies to academia.
  - modeless 5y ago
    It's been obvious for 10+ years now that AI was not only incredibly important but also a ginormous market opportunity up for grabs. To the extent that they were cash strapped they should have reprioritized. The effort would have paid for itself many times over. And that excuse no longer holds water now that AMD's market cap rivals Intel's and they're still not investing enough or making the right technical decisions. There must be deeper problems.
  - paulmd 5y ago
    Yeah, there was a lot of poor management from AMD though that led to them being in that situation, let's not pretend like it's just one day "oops we have no money". (And nor was it "all Intel's fault" either, AMD fucked up a lot in this time period, both technically and in their business decisions) At one point AMD was set to merge with NVIDIA but the board couldn't get over the sticking point of Jensen wanting to be CEO of the resulting company. Had the board swallowed their pride and let that happen, I doubt he would have led the company down the Bulldozer garden path. Instead AMD said no, and then way overpaid for ATI, which depleted their cash reserves and forced them to sell their foundries (which was ultimately probably the correct long-term move, but that isn't the proximate reason why they sold them - they were just out of cash at the time). Then they had no money and had to underfund their architectures, and move to "cost reduction" mechanisms in the design and implementation, and had a series of implementation problems and poor designs. Phenom was late (so much so that AMD had to put out a stopgap dual-socket "Quad FX" system just to try and compete against Core2 Quad) and Phenom had a major bug where part of the cache system had to be turned off, tanking performance by like >25%. And by the time Phenom II came out, AMD was putting non-SMT quadcores against Nehalem. By the time Bulldozer came out, the 8150 was going against Sandy Bridge, and by the time AMD fixed the worst of Bulldozer's performance problems, the 8350 was going against Ivy Bridge, and it still sucked anyway. AMD just executed extremely poorly in this period and a large part of that is the cash shortages resulting from buying ATI. The next decade of AMD's financial woes largely spring from that moment when the AMD board said "no" to the merger, and the chain of decisions and failures that resulted. And while consoles might have saved the company - they likely would not have been in that position in the first place had they said yes to the merger instead of emptying the bank account to buy ATI, instead of doing a merger-of-equals with NVIDIA. Also, AMD still isn't taking the steps that they can to increase their ML marketshare. They are actively reducing the support levels of their ROCm package - amateurs can easily get into the basics with a consumer NVIDIA gaming GPU while AMD forces you to buy a $5000 enterprise card.
catfishx 5y ago
Yes, it would also nice if they would implement stuff like CUDA (if thats not illegal) or provide alternatives, which is also used by many ML programs and forces you to buy a Nvidia card
- qayxc 5y ago
  They have ROCm and many popular frameworks support it. The problem is that NVIDIA has 100% of the mindshare and owns 97% of the dGPU market so it's a self-fullfilling prophecy that proprietary tech like CUDA will stay dominant in the foreseeable future. Hopefully Intel re-entering the dGPU market will put a dent in that and move the needle a little.
deathcakes 5y ago
They actually have already - PyTorch works straight up with ROCm, and so does Tensorflow. There is a little faffing to do to get it to work but they've made great progress in the last six months.
- zucker42 5y ago
  Except consumer GPUs don't officially support ROCm, despite consistent pressure from users for years. And there's no indication of when that situation will change. CUDA is successful because the same software works on low-powered laptop GPUs and expensive datacenter GPUs.

That would be nice. It really boggles the mind how thoroughly AMD has missed the boat on ML. And it really seems like they aren't on any trajectory to catch up even today. I've given up on them. Instead I'm hoping that Intel's imminent entry into the discrete GPU market does better. Nvidia is in desperate need of some competent competition in ML.

qayxc 5y ago
You seem to overlook the fact that NVIDIA poured immense resources into their ML software stack. This was back in AMD's Bulldozer days (2011), when the company struggled both financially and technologically. Meanwhile NVIDIA sponsored universities with graphics cards and had already developed their CUDA ecosystem (in 2007) when AMD was still busy with the ATI acquisition. In 2011 NVIDIA had an annual net income of about $500 million, while AMD had a net loss of $600 million at the same time and kept struggling for the following 5 years. In other words, NVIDIA already had an existing ecosystem of professional grade H/W accelerators and S/W infrastructure, when AMD was still a CPU manufacturer without a dedicated GPU division. When AMD acquired ATI, NVIDIA was already in the process of transitioning their GPGPU stack from data centre-only products to consumer hardware. AMD has powerful ML hardware today, but that's data centre and supercomputer only. They didn't miss the boat on ML - they were busy not drowning while NVIDIA was handing out goodies to academia.
- modeless 5y ago
  It's been obvious for 10+ years now that AI was not only incredibly important but also a ginormous market opportunity up for grabs. To the extent that they were cash strapped they should have reprioritized. The effort would have paid for itself many times over. And that excuse no longer holds water now that AMD's market cap rivals Intel's and they're still not investing enough or making the right technical decisions. There must be deeper problems.
- paulmd 5y ago
  Yeah, there was a lot of poor management from AMD though that led to them being in that situation, let's not pretend like it's just one day "oops we have no money". (And nor was it "all Intel's fault" either, AMD fucked up a lot in this time period, both technically and in their business decisions) At one point AMD was set to merge with NVIDIA but the board couldn't get over the sticking point of Jensen wanting to be CEO of the resulting company. Had the board swallowed their pride and let that happen, I doubt he would have led the company down the Bulldozer garden path. Instead AMD said no, and then way overpaid for ATI, which depleted their cash reserves and forced them to sell their foundries (which was ultimately probably the correct long-term move, but that isn't the proximate reason why they sold them - they were just out of cash at the time). Then they had no money and had to underfund their architectures, and move to "cost reduction" mechanisms in the design and implementation, and had a series of implementation problems and poor designs. Phenom was late (so much so that AMD had to put out a stopgap dual-socket "Quad FX" system just to try and compete against Core2 Quad) and Phenom had a major bug where part of the cache system had to be turned off, tanking performance by like >25%. And by the time Phenom II came out, AMD was putting non-SMT quadcores against Nehalem. By the time Bulldozer came out, the 8150 was going against Sandy Bridge, and by the time AMD fixed the worst of Bulldozer's performance problems, the 8350 was going against Ivy Bridge, and it still sucked anyway. AMD just executed extremely poorly in this period and a large part of that is the cash shortages resulting from buying ATI. The next decade of AMD's financial woes largely spring from that moment when the AMD board said "no" to the merger, and the chain of decisions and failures that resulted. And while consoles might have saved the company - they likely would not have been in that position in the first place had they said yes to the merger instead of emptying the bank account to buy ATI, instead of doing a merger-of-equals with NVIDIA. Also, AMD still isn't taking the steps that they can to increase their ML marketshare. They are actively reducing the support levels of their ROCm package - amateurs can easily get into the basics with a consumer NVIDIA gaming GPU while AMD forces you to buy a $5000 enterprise card.

catfishx 5y ago

Yes, it would also nice if they would implement stuff like CUDA (if thats not illegal) or provide alternatives, which is also used by many ML programs and forces you to buy a Nvidia card

qayxc 5y ago
They have ROCm and many popular frameworks support it. The problem is that NVIDIA has 100% of the mindshare and owns 97% of the dGPU market so it's a self-fullfilling prophecy that proprietary tech like CUDA will stay dominant in the foreseeable future. Hopefully Intel re-entering the dGPU market will put a dent in that and move the needle a little.

deathcakes 5y ago

They actually have already - PyTorch works straight up with ROCm, and so does Tensorflow. There is a little faffing to do to get it to work but they've made great progress in the last six months.

zucker42 5y ago
Except consumer GPUs don't officially support ROCm, despite consistent pressure from users for years. And there's no indication of when that situation will change. CUDA is successful because the same software works on low-powered laptop GPUs and expensive datacenter GPUs.