More comments from grxbstrd on Seeking Alpha:
"I think there’s room for many players in these exponentially growing markets but agree NVIDIA will probably be a prime beneficiary."
I've been following this market like a hawk with a sharp focus on competition since GPGPU started becoming real ~2014. This is opposed to the R&D experiments GPGPUs were ~2006-2013. Many challengers have made big claims, many have come, many have gone. Notable: TPU, Nervana, and Graphcore. TPU has been through 4 revisions, the last one Google didn't even bother to post performance numbers. This is after comparing their gen2 TPU to Maxwell generation GPUs and claiming orders of magnitudes faster and generating lots of PR around that idea. Billions spent and deployed and the scrappy GPUs out innovated them. Nervana was Intel's startup purchase, another ASIC that also was going to put Nvidia out of business. The founder bought and posted his new Ferrari and how they were going to destroy GPUs. Long story short, Nervana who? Intel then bought Habana Labs, another ASIC at 5X the cost of Nirvana, but father along. I was at the public unveiling when Habana came out of stealth mode in 2018 and bragged about destroying GPUs ( seekingalpha.com/... ). Now they are positioned as the low price leader at AWS for inference their sole design win. I wonder if anyone is using them? A giant disappointment for Intel. And Graphcore is my fav, also big claims and big marketing pushes about killing GPUs and orders of magnitude faster blah blah blah. Where is Graphcore today? Struggling for traction. Billions in valutions, but last I heard the CEO pushed out their IPO from "soon" to "a few more years". Cerebras, SambaNova, Wave Computing, Thin CI, and many others. They all have one thing in common: where are they? Yes a few of these guys have sold some systems, those are experiments, not production machines. The software and development work is obviously challenging as evidenced by the lack of startup results in MLPerf, the industry standard AI benchmark suite endorsed by most of these guys.
Nvidia has faced some of the biggest challengers the industry has to offer and from what I can tell, hasn't lost a scrap of market share. The barriers to entry are enormous. Nvidia themselves claim 90% of DC accelerator share, and BofA semi analyst Vivik Arya 96% IIRC.
Will AMD change that? Very doubtful, if they haven't by now they probably aren't ever going to.
Will Intel challenge in DC Acceleraters? Remains to be seen. I defer an opinion at this point until we see their offerings in public supposedly early next year. The biggest issue, as all these others have found, is the software stack. Optimizing compilers for an AI framework can take years. If you line that up against the idea Intel saw NVDA coming in 2007 and designed and then killed their own parallel processor, Larrabee, it doesn't look too good for them to now, finally, 14 years later, pull a rabbit out of the hat.
So yes, there is plenty of room for other players in a very large and growing market. Whether someone can gain a toe hold and turn it into real traction is a great big giant open question. Because to date there is just basically one platform, as you state, and one prime beneficiary (who isn't slowing). I'm still trying to identify a meaningful #2. cheers to you.
grxbstrd |
