Intel Won't Bring Its Falcon Shores AI Chip To Market (techcrunch.com) 8
During the company's fourth-quarter earnings call Thursday, Intel co-CEO Michelle Johnston Holthaus announced that Intel has decided to cancel its Falcon Shores AI chip. Instead, it'll opt to use it as an internal test chip while shifting focus to Jaguar Shores for AI data center solutions. TechCrunch reports: "AI data center ... is an attractive market for us," Holthaus said during the call. "[B]ut I am not happy with where we are today. We're not yet participating in the cloud-based AI data center market in a meaningful way ... One of the immediate actions I have taken is to simplify our roadmap and concentrate our resources." The focus instead will be on Jaguar Shores, which Holthaus called Intel's opportunity to "develop a system-level solution at rack scale ... to address the AI data center more broadly."
Holthaus tempered expectations for Falcon Shores last month, when she implied that it was an "iterative" step over the company's previous dedicated AI data center chip, Gaudi 3. "One of the things that we've learned from Gaudi is, it's not enough to just deliver the silicon," Holthaus said during Thursday's earnings call. "Falcon Shores will help us in that process of working on the system, networking, memory -- all those component[s]. But what customers really want is that full-scale rack solution, and so we're able to get to that with Jaguar Shores."
"As I think about our AI opportunity, my focus is on the problems our customers are trying to solve, most notably the need to lower the cost and increase the efficiency of compute," Holthaus said. "As such, a one-size-fits-all approach will not work, and I can see clear opportunities to leverage our core assets in new ways to drive the most compelling total cost of ownership across the continuum."
Holthaus tempered expectations for Falcon Shores last month, when she implied that it was an "iterative" step over the company's previous dedicated AI data center chip, Gaudi 3. "One of the things that we've learned from Gaudi is, it's not enough to just deliver the silicon," Holthaus said during Thursday's earnings call. "Falcon Shores will help us in that process of working on the system, networking, memory -- all those component[s]. But what customers really want is that full-scale rack solution, and so we're able to get to that with Jaguar Shores."
"As I think about our AI opportunity, my focus is on the problems our customers are trying to solve, most notably the need to lower the cost and increase the efficiency of compute," Holthaus said. "As such, a one-size-fits-all approach will not work, and I can see clear opportunities to leverage our core assets in new ways to drive the most compelling total cost of ownership across the continuum."
Pretty clear what the problem is (Score:1)
The problem is... (Score:4, Informative)
... if you're not within spitting distance of NVidia on performance, people aren't going to use your chips even if you could give them away for free. GPUs are only a portion of server costs. Server costs are only a portion of datacentre costs. In addition to (minor) opex, you also have lots of power consumption. So any decrease in performance competitiveness really starts to stack up quickly against the selection of your product.
It's not worth releasing something that's not market-viable.
Umm, for free they would (Score:2)
If you could save some 10s of millions on hardware with the only cost a few percentage increase in training time I can gaurantee they'd be used. The cost vs efficiency/speed equation operates in most areas of life.
If intel could bring a chip to market that is maybe only 90% the speed of Nvidea GPUs yet costs 50% of the price then they'd be on to a winner.
This is not a repeat... (Score:2)
... of the iAPX 432, or the i960, or XScale, or Itanium, or Xeon Phi, or Arc GPUs, or ...
Seems strategically odd. (Score:2)
If their own press release says that they don't have the 'full-scale rack solution' that customers want I'm not going to doubt them on that point; just wonder why they aren't still offering it in smaller units: previous marketing material talked up 'compute density in x86 socket', which suggests a 500w or less part; and even if it's more than that th