Intel Thunder Bay Is Officially Canceled, Linux Driver Code To Be Removed

[edwaleni]

Laundry list:

Intel WiMAX (discontinued)
Intel 5G LTE radios (shut down)
Intel - McAfee Security products ($7 billion +, sold to TPG)
Intel Optane (JV with Micron, sold to Asian company)
Intel WiFi (Still going)
Intel Networks (still going)
Intel Maps (former Telmap)
Intel Interactive (former Omek)
Intel Autonomous Driving (former Mobileye)
Intel Graphics (Still going)
Intel RAN (JV with Ericsson)
Intel - Zhaoxin (JV with VIA)
Intel Scientific (supercomputing) Still going

[coder]

Much like they killed Xeon Phi.. I still think it was a cool accelerator, as it allowed you to run simulations whith different branches on the differents threads, not like GPUs which don´t have an instruction decode per "thread" (WARPs).

Yeah, I liked the notion that you could run a general-purpose OS on it. At the time, it was one of the highest core-count monolithic processors that could run Linux. And it had in-package memory (HMC, not HBM) before practically anyone else had done that with a CPU. It also pioneered the mesh interconnect Intel has been using for the bigger Xeons since Skylake SP.

Since you could run Linux and put it in a normal Xeon CPU socket, I imagine some people might've used them for more standard server apps, if they simply needed tons of CPU threads.

It would have competed with todays EPYCs in the HPC world, which are quite comparable,

No, the CPU cores in it were merely dual-issue Silvermont + SMT4 (and AVX-512). This was the biggest disappointment. I've seen others post complaints about how hard it was to actually get decent utilization of them.

In some ways, Sierra Forest could be its spiritual successor (except for the apparent lack of AVX-512). Especially, if they make a version with HBM.

and memory bandwidth.

Maybe, if you add together its HMC + 6-channel DDR4. HMC wasn't as high-performing as HBM, though. And most people probably just used it in L4 cache mode.

Xeon Phi was an intriguing product. I liked the idea of a massively multicore general-purpose processor, but it apparently fell far short of being competitive with actual GPUs. That's what ultimately sealed its fate and ushered in the current Xe-HPC era.

[Spacefish]

Much like they killed Xeon Phi.. I still think it was a cool accelerator, as it allowed you to run simulations whith different branches on the differents threads, not like GPUs which don´t have an instruction decode per "thread" (WARPs).

It would have competed with todays EPYCs in the HPC world, which are quite comparable, x86 with AVX-512, aprox. same Core count and memory bandwidth.

[brad0]

True, but the sad part is what it probably means for Movidius. They seemed like a promising little company. IIRC, they were about to release their second-gen products, when Intel bought them.

They should've ended up in the hands of someone more serious about building embedded SoCs towards the IoT end of the spectrum.

That's the way I felt about Intel's Bigfoot Networks acquisition.

[coder]

And another project to add to the "stumble and faceplant" list.

True, but the sad part is what it probably means for Movidius. They seemed like a promising little company. IIRC, they were about to release their second-gen products, when Intel bought them.

They should've ended up in the hands of someone more serious about building embedded SoCs towards the IoT end of the spectrum.

[Paradigm Shifter]

And another project to add to the "stumble and faceplant" list.

[coder]

Intel was marketing this for the datacenter, except that now all hyperscaler cloud operators make their own hardware. Anybody can slap an ARM core on a PCB.

Uh, it's not the ARM core that's the focus, but yeah. The basic architecture of an AI accelerator is straight-forward enough that just about everyone thinks they can do a half-decent job. I think too many underestimated how much work is involved in doing it well.

[coder]

Intel has way too many examples of squandered potential with IP acquisitions.

IMO, it's the typical Silicon Valley Venture Capitalist mindset, where they throw a bunch of stuff at the wall and see what sticks. And by "sticking", I mean they're looking for products which can pull $Billions in annual revenue, and are willing to accept a rather low success rate to find them.

That means it's not enough for a product merely to be profitable. It either has to have significant growth potential or be truly strategic. Otherwise, they cut their losses and move on. If they didn't see GPUs as strategic, I'm sure they'd have gotten cut by now.

[coder]

The whole Movidius thing was a pretty wasted potential especially under Intel. Those VPUs are pretty much general purpose computing cores - resembling similarities with GPUs but with much lower energy footprint. Yet they would be ideal to implement various sort of portable appliances - only if they would release the SDK and enable making own code for them. But no.. Intel choose to bundle a per-compiled AI specific software library that runs on the powerful VPU to cripple it to a totally useless brick for the all-mighty AI.

A funny thing is that Movidius didn't even start out as AI-focused, but rather got there through more conventional computer vision. The first product I heard of them being integrated into was Google's Project Tango tablet.

Fun fact: they originally used a SPARC core for control & coordination, due to its royalty-free status. Of course, the heavy-lifting is done by VLIW DSP-like cores working mainly out of on-chip SRAM.

What are they always afraid of? That it would be useful for some? And that same vibe goes over all useful products today.

I'm sure they simply don't want the type of support burden that comes with programmable products.

I can tell you, having used VLIW DSPs from TI, that it's incredibly frustrating when you're give a buggy and incomplete toolchain, with very limited tech support.

I'm not really siding with them, but I can see how they would be reluctant to open up something unless doing so fit into a broader strategy. Intel is no longer a simple chip company.

[stormcrow]

Intel has way too many examples of squandered potential with IP acquisitions.

Just about all industry behemoths do. They buy a technology base without management really understanding what it's about an what the market audience wants... They only see a temporary increase in their customer base. They incompetently integrate the team(s), either by pairing them with the wrong teams internally or losing them entirely as they mass resign. Then they either never release a product or release products no one asked for or wants. A little more flailing ensues, the brand name ends up being associated with something unrelated in the company, and the purchase is written off on taxes. Rinse, repeat.