3 ms·

intel once tried to do this with larrabee[0] some years back. hopefully they learn from what went wrong there. [0] https://en.wikipedia.org/wiki/Larrabee_(mic

by abiox 9y ago

intel once tried to do this with larrabee[0] some years back. hopefully they learn from what went wrong there. [0] https://en.wikipedia.org/wiki/Larrabee_(microarchitecture) https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)

awalton 9y ago
Larrabee was an attempt to see if the x86 architecture could power a GPU. The answer was "not very likely", but it got turned into a product of its own anyways because it turned out to be very interesting for other compute-heavy use-cases. Larrabee's descendents became "Knights", which became the Xeon Phi product line. Keep in mind Intel currently builds GPUs - just of the integrated variety. What's new here is that Intel is deciding to build discrete (standalone, like those you'd plug into a PCIe port) GPUs.
- craftyguy 9y ago
  It was definitely being targeted to compete with other discrete graphics products, and at some point in the program they figured out that they would never meet the performance necessary to compere effectively. So in order to not have completely wasted several years of development, it was re-purposed as a product targeting HPC (the first generation Knights/Xeon Phi product)
  - awalton 9y ago
    Intel really doesn't mind "wasting" time on innovation - they make tens of billions of dollars a year and they're on top of the market. They can afford to go down blind avenues, especially when the research spills out so well, as it did in this case. It definitely wasn't a "saving throw" that Larrabee's architecture got repurposed. There were several teams at Intel working in similar directions - one team worked on a "cloud on a chip", one team worked on high bandwidth chip-to-chip interconnects, one team worked on on-chip networking... they all came together and formed the Knights Ferry research project, which then got turned into the Xeon Phi. The "core" of Larrabee, its quick little Pentium-derivatives, went on to be repurposed in the Quark product line and its lineage (e.g. the Intel PCH has a "Quark" inside). The 512-bit instruction set got parted out and became AVX512 in is various incarnations. They definitely got their money's worth out of Larrabee. Nobody is disagreeing with the fact that Larrabee didn't turn into a discrete GPU despite their attempts make it so. (It's also not surprising, seeing the carriage turn back into a pumpkin with Cell and other Many Core architectures fail to pan out to be good at graphics workloads). But that's a separate issue from Intel building GPUs, since they have a completely other team that works on building productized and shipped GPUs.

by abiox 9y ago

awalton 9y ago

Larrabee was an attempt to see if the x86 architecture could power a GPU. The answer was "not very likely", but it got turned into a product of its own anyways because it turned out to be very interesting for other compute-heavy use-cases. Larrabee's descendents became "Knights", which became the Xeon Phi product line. Keep in mind Intel currently builds GPUs - just of the integrated variety. What's new here is that Intel is deciding to build discrete (standalone, like those you'd plug into a PCIe port) GPUs.

craftyguy 9y ago
It was definitely being targeted to compete with other discrete graphics products, and at some point in the program they figured out that they would never meet the performance necessary to compere effectively. So in order to not have completely wasted several years of development, it was re-purposed as a product targeting HPC (the first generation Knights/Xeon Phi product)
- awalton 9y ago
  Intel really doesn't mind "wasting" time on innovation - they make tens of billions of dollars a year and they're on top of the market. They can afford to go down blind avenues, especially when the research spills out so well, as it did in this case. It definitely wasn't a "saving throw" that Larrabee's architecture got repurposed. There were several teams at Intel working in similar directions - one team worked on a "cloud on a chip", one team worked on high bandwidth chip-to-chip interconnects, one team worked on on-chip networking... they all came together and formed the Knights Ferry research project, which then got turned into the Xeon Phi. The "core" of Larrabee, its quick little Pentium-derivatives, went on to be repurposed in the Quark product line and its lineage (e.g. the Intel PCH has a "Quark" inside). The 512-bit instruction set got parted out and became AVX512 in is various incarnations. They definitely got their money's worth out of Larrabee. Nobody is disagreeing with the fact that Larrabee didn't turn into a discrete GPU despite their attempts make it so. (It's also not surprising, seeing the carriage turn back into a pumpkin with Cell and other Many Core architectures fail to pan out to be good at graphics workloads). But that's a separate issue from Intel building GPUs, since they have a completely other team that works on building productized and shipped GPUs.