3 ms·

Do silicon vendors give internal teams room for innovation experiments, beyond competition for markets defined by Apple supply chains? Even if there's no innova

Do silicon vendors give internal teams room for innovation experiments, beyond competition for markets defined by Apple supply chains? Even if there's no innovation on use cases, how about extending some products sideways, e.g. x86+FPGA version of Zynq, https://www.amd.com/en/products/adaptive-socs-and-fpgas/soc/zynq-7000.html https://www.amd.com/en/products/adaptive-socs-and-fpgas/soc/... > AMD Zynq 7000 SoC family integrates the software programmability of an Arm-based processor with the hardware programmability of an FPGA, enabling key analytics and hardware acceleration while integrating CPU, DSP, ASSP, and mixed signal functionality on a single device. There are small manufacturers [1][2] bending AMD "mobile" SoCs into a plethora of form factors. Now that we have OSS FPGA toolchains (LiteX) and modular PC enclosures (Framework), AMD can offer flexible SoCs for experiments that unlock new markets that can grow in volume, instead of me-too chasing Apple/Qualcomm. [1] https://cwwk.net https://cwwk.net [2] https://aoostar.com https://aoostar.com

jauntywundrkind 2y ago
The fpga market is sad as hell man. There seems to be a gentleman's agreement to not let the field advance. Zynq is from 2011 and has barely been revved. The FPGA market has climbed ever more up-market, building only ever bigger better but at exponentially vaster stratospheric price points. There's a couple mid-tier Lattice parts that have emerged, and Gowin has some real disruption from below just starting, but big playerskke Xilinx/AMD and Altera/Intel (since re-spun off) have made FPGAs ever more unattractive & unaffordable year by year. It seems very unlikely that the FPGA people actually want their parts accessible useful or commoditied. There is a very strong demand for FPGAs that justifies their multi-thousand dollar price points for ever pretty modest sized chips (which can be scaled up to huge house-priced sizes too); nothing else gives you what fpgas do. There's no substitute. It would be too risky for these folks to actually allow reconfigurable computing to be in reach. Some of the SmartNIC offerings that hyperscalers and hyperscalers alone use have nice but kind of specialized FPGAs onboard. Marvell/Nvidias's Bluefield rocks. But again it's a $500+ network card with capabilities most folks won't be able to use well. It's especially damned a proposition because there's no solutions available for use & the necessary AMD Vivado software needed to program the fpga is thousands of dollars and that's the starting cost before you start signing contracts to get access to the dram controllers or to Ethernet phy. Every now and then there's something that seems like a ray of hope in the gloom - for ex AMD's nanotube ebpf -> fpga converter (https://www.phoronix.com/news/AMD-Xilinx-Nanotube-Compiler https://www.phoronix.com/news/AMD-Xilinx-Nanotube-Compiler) - but you need that expensive closed tool chain to start and the efforts usually seem quickly abandoned after a code drop or two, no real effort to make it happen. It's hellworld. It's so sad to see an industry so near to the rest of computing and the pax Romana of "PC compatible" interoperability & mutual growth, but so unable to cross the chasm & flourish, just in such a closed off isolated locked-in-its-head niche darkworld. I get it, there's no evidence the volume isn't there so chasing whales is what you have to do. FPGAs are such amazing & interesting systems, but it's been a terrible decade of them getting nowhere. Zynq is an ancient relic reminding us of hope lost. I feel like the best case possibility is that this NPU AI thing maybe breaths some life in; XDNA is supposedly a somewhat special small FPGA, and that maybe possibly over time could help bring FPGAs out of the dark & into the light of day. One could hope.
- transpute 2y ago
  > FPGAs are such amazing & interesting systems, but it's been a terrible decade of them getting nowhere.. best case possibility is that this NPU AI thing maybe breaths some life Cut to sci-fi film as engineers in 4024 unearth a colossal landfill of 2024 AI "GPUs". They place one "GPU" into an additive fabrication device for automated scanning, interface construction, electrical and software analysis. Later, their AI announces that it has rewritten the "GPU" firmware, making the underlying hardware usable for general-purpose digital logic.
- duskwuff 2y ago
  > Zynq is from 2011 and has barely been revved. What do you mean? There's a bunch of newer Zynq parts in the UltraScale and UltraScale+ lines, as well as Versal - if you're only looking at the Zynq 7000 series, that's why you aren't seeing anything.

jauntywundrkind 2y ago

The fpga market is sad as hell man. There seems to be a gentleman's agreement to not let the field advance. Zynq is from 2011 and has barely been revved. The FPGA market has climbed ever more up-market, building only ever bigger better but at exponentially vaster stratospheric price points. There's a couple mid-tier Lattice parts that have emerged, and Gowin has some real disruption from below just starting, but big playerskke Xilinx/AMD and Altera/Intel (since re-spun off) have made FPGAs ever more unattractive & unaffordable year by year. It seems very unlikely that the FPGA people actually want their parts accessible useful or commoditied. There is a very strong demand for FPGAs that justifies their multi-thousand dollar price points for ever pretty modest sized chips (which can be scaled up to huge house-priced sizes too); nothing else gives you what fpgas do. There's no substitute. It would be too risky for these folks to actually allow reconfigurable computing to be in reach. Some of the SmartNIC offerings that hyperscalers and hyperscalers alone use have nice but kind of specialized FPGAs onboard. Marvell/Nvidias's Bluefield rocks. But again it's a $500+ network card with capabilities most folks won't be able to use well. It's especially damned a proposition because there's no solutions available for use & the necessary AMD Vivado software needed to program the fpga is thousands of dollars and that's the starting cost before you start signing contracts to get access to the dram controllers or to Ethernet phy. Every now and then there's something that seems like a ray of hope in the gloom - for ex AMD's nanotube ebpf -> fpga converter (https://www.phoronix.com/news/AMD-Xilinx-Nanotube-Compiler https://www.phoronix.com/news/AMD-Xilinx-Nanotube-Compiler) - but you need that expensive closed tool chain to start and the efforts usually seem quickly abandoned after a code drop or two, no real effort to make it happen. It's hellworld. It's so sad to see an industry so near to the rest of computing and the pax Romana of "PC compatible" interoperability & mutual growth, but so unable to cross the chasm & flourish, just in such a closed off isolated locked-in-its-head niche darkworld. I get it, there's no evidence the volume isn't there so chasing whales is what you have to do. FPGAs are such amazing & interesting systems, but it's been a terrible decade of them getting nowhere. Zynq is an ancient relic reminding us of hope lost. I feel like the best case possibility is that this NPU AI thing maybe breaths some life in; XDNA is supposedly a somewhat special small FPGA, and that maybe possibly over time could help bring FPGAs out of the dark & into the light of day. One could hope.

transpute 2y ago
> FPGAs are such amazing & interesting systems, but it's been a terrible decade of them getting nowhere.. best case possibility is that this NPU AI thing maybe breaths some life Cut to sci-fi film as engineers in 4024 unearth a colossal landfill of 2024 AI "GPUs". They place one "GPU" into an additive fabrication device for automated scanning, interface construction, electrical and software analysis. Later, their AI announces that it has rewritten the "GPU" firmware, making the underlying hardware usable for general-purpose digital logic.
duskwuff 2y ago
> Zynq is from 2011 and has barely been revved. What do you mean? There's a bunch of newer Zynq parts in the UltraScale and UltraScale+ lines, as well as Versal - if you're only looking at the Zynq 7000 series, that's why you aren't seeing anything.