5 ms·

so basically they're ASICs? Would love some tech details, but it seems that the paper wont be published until 5pm today

monocasa 10y ago
https://drive.google.com/file/d/0Bx4hafXDDq2EMzRNcy1vSUxtcEk/view https://drive.google.com/file/d/0Bx4hafXDDq2EMzRNcy1vSUxtcEk...
- Cyph0n 10y ago
  I wonder why they didn't take the time to typeset it properly using, e.g., LaTeX. I'm hoping that the arXiv version will look more professional.
revelation 10y ago
That's like saying "so basically it's a chip". Well yes, it is a chip.
- Cyph0n 10y ago
  No, an integrated circuit (or chip) != an ASIC. An ASIC is a chip, but it is usually designed for a specific application, hence the name.
dom0 10y ago
Most accelerator-esque chips are often considered some kind of ASIC. CPUs are not ASICs because ??? ... never heard a good reason there. It's a fuzzy (but not fluffy) term.
- Cyph0n 10y ago
  Because CPUs are not application-specific? ASICs are typically designed from the ground-up to do a specific thing really well.
  - dom0 10y ago
    I'll bite: In which way is a computation engine (CPU [not ASIC], GPU [ASIC], Xeon Phi [ASIC]) more general purpose than a computation engine for AI [ASIC], GPU [ASIC] or Xeon Phi [ASIC, even though some of these are usable as a host processor?] For other parts it's even less clear: Flash or hard drive controllers are pretty normal micros with some dedicated hardware bolted on - clearly ASIC, but most of it was not designed for the "AS" part, and you could just ignore the flash and SATA interfaces and use it as a regular micro. So the distinction, if any, can't be about volume (since a lot of them are large volume parts), nor about functionality, but some fuzzy distinction by narrowness of intended use of the part (- but then again, GPUs). And how does it apply to other domains of chips? Is a TDA7000 an ASIC? :)
    - Cyph0n 10y ago
      Firstly, on what basis are the Xeon Phi and GPUs ASICs? Secondly, I'd argue that a GPU is definitely more general-purpose than the TPU. Just look at the block diagram shown in the linked paper: a GPU is orders of magnitude more complex than that! The added complexity is a result of a GPU having to support a wider variety of workloads. As for Flash controllers etc., if they incorporate a MCU or CPU, then they are simply not ASICs? I believe that the term ASIC itself is quite overloaded. From what I've seen, many people use the term ASIC to refer to any IC that is not reconfigurable (i.e., FPGA). Going by that definition, an ASIC is any circuit that is custom designed and fabbed on a wafer. Naturally, this would include a CPU, a GPU, and whatever else you can think of. The way I like to think of it is that an ASIC is a circuit designed to perform a specific task as efficiently as possible. Note that I used to word circuit; in other words, an ASIC could be part of some larger design. Some examples off the top of my head: - Digital camera CMOS sensor - Video decryption chip (e.g., in a cable box) - Active noise cancellation chip (if custom and not a DSP) - Full-custom TPM - Full-custom RSA-2048 engine - High-performance Ethernet switch controller - CPU cache controller (ASIC that is part of a CPU)
      - dom0 10y ago
        Counterpoint to complexity: Simple CPUs (scalar, in-order) are likely less complex than the TPU. E.g. here's a super-scalar in-order core: https://patentimages.storage.googleapis.com/US6311261B1/US06311261-20011030-D00011.png https://patentimages.storage.googleapis.com/US6311261B1/US06... > I believe that the term ASIC itself is quite overloaded. I think we're arguing the same thing, just from different angles :)
- MichaelGG 10y ago
  The AS is "Application Specific". CPUs are general, unless you wanna say a CPU is an ASIC for executing a general purpose instruction set as well as managing a bunch of system capabilities.
- kyboren 10y ago
  "ASIC" is usually used as a stand-in for "standard cell ASIC", meaning (with some hand waving) that the chip is implemented with suboptimal, general building blocks. In contrast, Intel CPUs are traditionally "full custom", where the chip is designed to be near-optimal for this design down to the transistor level. Full custom chips have the highest NRE costs and longest development cycles. Standard cell ASICs are less energy efficient, slower, and provide generally less logic density, but are cheaper and quicker to develop.

https://drive.google.com/file/d/0Bx4hafXDDq2EMzRNcy1vSUxtcEk/view https://drive.google.com/file/d/0Bx4hafXDDq2EMzRNcy1vSUxtcEk...

Cyph0n 10y ago
I wonder why they didn't take the time to typeset it properly using, e.g., LaTeX. I'm hoping that the arXiv version will look more professional.

revelation 10y ago

That's like saying "so basically it's a chip". Well yes, it is a chip.

Cyph0n 10y ago
No, an integrated circuit (or chip) != an ASIC. An ASIC is a chip, but it is usually designed for a specific application, hence the name.

dom0 10y ago

Most accelerator-esque chips are often considered some kind of ASIC. CPUs are not ASICs because ??? ... never heard a good reason there. It's a fuzzy (but not fluffy) term.

Cyph0n 10y ago
Because CPUs are not application-specific? ASICs are typically designed from the ground-up to do a specific thing really well.
- dom0 10y ago
  I'll bite: In which way is a computation engine (CPU [not ASIC], GPU [ASIC], Xeon Phi [ASIC]) more general purpose than a computation engine for AI [ASIC], GPU [ASIC] or Xeon Phi [ASIC, even though some of these are usable as a host processor?] For other parts it's even less clear: Flash or hard drive controllers are pretty normal micros with some dedicated hardware bolted on - clearly ASIC, but most of it was not designed for the "AS" part, and you could just ignore the flash and SATA interfaces and use it as a regular micro. So the distinction, if any, can't be about volume (since a lot of them are large volume parts), nor about functionality, but some fuzzy distinction by narrowness of intended use of the part (- but then again, GPUs). And how does it apply to other domains of chips? Is a TDA7000 an ASIC? :)
  - Cyph0n 10y ago
    Firstly, on what basis are the Xeon Phi and GPUs ASICs? Secondly, I'd argue that a GPU is definitely more general-purpose than the TPU. Just look at the block diagram shown in the linked paper: a GPU is orders of magnitude more complex than that! The added complexity is a result of a GPU having to support a wider variety of workloads. As for Flash controllers etc., if they incorporate a MCU or CPU, then they are simply not ASICs? I believe that the term ASIC itself is quite overloaded. From what I've seen, many people use the term ASIC to refer to any IC that is not reconfigurable (i.e., FPGA). Going by that definition, an ASIC is any circuit that is custom designed and fabbed on a wafer. Naturally, this would include a CPU, a GPU, and whatever else you can think of. The way I like to think of it is that an ASIC is a circuit designed to perform a specific task as efficiently as possible. Note that I used to word circuit; in other words, an ASIC could be part of some larger design. Some examples off the top of my head: - Digital camera CMOS sensor - Video decryption chip (e.g., in a cable box) - Active noise cancellation chip (if custom and not a DSP) - Full-custom TPM - Full-custom RSA-2048 engine - High-performance Ethernet switch controller - CPU cache controller (ASIC that is part of a CPU)
    - dom0 10y ago
      Counterpoint to complexity: Simple CPUs (scalar, in-order) are likely less complex than the TPU. E.g. here's a super-scalar in-order core: https://patentimages.storage.googleapis.com/US6311261B1/US06311261-20011030-D00011.png https://patentimages.storage.googleapis.com/US6311261B1/US06... > I believe that the term ASIC itself is quite overloaded. I think we're arguing the same thing, just from different angles :)
MichaelGG 10y ago
The AS is "Application Specific". CPUs are general, unless you wanna say a CPU is an ASIC for executing a general purpose instruction set as well as managing a bunch of system capabilities.
kyboren 10y ago
"ASIC" is usually used as a stand-in for "standard cell ASIC", meaning (with some hand waving) that the chip is implemented with suboptimal, general building blocks. In contrast, Intel CPUs are traditionally "full custom", where the chip is designed to be near-optimal for this design down to the transistor level. Full custom chips have the highest NRE costs and longest development cycles. Standard cell ASICs are less energy efficient, slower, and provide generally less logic density, but are cheaper and quicker to develop.