Thoughts about Tachyum

Tachyum.txt

Thoughts on Tachyum (8/7/2020)

Tachyum is a CPU startup that claims that they will
release a VLIW CPU for general purpose use[0], that
would also be faster than a GPU for even AI workloads.

The release of a production processor is planned according
to the company in 2021. However, the company itself says 
that they do not have the design working as an FPGA at
this stage, which is a major contradiction. [1]

For a CPU design that is claimed to be released next year,
there is no upstream software work done for it, which does
not match at all, especially for a server design. That alone
counts as a major mismatch with what is said.

Faster than a GPU at AI training and inference... with only
conventional DRAM. If the design itself has the CPUs keeping
up, the system as a whole won't because of DRAM bandwidth 
limitations alone. And that's just one of my doubts about it.

Claiming that they'll be faster than any future Arm and RISC-V
chips is very disingenious, especially the former...

"Tachyum customers consistently indicate that they would run 100%
 native applications within 9-18 months of transitioning to the
 Tachyum platform". No, that will not happen. I worked on porting
software to Arm for a while, and the world is just not like that.

If Tachyum is making something at this point, they should at
least put proper ISA documentation forward already for people
to start porting, and committing to have chips available for
testing in a reasonable timeframe.

Building a hardware platform with your own ISA is hard enough,
if you are going to handle porting the software ecosystem yourself
as a startup, that moves it to the borderline impossible.

[0]: That claim has been made countless times, however, to
my knowledge, I didn't see a device fullfilling those dreams,
with Intel and Nvidia coming closest, with much more resources.

[1]: https://www.hpcwire.com/off-the-wire/tachyum-shows-prodigy
-running-existing-x86-arm-and-risc-v-software/