News Cerebras files for IPO, shows rapid revenue growth and declining losses

[Admin]

Cerebras proposes IPO at unknown valuation as its revenue grows, and losses decline.

Cerebras files for IPO, shows rapid revenue growth and declining losses : Read more

 

[bit_user]

Relying on essentially one big customer, based in the UAE, a country which is under U.S. export restrictions, is a risk and Cerebras makes no secrets of that.

Forget the fact that it's a foreign customer - relying on a single customer is bad for other big reasons. Having multiple customers helps spread mindshare, creates a developer community, fosters collaboration on integrating support into open source projects and OS distros, and ultimately helps drive demand at other customers. So, I think it was short-sighted of them, if they failed to set aside a chunk of their production volume to use for courting other customers.

Speaking of their architecture more generally, a challenge I foresee them facing is the issue of SRAM-scaling. For good performance, they're very dependent on storing weights in the wafer, not in separate DRAM. While their compute logic will benefit from smaller manufacturing nodes, they won't get much more SRAM density. Perhaps they can address this by die-stacking like AMD did, but I'll bet it's a lot harder to do on a wafer-sized slab of silicon than on individual chiplets. Also, it could result in some of the tiles becoming defective, which I'm sure isn't an insurmountable problem, but not one you'd like to have.

Anyway, I think it'd be a lot of fun to program, though I doubt I'll ever have the chance. Good luck to them!

 

[gg83]

Forget the fact that it's a foreign customer - relying on a single customer is bad for other big reasons. Having multiple customers helps spread mindshare, creates a developer community, fosters collaboration on integrating support into open source projects and OS distros, and ultimately helps drive demand at other customers. So, I think it was short-sighted of them, if they failed to set aside a chunk of their production volume to use for courting other customers.

Speaking of their architecture more generally, a challenge I foresee them facing is the issue of SRAM-scaling. For good performance, they're very dependent on storing weights in the wafer, not in separate DRAM. While their compute logic will benefit from smaller manufacturing nodes, they won't get much more SRAM density. Perhaps they can address this by die-stacking like AMD did, but I'll bet it's a lot harder to do on a wafer-sized slab of silicon than on individual chiplets. Also, it could result in some of the tiles becoming defective, which I'm sure isn't an insurmountable problem, but not one you'd like to have.

Anyway, I think it'd be a lot of fun to program, though I doubt I'll ever have the chance. Good luck to them!

A 3d vcache over a wafer scale processor would be amazing! My using slightly mature/perfected nodes would make sense I think. I definitely wouldn't buy into the ipo though.