News In a bid to compete with Nvidia, Jeff Bezos and Samsung invest $700 million in AI chip startup Tenstorrent

[Admin]

The new investment will help Tenstorrent to build an engineering team, invest in its global supply chain and build large artificial intelligence training servers.

In a bid to compete with Nvidia, Jeff Bezos and Samsung invest $700 million in AI chip startup Tenstorrent : Read more

 

[bit_user]

Congrats to Tenstorrent!

I'm a bit surprised Altman/OpenAI didn't snap them up, but instead worked with Broadcom on their AI accelerator. I don't actually know anything about Broadcom's solution, but what I've read about Tenstorrent's reminds me a fair bit of Cerebras' WSE, in that they're both tile-oriented dataflow architectures.

I think the article grossly overemphasizes the RISC-V aspect. It seems to me like a plan B for the business, in case their AI gambit doesn't work out. Their Tensix cores, at least in all the generations I've so far read about, aren't RISC-V, but much more DSP-like. Their original motivation for building RISC-V cores was indeed to use in their products, but mostly at the periphery and not as the main compute workhorses. Then, they (or their investors) probably thought they could keep developing the RISC-V cores and potentially license them, as an alternate revenue stream.

 

[DS426]

...

I think the article grossly overemphasizes the RISC-V aspect. It seems to me like a plan B for the business, in case their AI gambit doesn't work out. Their Tensix cores, at least in all the generations I've so far read about, aren't RISC-V, but much more DSP-like. Their original motivation for building RISC-V cores was indeed to use in their products, but mostly at the periphery and not as the main compute workhorses. Then, they (or their investors) probably thought they could keep developing the RISC-V cores and potentially license them, as an alternate revenue stream.

It also depends on how much you think Jim Keller is a believer in RISC-V. For example, if he really does think it will attract great engineering and other talent, that could make it worthwhile as otherwise nVidia and the other top players will continue to snap up the bulk of great talent in the world.

Arm zoomed into the datacenter world while RISC-V is still a relatively small player, so it kind of goes back to the problem of: build it and they will come, or let the demand and ecosystems become sufficiently solid before trying to make a big splash?

 

[bit_user]

It also depends on how much you think Jim Keller is a believer in RISC-V. For example, if he really does think it will attract great engineering and other talent, that could make it worthwhile as otherwise nVidia and the other top players will continue to snap up the bulk of great talent in the world.

What he's previously said about it is that the ISA has the right features to enable an efficient implementation and that he appreciates being able to add custom extensions to support their more proprietary hardware blocks. The icing on the cake is probably the lack of royalties for an architecture license. This makes it perfect for embedded applications, like what they were doing with it.

I have no idea about the "attracting great engineering talent" aspect. In my opinion, what attracts great talent is a great team and good prospects for the business. Right now, I think they probably don't have much trouble finding the talent they need.

Arm zoomed into the datacenter world

I wouldn't call it a zoom. It took many years before they achieved significant scale. Entire teams and companies intent on building ARM server CPUs have come & gone, in the time it's taken ARM to achieve a substantial foothold in that market.

while RISC-V is still a relatively small player, so it kind of goes back to the problem of: build it and they will come, or let the demand and ecosystems become sufficiently solid before trying to make a big splash?

It will happen for geopolitical reasons, if not others. So, the trick is to straddle the embedded and server worlds so that you're positioned to ride the wave when RISC-V gets embraced in datacenters, but not betting your entire business on it by a specific point in time. RISC-V has already achieved scale in the embedded market and there's money to be made there.

 

[gg83]

Intel Foundry wasn't mentioned. I think there is a weird conspiracy going on with Intel. Something needs to happen with A18 or we will be ten years behind TSMC. It is the Crown Jewel of the Taiwanese government and the US needs that with Intel Foundry.

 

[bit_user]

Intel Foundry wasn't mentioned. I think there is a weird conspiracy going on with Intel.

Jim Keller seems to have been in TSMC's orbit, for a while. My guess is that everyone at Tenstorrent has experience dealing with TSMC, while nobody there used IFS (because it's basically a new thing). Given their relatively meager initial volumes, they probably view it as not being worth the risk and learning curve to try and use IFS vs. going with the known quantity that is TSMC.

Startups often take a somewhat counterintuitive view towards risks. As hugely risky endeavors, I think the good ones try to be very selective about where the business takes risks, and then go with tried-and-true in every other case. The only gambles they tend to make are on things fundamental to the business plan. As far as that goes, if IFS only offers them a relatively small savings on manufacturing costs, it's not a key enabler for their business plan. However, if they suffer major delays because of differences between IFS and TSMC or because IFS fails to deliver on its promises, that could easily sink Tenstorrent.

So, when Tenstorrent becomes big and successful, has a good revenue ramp, established customer base, and substantial production volumes, then I think they'll start to worry about cost optimizations and take a hard look at IFS. Right now, I wouldn't do it in their shoes.

Something needs to happen with A18 or we will be ten years behind TSMC.

Intel already has big customers lined up for it, including themselves. Any volume Tenstorrent can bring them isn't going to make much difference, I think. Intel needs to deliver on its promises. Once they've done that, for a few generations, then I think they'll find more customers lining up at their doors.