Jim Keller, noted hardware engineer of AMD’s Athlon 64 and Zen architectures, is apparently up to no good again. After leaving AMD, he joined Tesla for a short time before starting his own company Tenstorrent. Once on his own he led development on two new processors based on RISC V, for what seems to be a general purpose CPU (Ascalon) and one ntended for AI (Tensix.)
This is quite interesting news to me. Has anyone had any experience with either of these new processors? Keller’s company is claiming performance previously unheard of for a RISC V design. Has anyone worked with any of these chips yet? Do these claims hold up?
Apparently a developer board called Atlantis will be available Real Soon Now ™. It seems that Tenstorrent has cloud access up and running so that people can ‘kick the tires.’
I know this doesn’t answer your question and it’s likely a marketing stunt, but a guy called “TechTechPotato” on youtube did a few interviews and coverages about Tenstorrent, thought you might get some value out of it.
Tech Tech Potato is a terrible name but Dr Ian Cutress who runs the channel is very well respected in the tech community. Used to work at Anandtech before going out on his own. He gets one on one interviews many YouTubers would only dream about.
I am not talking about him. I am talking about the naming of the chips. An American company naming their chip IP after a city in Isreal gives me vibes of a certain group of people in American that seem shady.
I hope with their processor called Atlantis it’s just a weird fable for ancient or mythological names.
It’s really great to see a company bringing RISC-V into the high end of the market.
I’ve not got any experience with Tenstorrent’s products I’m afraid. But I’d like to note that the criticisms of RISC-V I hear are very similar in nature to the criticisms I used to hear about Linux back in the 2000s.
Lots of “it lacks this specific feature”, or “it doesn’t do this as well as $competitor”.
All of which was true, but none of which mattered in the long run. Free is a price that scales insanely well. A “Good Enough” + “Free” is a proven market winner.
For every objection that traditional UNIX systems raised Linux either improved and gained the necessary feature, or the market changed and the feature was no longer important.
With the RISC-V RVA23 profile I think they’re on the cusp of hitting that sweet spot where the market’s issues have been addressed.
The compiler support, the software support, and the hardware are all “almost there” at the moment.
x86 has too much inertia to die anytime soon. But ARM have been annoying major customers with stupid lawsuits and announcing directly competing products. MIPS still exists, but nobody seems to care unless they have an existing investment in the platform.
IMO RISC-V is probably the future, and it looks like Tenstorrent will be a big part of that future.
I mean RISC-V is still in development, they only finalized the specification for the vector extension some time ago. The vector extension specification looks promising as it can handle variable length vectors too, and vector math is kinda important in the AI day and age and as we’ve seen with AVX can greatly enhance speed for certain types of workloads.
I feel like the competition might be a little greater with RISC-V since you can start implementing the specification without needing to pay licensing fees. I feel to this might be an advantage.
Though of course the RISC-V CPUs need to get into usable territory from the performance for desktops, notebooks and servers. That’s why this announcement looks good if the performance numbers match up.
Yes, the RV is most interesting and with the vector functions, it should also be able to do GPU/AI type workloads.
Its also the choice for China and EU each own decouple from US tech dominance, indl the US tech monopolies. EU has as running HPC program, some covert you software to RV, for cloud computing, so rack servers for hyperscalers, and also a vehicle platform. For EU it means we don’t have to do a GPU ourselves.
India also has some activity…
It could be interesting to see some hardware… Personally I am not going to replace US dependencies for Chinese, just yet, but the cheap prolific RV cores can be used for everything from MCUs and small routers, tvboxes, and will also be good choice for home servers, that no longer are running windows.
riscv is well established in microprocessors and almost all new graphics cards and storage devices run riscv cores inside them.
all the esp32 chips run riscv
literally the only place they havent taken over yet is high performance computing but that is only a matter of time.
I think what is going to happen is a ground up takeover like X86, it will start on the cheapest systems then slowly work its way up as it gains performance and features until it pushes out all the more expensive competition. ARM is going to be the biggest looser to begin with as their small cores are easiest to replace in embedded devices. amd and intel will be safe for a few years yet for high performance single thread performance, but almost everything else can be replaced easily with a bunch of cheap riscv cores.
I did say ARM had been annoying major customers with stupid lawsuits. If this deal goes through, then it will be the start of ARM entering the “find out” phase of the sequence…
More seriously, I’m not sure I’m all that keen on a Qualcomm acquisition. I mean have they actually done anything innovative lately? Maybe they have and I’ve been under a rock…
Hmm… given their proprietary nature on everything else, I am not convinced this will help RISC-V as Qualcomm is more interested in vendor lock-in mechanisms. See: mobile phones / modems.
If anything I could potentially see their involvement as a negative, in terms of ruining the more open ecosystem of RISC-V.
But also I have no idea what I am talking about. So it’s best to discard my uneducated opinion.
There needs to be a group made up by the big cpu makers to settle on a list of instructions that will be in every desktop chip but that will never happen so the next best thing is one dominant maker comes number 1 and everyone else follows so we at least get some kind of standard out of it like what intel done so “intel compatible” was the standard.
When it comes to things like radios (WiFi, bluetooth) Qualcomm only need to provide documentation and other support to their paying customers, making lock-in easy.
But general purpose CPUs need these pesky things called compilers. And compiler writers want documentation. In fact compiler writers need documentation and openness to the point where if they don’t have it, you don’t get a good compiler.
Which means software won’t run very well, if at all, on your CPU.
That means Qualcomm has a choice:
Spend 10 billion dollars and then spend even more becoming a very niche compiler vendor and hope for massive vertical sales
Provide documentation and errata, and cooperate with RISC-V International, so that the software ecosystem actually runs on their chips
It’s not a difficult decision.
I’m sure at some point Qualcomm will be found trying to wedge a patented technology into the RISC-V standards, but the other members will hopefully keep them in check. That, I hope, will be the worst of it.
There is such a group, it’s called the “x86 Ecosystem Advisory Group”, founded by Intel and AMD in October 2024.
For RISC-V, the RISC-V International non-profit company that owns the trademarks and steers the specs publishes what it calls Application Profiles. These define which subsets of instructions might be needed for a particular purpose (low power chip, desktop chip, etc).
RVA23 is probably the best known of them, but more will come in the future.
If Qualcomm does buy Tenstorrent, I’d expect them to join RISC-V International and start contributing to the future profiles. There’s no good reason for them not to.
