News Tachyum announces 96-Core Prodigy-based AI workstation — $5,000 system also includes 1TB of memory

Status
Not open for further replies.
The fatal error Tachyum made was to try to be everything to everyone - a master of all trades. Had they focused on just a couple key niches, that might've been enough for them to have shipped product, by now, and got their legs under them. As it is, I'm sure the only way they're hanging on is by government subsidies and no-bid contracts.

Tachyum stated that a system powered by a single 96-core Prodigy processor with 1TB of RAM can run inference on a ChatGPT4 model with 1.7 trillion parameters, 'whereas it requires 52 Nvidia H100 GPUs to run the same thing at significantly higher cost and power consumption.'
The key detail he's glossing over is that running it on a Prodigy won't have nearly the same throughput as using H100's. He's really just relying on the memory capacity his CPU can support to establish that wild differential. But, if you just needed more memory capacity, it would seem to make much more sense to go with one of the x86 processors. With 256 cores, I expect the most cost effective option would currently be AMD's EPYC 9754 (2P).

As far as the Prodigy ATX Platform is concerned, we have some reasonable doubts about its economic viability for the company.
Thank you for doing this analysis. $5k sounded awfully cheap, to me.

This would be one kickstarter I wouldn't even be tempted by!
 
Last edited:
  • Like
Reactions: atomicWAR
The Duke 'Nuke-Em of processors?
I think you mean:

...which was first demo'd in 1998 and didn't ship until 2011.

I played the original Duke Nukem', which came out around the time Quake launched. Duke Nukem' was more advanced than Doom, but couldn't hold a candle to the graphics & gameplay in Quake. However, it was funny and irreverent enough to have still had an impact.
 
  • Like
Reactions: atomicWAR
I played the original Duke Nukem', which came out around the time Quake launched.
Duke Nukem 1 > 2 > 3D which I think is the one you mean. But I'll let you off :)

I remember the (initial) insane hype around Duke Nukem Forever. It was quite similar to Daikatana, and both games ended up having a lot in common: they both ended up being complete and utter trash. Good times.

As for this, to me it seems like it has hints of them trying to leap on the AI bandwagon just to drum up some kind of interest in the product (though I admittedly don't know enough of the details to be making any kind of assessment).
 
Cautiously interested.

Whenever a tech company uses the word "democratize" about its product, that's often a red flag.

But is the CPU any good ... as a CPU?
It's a proprietary ISA, so it's sort of a guess at this stage what that core count and clockspeed will translate to in actual practice, and whether software will be able to take full advantage of all the numbers they've thrown out. Doing a little searching, apparenly branch prediction is not looking to be as good as Intel/AMD/ARM and they've scaled up FP32 performance more than they have memory bandwidth, so there may be bottlenecks depending on what you do and how you do it.

...also a question whether people will write software to run on it, given it's status as vaporware. If you can't buy it and no one writes anything for it, it's not a good CPU it's just a good whitepaper.

Supposedly it will run unmodified X86, ARM, or RISC-V binaries at "similar to running native" through using a higher clockspeed to offset a predicted 30-40% translation overhead... but somehow that feels both dubious and unimpressive?
 
If it was real imagine one of these in each person's house sitting somewhere like a router. Doing things in the background.

Have an option of NO wifi/bluetooth/ethernet. For the people totally afraid of Skynet taking over. A small AI doing stuff for you and your home would be great.
 
Status
Not open for further replies.