Among its outlandish claims, the company stated just one of its $23,000 processors could match the AI training performance of a highly sophisticated array of 52 Nvidia H200 GPUs — currently among the best GPUs out there. Banking that many GPUs alongside seven Supermicro GPU servers would cost $2,349,028, the company claimed, versus a single Prodigy pocket system with 2TB DDR5 DRAM.
Something to watch but seems crazy that one processor could match 52 of NVidia’s top of the line GPU’s
Andy
Very very unlikely. It’s some unheard of startup Tachyum. They are probably trying to get venture capital. But I simply do not believe the claim, at all. I doubt AMD does either, let alone Nvidia.
I visited their website;
Tachyum will unlock unprecedented performance, power efficiency, and cost advantages, to solve the most complex problems in big data analytics, deep learning, mobile, and large scale computing. The deep impact will be felt across a broad spectrum of markets and industries.
TachyumCustomer & Partner Portal
Currently for pre-approved partners only.
To access ourTachyumCustomer & Partner Portal and license use of our emulation software, please get in touch with us first at contactus.sk@tachyum.com and tell us about your company and use case.
No products, no customers.
The Captain
Sounds like a chance to get some investment dollars on a company before it explodes however I can’t seem to find a symbol. Hope they aren’t privately held…doc
Yes Doc they are private and their last funding round was around 24 million so very small company.
Andy
So, from the Tachyum website their high-end Prodigy processor boasts 24 AI Petaflops. This is a single socket which can be configured in racks of 256 to get 6 Exaflops per rack, and they have designs with 64 such racks.
How does this compare to NVIDIA’s and AMD’s best products right now?
From https://www.tachyum.com/datasheets/Prodigy%20PB%2016192%20v1.01_231202.pdf:
When Tachyum is actually shipping product that’s in customers’ hands and delivering effective pilots we can know whether they’re a factor.
No tapeout yet. Claims about running various forms of code… on their emulators.
Maybe it’s vapor, maybe it’s Maybelline
Sigh! I don’t care how fast it can do arithmetic in the CPU, even how fast it is at doing saxpy (or daxpy). I want to know how fast it can move data into and results out of the CPU. Well, I guess if the math is very slow compared to data motion… Um, then I don’t care either.
Given these results are due to simulations, I suspect they are in for a rude awakening when they run software on real hardware. Main memory is slow. Even L1 or L2 cache is slow compared to the hardware registers. The trick to writing fast low-level math software is to hide the latency of memory fetches as much as possible. True, there is software which has lots of CPU operations per data fetch. But such software, other than when testing simulators 
runs so fast that you’ don’t care.
emphasized text[quote=“eachus, post:8, topic:101116”]
Sigh! I don’t care how fast it can do arithmetic in the CPU, even how fast it is at doing saxpy (or daxpy). I want to know how fast it can move data into and results out of the CPU. Well, I guess if the math is very slow compared to data motion… Um, then I don’t care either.
[/quote]
This is an old problem. Vendors like to highlight magic numbers but semiconductor circuits are so complex that all that really counts are benchmark tests, how fast/well can they do THIS JOB?
The Captain
More skepticism regarding Tachyums delivery date and on delivery price :
Extraordinary claims require extraordinary evidence…
