Tachyum has unveiled its 2nm Prodigy chips, which aim to offer up to 1024 cores and support super-fast DDR5 memory, and will compete against NVIDIA's Rubin Ultra.
Tachyum Upgrades Prodigy Chips To 2nm, Now Planned To Offer 1024 Cores, 6 GHz Clocks, 1 GB of Combined Cache, & Should Comfortably Tackle NVIDIA's Rubin Ultra (On Paper)
Tachyum's General-Purpose "Tachyum" chip project has faced several delays and several design revisions. Last time we heard about Prodigy, it was poised to feature 256 cores per chiplet and shipments by 2027. It was seen as a big upgrade from the previously disclosed 128 cores and 5.7 GHz clocks based on a 5nm process tech, which itself was upgraded to 192 cores, but never saw the light of day.
Today, Tachyum is announcing its 2nm Prodigy chips, another update to the Prodigy platform, which has gone without seeing any shipment, let alone a proper launch.
Starting with the specifications, the Prodigy 2nm from Tachyum is going to push things further with up to 1024 64-bit cores on a single socket. These cores will be operating at clock speeds of up to 6.0 GHz, and they can scale to 16-socket systems for up to 8,192 CPU cores (1024 core SKU supports 8-socket config).
Specs at a glance:
- 2nm Architecture (Yet to be Fabbed)
- Up To 1024 64-bit Cores
- Up To 6 GHz Clock Speeds
- Up To 1 GB of LLC
- Up To 1600W TDP
- Up To DDR5-17,600 MT/s Speed Support
- Up To 48 TB DDR5 Capacity Per Socket
- Up to 128 PCIe 7.0 Lanes
As per Tachyum, their 64-bit microarchitecture for Prodigy 2nm chips is going to support the latest matrix and vector extensions, tailored for high-performance AI and HPC use. It utilizes an Out-of-Order architecture with 8 instructions per clock.
The chip itself packs 128 KB of I-Cache, 64 KB of D-Cache (both with ECC), and 1 GB of L2+L3 cache. The SKUs range from 32, 64, 96, 128, 256, 320, 384, 448, 512, 768, 1024 core configs, with TDPs ranging from 30, 70, 140, 150, 300, 420, 550, 645, 800, 1000, and up to 1600W.
Moving over to platform details, the Prodigy 2nm chips will support up to 24 DDR5 channels with speeds of up to 17,600 MT/s, and up to 48 TB capacity per socket. I/O will include 128 PCIe 7.0 lanes and 64 PCIe controllers in total. Neither the DDR5-17600 spec nor PCIe 7.0 is common in the existing server market, so it is unlikely that the platform Tachyum is talking about today will be shipped by 2027, and even by 2030, it will be a miracle if they can get something like this out.
With the specifications out of the way, let's talk about some of the performance numbers that Tachyum is sharing. And right off the bat, Tachyum compares its Prodigy 2nm chips against NVIDIA's Rubin Ultra GPU platform, which is positioned for launch by 2027.
Tachyum claims that Prodigy 2 will be the first ever chip to exceed 1000 PFLOPs on inference, compared to NVIDIA Rubin's 50 PFLOPs.
They also claim that Prodigy Ultimate offers 21.3x higher AI rack performance versus NVIDIA Rubin Ultra (NVL756), while Prodigy Premium will offer 25.9x higher AI rack performance vs NVIDIA Rubin (NVL144). They don't go into much detail on what Prodigy Premium and Prodigy Ultimate are; they might be the different SKUs or entirely different chips.
All of this just goes to show that Tachyum has once again laid out some big plans for its Prodigy chips, but all of them are on paper, and the company is yet to even tape-out the Prodigy 2nm chips, though they claim they are ready to achieve this milestone thanks to a recent $220m investment. It looks like we will have to wait and see how much of this ends up being true and if Tachyum will really be able to deliver a 2nm chip on time, or if all of this is just big talk.
Follow Wccftech on Google to get more of our news coverage in your feeds.
