Be skeptical: Tachyum's Universal Processor builds huge castles in the air


Tachyum has been wanting to build the Über processor for over four years. But the new announcements are also based on simulations and no tangible hardware. The targeted specifications are more than impressive, but that's what they were two years ago and whether the practice corresponds to the simulated theory remains to be seen.

Too good to be true?

If something sounds too good to be true, it often is. Two years ago, Tachyum introduced a 64-core processor that was supposed to have a DDR4/DDR5 memory interface with eight channels, up to 72 PCIe 5.0 lanes and also natively up to dual 400 Gigabit Ethernet . Various smaller variants should be able to hold a candle to almost every current Intel product on the market, because each of the CPU cores should be faster than that of a current Xeon, but also smaller than an ARM core. The product should appear in 2021. But it isn't.

Now Tachyum follows up with another announcement. It now reports up to 128 cores in various variants with clock rates of up to 5.7 GHz, a 16-channel DDR5-7200 memory interface and again many PCIe 5.0 lanes. All other key data have also grown enormously, TSMC is to take over series production in N5. And the competition?

The new variant in the largest expansion stage is said to be four times faster than the best Intel Xeon, three times faster than all AMD Epyc – and the company even wants to outperform Nvidia's new Hopper GPU, optionally being three times or even six times faster. In every area, you want to be ahead of every possible competitor on paper. Sounds like a jack of all trades. But as we all know, there is no such thing.

Among the highlights of the newly launched Prodigy processor are:

  • 128 high-performance unified 64-bit cores running up to 5.7 GHz
  • 16 DDR5 memory controllers
  • < li>64 PCIe 5.0 lanes

  • Multiprocessor support for 4-socket and 2-socket platforms
  • Rack solutions for both air-cooled and liquid-cooled data centers
  • SPECrate 2017 Integer performance of around 4x Intel 8380 and around 3x AMD 7763HPC
  • Double-Precision Floating-Point performance is 3x NVIDIA H100
  • AI FP8 performance is 6x NVIDIA H100

x86, ARM, and RISC-V code per emulator only

A catch with the processor is that the CPU only runs with its own ISA. Code written for x86, ARM or RISC-V processors only works via emulation. The manufacturer explained two years ago that the huge raw performance is still sufficient to more than compensate for emulation losses and continue to offer the fastest solution on the market. But so far there is nothing more than the insurance from Tachyum. However, new Linux emulations in an FPGA environment should continue to attract potential investors.

The fact that the startup actually manages to produce real hardware that then also works is clear the roadmap but before the question of software compatibility and the performance of an emulation. Then comes the software topic, without which nothing works today – and with such a special solution, a lot more work has to be invested there than in other areas.

What the coming years will show. The target group in the HPC segment is currently still more than skeptical about the topic.