News Tachyum's Monster 128 Core 5.7GHz 'Universal Processor' Does Everything

[Admin]

Tachyum builds a 128-core super processor that can handle a multitude of high end applications on a single chip.

Tachyum's Monster 128 Core 5.7GHz 'Universal Processor' Does Everything : Read more

 

[thisisaname]

Was looking good until I read at the end "Production starts in 2023, so we should see actual benchmarks of these chips sometime next year. " So a dream CPU in more than one meaning of the word.

 

[atmapuri]

AVX512 looks shy in compare to 2x1024bit vector and 1x4096bit matrix registers. Yes, On paper one core is 4x faster than Intel for vector math. 16x50GB/s = 800GB/s memory bandwidth, if you can afford to run minimum 32 threads to make use of that.

 

[SunMaster]

It'll be very interesting to see how this materialize into the real world.

 

[mdd1963]

If this CPU was x86-64 compatible, and one could install Windows Server 2022 Datacenter, it should cost only $36K in Windows licensing fees, at 16 cores per license, ... x 8! SQL Server Enterprise would run just $879, 872!

I 'd assume all these cores would need to be spread out into an area the size of a piece of bread, or two.... at a minimum...

 

[jasonf2]

But will it play Crysis?

 

[jasonf2]

If this CPU was x86-64 compatible, and one could install Windows Server 2022 Datacenter, it should cost only $36K in Windows licensing fees, at 16 cores per license, ... x 8! SQL Server Enterprise would run just $879, 872!

I 'd assume all these cores would need to be spread out into an area the size of a piece of bread, or two.... at a minimum...

I am pretty sure looking at the arc on this thing it is going to require a custom OS. By the time that port is done that $900,000 figure you just threw will look like a deal.

 

[Friesiansam]

If the reality lives up to the bold claims, it will be a highly impressive chip but, if something seems to good to be true, it usually is.

 

[farnell121]

"The Prodigy chip can also run binaries for x86, ARM, RISC-V, and ISA".

I'm assuming you meant POWER ISA, instead of just Instruction Set Architecture, as that would be redundant.

 

[Historical Fidelity]

I’m calling bs on this. Factors more powerful, factors more energy efficient, yet the silicon is 5400mm2 in area. Doesn’t add up, but then again you gotta try the snake oil before you can judge haha

 

[hotaru.hino]

Tachyum boldly claims the Prodigy supercomputer chip offers four times the performance of Intel's fastest Xeon on the market and triple the raw performance of Nvidia's H100 in high-performance computing applications. All while being 10 times more power efficient.

Where have I heard this before and it ended up falling flat on its face?

 

[Sippincider]

But will it play Crysis?

But will it run Cities:Skylines with a population ~1 million?

 

[D Gates]

"The Prodigy chip can also run binaries for x86, ARM, RISC-V, and ISA".

I'm assuming you meant POWER ISA, instead of just Instruction Set Architecture, as that would be redundant.

From Tachyon's site, it's their native ISA. It's unnamed, so I assume it's unique to Prodigy.

 

[D Gates]

I am pretty sure looking at the arc on this thing it is going to require a custom OS. By the time that port is done that $900,000 figure you just threw will look like a deal.

It uses QEMU for ISA translation so it should be able to run a bunch of operating systems out of the box. Even ones like Windows and Android. On their YouTube they show FreeBSD and Linux. They said they ported FreeBSD to run natively and I assume it's the same for Linux.

 

[Raj_Parihar]

a single Nvidia A100 is only capable of 5 AI PetaFLOPS.

Really. I thought it was 640 TOPS which is much higher than 5 PetaFLOPS.

Also this is all BS! I am wondering what are they smoking. Must be of really good quality!

 

[jp7189]

I'll assume these claims are coming from simulations. That's a lot of silicon so who knows maybe the performance claims can be met, but I'm not so sure about the efficiency claims. The thing I'm most stumbling over is more fundamental. My mind boggles over running the traces to support 16x DDR5 channels at what is likely to be the top end speed for 2023 AND 64x PCIe 5 lanes.

 

[jasonf2]

Where have I heard this before and it ended up falling flat on its face?

Transmeta?

 

[Colif]

So is going to be used to design Deep thought?

 

[NinoPino]

Transmeta?

Transmeta was efficient and innovative at the time but not very performant.
I think he is referring to Intel 432 or similar.