Tuesday, May 5th 2020

Tachyum Prodigy is a Small 128-core Processor with Crazy I/O Options, 64-core Sibling Enroute Production

Silicon Valley startup Tachyum, founded in 2016, is ready with its crowning product, the Tachyum Prodigy. The startup recently received an investment from the Slovak government in hopes of job-creation in the country. The Prodigy is what its makers call "a universal processor," which "outperforms the fastest Xeon at 10X lower power." The company won't mention what machine architecture it uses (whether it's Arm or MIPS, or its own architecture). Its data-sheet is otherwise full of specs that scream at you.

To begin with, its top trim, the Prodigy T16128, packs 128 cores on a single package, complete with 64-bit address space, 512-bit vector extensions, matrix multiplication fixed-function hardware that accelerate AI/ML, and 4 IPC at up to 4.00 GHz core clock. Tachyum began the processor's software-side support, with an FPGA emulator in December 2019 (so you can emulate the processor on an FPGA and begin developing for it), C/C++ and Fortran compilers; debuggers and profilers, tensorflow compilers, and a Linux distribution that's optimized it. The I/O capabilities of this chip are something else.
For starters, the chip features two 400 Gbps PHYs integrated, along with a PCI-Express gen 5.0 x48 root-complex split between four SerDes. It also has a gargantuan 12-channel DDR5 memory interface (albeit limited to 1 DIMM per channel, and 512 GB cumulative memory size).

For now, Tachyum has two performance claims for the T16128. 16 TFLOPs HPC throughput (likely single-precision), and 262 TFLOPs (equivalent) AI training and inference performance, thanks to its matrix multipliers integrated into the core.

All of this crams into a 65 mm x 65 mm FCBGA package with an integrated heatspreader (slightly bigger than LGA2066 package, although smaller than an SP3 package). Underneath the IHS, the die has been fabricated on TSMC 7 nm process.

Tachyum has reportedly already taped out a smaller 64-core variant of this chip, the Prodigy T864. It features a slightly different I/O, with an 8-channel DDR4/DDR5 memory interface, an optional HBM3 interface, a wider 72-lane PCI-Express gen 5.0 root complex, and two 400 Gbps PHY interfaces. Built on the 7 nm process like its 128-core sibling, this chip's die measures 290 mm², and has a TDP rating of 180 W. ComputerBase reports that the 64-core variant will enter production within 2020.
Source: ComputerBase.de
Add your own comment

51 Comments on Tachyum Prodigy is a Small 128-core Processor with Crazy I/O Options, 64-core Sibling Enroute Production

#1
btarunr
Editor & Senior Moderator
I'm really, really looking forward to a proof of concept.
Posted on Reply
#3
ratirt
Wonder what is the price. I'd like to see the performance charts and some comparisons. No matter how you look at it, it seems impressive if true.
Posted on Reply
#4
hat
Enthusiast
Seems like a pipe dream to me...
Posted on Reply
#5
londiste
They are avoiding the question of ISA.
"Out of order execution with compiler" implies this is an in-order architecture.
"Faster than Xeon, smaller than ARM?" In what operations would exactly that be?

Sounds like an HPC chip, more akin to GPU than a CPU.
Or a pipe dream.

4GHz on TSMC 7nm? AMD seems to put a lot of work in to achieve that speed.
Small wires help with each individual core, sure, but how is inter-core traffic organized?
Posted on Reply
#7
wishgranter042
TheGuruStud
Reeks of investor scam to me.
take a look on who its behind.. no greenhorns...

www.tachyum.com/company
londiste
They are avoiding the question of ISA.
"Out of order execution with compiler" implies this is an in-order architecture.
"Faster than Xeon, smaller than ARM?" In what operations would exactly that be?

Sounds like an HPC chip, more akin to GPU than a CPU.
Or a pipe dream.

4GHz on TSMC 7nm? AMD seems to put a lot of work in to achieve that speed.
Small wires help with each individual core, sure, but how is inter-core traffic organized?
www.tachyum.com/resources
Posted on Reply
#8
londiste
VLIW instruction set with custom vector and matrix-multiplication instructions
This isn't the most efficient thing or the easiest thing to get to work for general compute applications. HPC/AI - sure.

In-order architecture.
Topology is mesh, with IO ring.
Outperforms Xeon on SpecInt & SpecFP 2006 benchmarks
- One 3.5GHz Xeon E5-2687W v4 core vs one Prodigy core, same GCC 7.2
Posted on Reply
#10
AnarchoPrimitiv
If it's too good to be true.... It probably is.... I'm sure that If this is real and IF it works, that it's going to have a very specific use case... And was that a single monolithic die?
Posted on Reply
#11
Caring1
nemesis.ie
@btarunr The correct term is "Slovak".
Nah, Slovakian is correct.
Posted on Reply
#12
ratirt
Caring1
Nah, Slovakian is correct.
No it isn't. Czechian and Polishian are not correct so as Slovakian. :)
Posted on Reply
#13
Aquinus
Resident Wat-man
ratirt
No it isn't. Czechian and Polishian are not correct so as Slovakian. :)
I take it that you never heard of the term "Czechoslovakian" before the two countries split.
Posted on Reply
#14
DeathtoGnomes
This might go to market soon enough but how long until the next one?
Posted on Reply
#15
ratirt
Aquinus
I take it that you never heard of the term "Czechoslovakian" before the two countries split.
It was Czechoslovakia as a country not Czechoslovakian. It was a combination of the Czech and Slovakia as one country. There never was Czechoslovakian language. Besides Czech and Slovak (from Slovakia) are different languages. Similar but different.
Lets get back to the topic of this thread. You can always educate yourself about Slavic languages at your own time instead of duplicating someone's mistakes. :)
Posted on Reply
#16
ARF
ratirt
It was Czechoslovakia as a country not Czechoslovakian. It was a combination of the Czech and Slovakia as one country. There never was Czechoslovakian language. Besides Czech and Slovak (from Slovakia) are different languages. Similar but different.
Lets get back to the topic of this thread. You can always educate yourself about Slavic languages at your own time instead of duplicating someone's mistakes. :)
Dialects, to be precise. They are extremely close - maybe as close as how American is to British English.
nemesis.ie
@btarunr The correct term is "Slovak".
According to Wikipedia, the correct demonym is indeed Slovak en.wikipedia.org/wiki/Slovakia
Posted on Reply
#17
londiste
AnarchoPrimitiv
If it's too good to be true.... It probably is.... I'm sure that If this is real and IF it works, that it's going to have a very specific use case... And was that a single monolithic die?
Information is not too clear on that. They have (or plan) a 64-core die and 128-core CPU is a package, meaning most likely two 64-core dies on one package.
Posted on Reply
#18
ratirt
ARF
Dialects, to be precise. They are extremely close - maybe as close as how American is to British English.
Yes they are intelligible but they are recognized as 2 different languages. During the Czechoslovakia time, government tried to combine these two but failed. These are not dialects but two different language and that's how linguists recognize these two. Dialects are within the same language. You have different dialects in Czech language depending on the region so as different dialects in Slovak languages. Intelligibility of the two languages changes depending on the region a person lives. If you take western Czech person and eastern Slovak person to talk with each other, they would have had a problem understanding each other. On the other hand, Slovak people from eastern and western side of Slovakia would have no problems understanding each other.
Please lets get back to the topic.
Posted on Reply
#19
Vya Domus
"Out of order execution with compiler"

:roll:

That's a good one.

"SMT support with time sharing"
Posted on Reply
#20
yeeeeman
Out of order execution with compiler - this means the core is in order. No wonder it can cram 128 cores in 180W.
Posted on Reply
#21
Caring1
nemesis.ie
@btarunr The correct term is "Slovak".
Now the article doesn't make sense as it's been changed.
Stop confusing the poor man with your misunderstanding of the English language.
A Slovak is a native of Slovakia, Slovakian describes something or someone from Slovakia.
Posted on Reply
#22
Muser99
Looks like an EU pipedream aiming to be independent of Intel/US?
Posted on Reply
#23
ratirt
Muser99
Looks like an EU pipedream aiming to be independent of Intel/US?
Is that how you take innovations and something new not coming from US? It would seem that innovations are not desirable. I'd like to know more about this Techyum thing. 128 cores is pretty slick if you ask me. Wonder why it uses all the newest tech stuff. PCIe5 and DDR5 etc. Hopefully this isn't a marketing scheme.
Posted on Reply
#24
Aquinus
Resident Wat-man
A have a legit question though, what ISA does this chip use? Is it a custom homegrown ISA designed specifically for this chip? If that's the case, adoption is probably going to be pretty rough because there will be very little tooling for developing for it, even if the specs sound pretty impressive. We still need to be able to use it, not just oogle at its specs.
Posted on Reply
#25
londiste
Aquinus
A have a legit question though, what ISA does this chip use? Is it a custom homegrown ISA designed specifically for this chip? If that's the case, adoption is probably going to be pretty rough because there will be very little tooling for developing for it, even if the specs sound pretty impressive. We still need to be able to use it, not just oogle at its specs.
wishgranter042
www.tachyum.com/resources
VLIW instruction set with custom vector and matrix-multiplication instructions
Kanter also pretty directly says compiler and other tools are essential to this. I cannot not think about Itanium when looking at this.
Posted on Reply
Add your own comment