Flow Computing Claims its PPU can Boost CPU Performance by 100x

[TheLostSwede]

A Finnish company called Flow Computing is making some very bold claims when it comes to its new IP. The company has developed what it calls a Parallel Processing Unit or a PPU, which the company claims can boost the performance of any CPU by a hundred times its current performance. Furthermore, the company claims that its PPU can double the performance of any current code execution, with no need for any kind of optimisation for its PPU. The PPU can be integrated into new processors, but it can also be designed as a discrete chip that can be added to any current hardware and Flow Computing claims the performance benefits will be the same in both instances.

Flow Computing is a spinoff from the VTT Technical Research Centre of Finland and the company emerged from stealth mode last week with around €4 million in funding. Flow Computing doesn't intend to make its PPU by itself, but instead, the company wants to licence its tech to third parties, to give everyone an equal opportunity to take advantage of what's on offer. At this point in time, Flow Computing hasn't made any custom silicon, instead the company has validated its PPU using an FPGA tested against various Intel CPUs. As such, there are numbers to back up its claims and we've provided links below to a whitepaper and an FAQ for those that are interested in doing a deep dive into its claims. Flow Computing appears to have a few different implementations of its PPU, ranging from 16 to 256 cores, with the latter being for high-end computers, but the basic is said to be suitable for something as basic as a smartwatch. Time will tell if Flow Computing will be able to deliver on its claims and it'll be an interesting company to follow.



View at TechPowerUp Main Site | Source

 

[TumbleGeorge]

Large ratios like 100X seem to be a favorite of scammers lately.

 

[Steevo]

Physics Processing Unit?

I feel like we have seen this stuff before. How long before Nvidia buys them?

 

[natr0n]

It's like the original physx card all over again. People were testing games it wasnt made for I remember.

 

[DarkDreams]

Perfect example of how useless the i7 moniker is. As one can see in slide 3 fine print, they are using an i7 from a macbook pro with 4 cores clocked at 3.2 GHz which I think would make it a Clarksfield or Sandy Bridge CPU. That's FIFTEEN year old hardware. Thats how you get to be 100x faster

 

[xorbe]

Likely another IP company with no product looking to claim everyone owes them license fees eventually.

 

[Punkenjoy]

Extraordinary claims require extraordinary evidence

 

[TheLostSwede]

Perfect example of how useless the i7 moniker is. As one can see in slide 3 fine print, they are using an i7 from a macbook pro with 4 cores clocked at 3.2 GHz which I think would make it a Clarksfield or Sandy Bridge CPU. That's FIFTEEN year old hardware. Thats how you get to be 100x faster

Good catch.

 

[vmetodiev]

I don't know why that reminds me that much of Tachyum.

 

[ScaLibBDP]

>>...Flow Computing hasn't made any custom silicon, instead the company has validated its PPU
>>using an FPGA tested against various Intel CPUs...

I stopped reading after that statement and here is my comment:

This is a Classical Example of Technological Scam, similar to Tachyum claims about their Universal Processing Unit. Where is the Tachyum now?

Please read Fine Print text on the Slide #3...

Apple computer owners What do you think?

I don't know why that reminds me that much of Tachyum.

I stopped following Any news about Tachyum since August 2023...

Large ratios like 100X seem to be a favorite of scammers lately.

NVIDIA designed an "acceleration technology" many-many years ago and called it GPU. Everybody knows that GPUs really accelerate processing.

Almost 4 years ago I've attached 4 External GPUs to a Dell Precision M4700 Mobile Workstation and accelerated a vector addition processing by 125 times.

Please take a look at the Video Technical Report:.

Dell Precision M4700 Mobile Workstation accelerated 125x to 11.155 TFLOPs ( VTR-088 )
Published on 21.10.2020

If it would be so easy to increase performance by 100 times without external accelerators, like GPUs or FPGAs, everybody would do it. Period.

 

[freeagent]

How long before Nvidia buys them?

They did already, that's how we have it on GPU now

 

[TumbleGeorge]

Please take a look at the Video Technical Report:.

A specific type of tasks suitable for GPU acceleration with cuda is accelerated on a very old laptop that does not natively support this technology. Very, very different from claims of universal 100X accelerations.

 

[Assimilator]

Bullshit and grift, designed to lure gullible investors so they can take the money, embezzle it, then tell those investors "lol sorry guess the prodcut doesn't work, haha, too bad" and close the company down.

 

[TheLostSwede]

Bullshit and grift, designed to lure gullible investors so they can take the money, embezzle it, then tell those investors "lol sorry guess the prodcut doesn't work, haha, too bad" and close the company down.

It's a Finnish company, not a US company.

 

[Assimilator]

It's a Finnish company, not a US company.

Nationality is no indication of lack of willingness to grift.

 

[ScaLibBDP]

Bullshit and grift, designed to lure gullible investors so they can take the money, embezzle it, then tell those investors "lol sorry guess the prodcut doesn't work, haha, too bad" and close the company down.

Take a look if interested.

Flow Computing company was founded by 3 academia guys.

I've read two documents from

Did startup Flow Computing just make CPUs 100x faster? Here’s the white paper and FAQs

Read for yourself.

www.theverge.com

Document 1: https://s3.documentcloud.org/docume...faq-under-embargo-until-11-jun-at-9am-edt.pdf
Document 2: https://s3.documentcloud.org/docume...-design-goals-and-hw-sw-advantages-2024-5.pdf

They've claimed that No publications or Github codes are available however there is a very good publication of a Thick Control Flow ( TCF ) architecture at:

Parallel processing unit performance

The performance and programmability comparison of Flow’s F16 vs. Intel Skylake Core i7 and Xeon W processors assuming the same clock rate.

flow-computing.com

Performance and Programmability Comparison of the Thick Control Flow Architecture and Current Multicore Processors

https://doi.org/10.1007/s11227-021-03985-0

My personal opinion, as a Software Engineer, is Not changed: It looks like a High-Tech-Scam similar to Universal Processing Unit of Tachyum.

 

[persondb]

PS3/Cell Broadband Engine 2.0

They even used the same names with the only difference being the PPU as the CPU Core and the SPU as the Vector/Parallel unit

 

[Steevo]

They did already, that's how we have it on GPU now

I mean this new company, I know Nvidia bouht Physx. Back in the day many of us worked on cracking the Physx library to make it work on red card for the lulz, its how I found out most "real time" Physx was still pre-cooked, and it was literally titled "pre-cooked" in Nvidias documentation and programming API guide. Scientific acceleration happened first on red GPU's, I remember running Folding at Home on ATI only hardware with VLIW architecture not because it was easy, but because Nvidia gimped their color/math hardware for performance and wasn't able to run it without errors.


This flow computing PPU is either

1) A scam
2) Real world Dunning-Kruger effect
3) Revolutionary

I am going with 50/49.9/.1 as the chances for each.

 

[ScaLibBDP]

I mean this new company, I know Nvidia bouht Physx. Back in the day many of us worked on cracking the Physx library to make it work on red card for the lulz, its how I found out most "real time" Physx was still pre-cooked, and it was literally titled "pre-cooked" in Nvidias documentation and programming API guide. Scientific acceleration happened first on red GPU's, I remember running Folding at Home on ATI only hardware with VLIW architecture not because it was easy, but because Nvidia gimped their color/math hardware for performance and wasn't able to run it without errors.


This flow computing PPU is either

1) A scam
2) Real world Dunning-Kruger effect
3) Revolutionary

I am going with 50/49.9/.1 as the chances for each.

Have a quick look at the publication ( weblink in my previous post / I've read it today / 32 pages in total ).

These academia guys made All conclusions after running a set of benchmarks in a simulator with assumptions like "memory latency is hidden", or Emulated Shared Memory ( ESM ), and compared results against two real Intel CPUs.