Start up company Tachyum claims to design a processor that can replace cpus and gpus

[spartaman64]

Quote

Just over a month ago, Tachyum, a small tech startup, announced a new processor family, Prodigy, which it calls the world's first universal processor platform. Beyond that new label though are a number of hard-to-believe claims that the company has put forward including the fact that it delivers a 10x performance improvement on conventional processors.

1

Quote

1. How is it different from the Heterogeneous System Architecture (HSA) Foundation?

Tachyum’s Prodigy processor is a new and innovative processor architecture, developed using a hardware and software codesign from day-1. It has a single programming model, a single instruction stream, fully coherent memory, and fully coherent inter-core communication. We have also added data parallelism to our processor programming model in order to better address certain AI applications. 

1

Quote

2. Your press release mentions an AI chip, GPU and CPU as being part of that family. Can you tell us more?

Tachyum’s Prodigy processor is a single unified architecture, which exhibits out of order performance, with processor hardware (read: transistor counts / core size) similar to simple in-order execution machines. We have accomplished this by offloading to our compiler, tasks traditionally implemented in CPU hardware. The resulting IPC, clock speed and power reduction improvements not only offer a compelling value proposition in our core market, Hyperscale Data Centers, but they also enable Tachyum’s Prodigy processor to exceed NVIDIA Volta performance on Neural Nets.

5

Quote

Tachyum has developed an innovative processor architecture that offers a disruptive value proposition across multiple application domains. We have also included in the Prodigy architecture certain architectural improvements to enhance its performance on AI workloads, such as compressed 8-bit floating-point coefficients and matrix multiply-add operations.

Quote

5. You say that building a candidate for the human brain project will take less than three years with about 250,000 of your chips, is it safe to assume that this processor has a peak performance of 4 teraflops?

Tachyum’s 64 core Prodigy processor, will generate ~128TFLOPS. We claim that in 2020, with volume production of Prodigy underway, system integrators will be able to assemble ~250,000 Prodigy processors into a network capable of running human brain sized neural nets. The processing density of Prodigy, combined with its disruptive low power consumption, enable these systems to be built beginning in 2020. 

1

Quote

4. Why can Tachyum succeed where larger organizations have failed?

I would have to say it is due to my technical instincts born from hard won experience, along with learning from others’ mistakes, and disciplining myself to work only on important challenges. Even in large companies such as Intel and NVIDIA, the real innovation usually springs from a small group of innovators. At SandForce which I founded, my competitors were Intel, Samsung, Toshiba, Sandisk, Micron, Western Digital, LSI, Seagate, and many others with 1000s [of] engineers, and with less than 100 employees we won.

Having the instinct to go in right direction is important, hiring best team, learning from mistakes of competitors, and working only on important stuff. An example is the Intel Itanium that failed on compilers, so we developed compilers first and build architecture around compilers. Even at nVidia, key innovation was done with a couple [of] groups with ten engineers, not thousands. If you have team of “gods”, [head] count is not that critical and can be filled with contractors.

6

Quote

3. Your company is promising to durably disrupt the compute market with some, frankly extraordinary claims. How did you achieve so much by spending (in relative terms) so little in R&D compared to the likes of Samsung, Nvidia or Intel.

The Prodigy architecture is the result of decades of experience that I developed designing processors (e.g. Playstation 2, Tesla), flash memory controllers (Sandforce), and flash based systems (Skyera). Several years of self-funded R&D preceded Tachyum’s emergence from stealth mode. I have always been interested in solving “device physics” challenges, such as reliability issues in dual level cell flash memory, as I did at Sandforce. Prodigy is another example of that. With the decade long stagnation of processor clock speed, due in large part to slow wires relative to transistor switching speed, and coupled with CPU architectures which were designed when wires were infinitely fast compared to transistors, a fresh look at an optimal 21st century processor architecture was warranted. We started from a clean sheet of paper with a design philosophy of reducing the number of slow wires on a chip, and reducing the average length of existing wires. The result is breakthrough performance and low power consumption.

8

Quote

7. We spoke a lot about hardware but what the role of software in your plans? 

Our smart compiler is essential to the Prodigy solution, handling many processor tasks traditionally handled in hardware. We have GCC today, and will provide LLVM next year as well as Java JIT. Linux and FreeBSD will be natively supported. We will work closely with application developers to insure they can fully exploit the performance characteristics of Prodigy for both data center apps, as well as AI applications in all domains. 

I don't know much about processor architecture but from what I understand they are saying their processor can do a mix of what cpus are gpus do and it offloads some cpu tasks that it can't do to a compiler so in effect it can replace cpus and gpus. I don't quite buy what they say about only needing only a few employees for key innovations because why would you hire a large number of them if they are unneeded. I would think nvidia when designing a processor would want to be "reducing the number of slow wires on a chip, and reducing the average length of existing wires" so I doubt this was something that only they thought of. I'm skeptical of their claims of being superior to current processor companies like nvidia and intel but I guess only time will tell.

 

https://www.techradar.com/news/this-startup-wants-to-kill-the-cpu-and-gpu-in-one-go

[Tabs]

Until hardware actually exists that can be demonstrated easily by a third party, it's just another VC funding grab. I guarantee they'll put out a call for funding in the next 6 months.

 

[M.A.P]

It is a good thing to replace both cpu and gpu basically an apu though it is amd term. They are not the first to do that thing, amd apus are already able to replace low mid range cpu + gpu combos for ex 2200g and 2400g and depending on the workload some customers are good with the igpu for intel consumer grade cpus but here is the main part, what is the price and also I'm sure they'll call for funding and also it will take many months for them to get mobo partners and also the important thing the license for both x86 and x64 is closed and is patented by intel and amd respectively and this is the main reason why we are not seeing any new cpu. So sure after going through this process the price is going to be high. And also it might be a scam like smach z. It totally depends on the price they tag them.

[AskTJ]

Probably another startup that has ridiculous claims, will raise money and run away with it.

[Guest]

It's really just too early to tell what this will be, but good luck to them.

[Dabombinable]

Sony were thinking of this with the PS3. They dropped the idea because it'd be difficult to program for.

[Dabombinable]

[Froody129]

Lmao I wanna know how powerful the human brain is...

 

"Out of order" processing sounds like it's processing in parallel... Sooo it's a multicore chip? How would they deal with different workloads where parallel (graphics) or sequential (traditional CPU tasks) performance is preferred? 

 

 

 

[DuckDodgers]

The vague description of this architecture is quite similar to Intel's EPYC (Intanium) ISA: a simplified VLIW pipeline where proper instruction scheduling is offloaded to the compiler.

The only successful implementation of this ISA type was AMD's TeraScale series of GPUs and those required continuous driver/compiler tuning, despite being dedicated graphics hardware and not universal (general purpose) architecture.

[NMS]

I call BS on this so much!

Not only I highly doubt that they would be twice as good, performance-wise, but tenfold?! Those are some very out-of-this-world claims.

[AlTech]

10 minutes ago, Froody129 said:

Lmao I wanna know how powerful the human brain is...

 

"Out of order" processing sounds like it's processing in parallel

1

Actually no.

Quote

... Sooo it's a multicore chip? How would they deal with different workloads where parallel (graphics) or sequential (traditional CPU tasks) performance is preferred? 

 

That's not it either.

 

In order execution is when a CPU executes instructions in exactly the order it receives them. This is generally not preferable for situations where performance matters as it is not super fast. It does use less power than out of order execution which is the only thing it has going for it.

 

Out of order execution is when a CPU is able to use instruction cycles which would otherwise be wasted to execute other instructions which are able to be run immediately.

 

Thie Wikipedia article is wayyy more detailed as it's a complicated subject. https://en.wikipedia.org/wiki/Out-of-order_execution

 

The only inherent issues with out of order execution are that it costs more to implement and it uses more power. A real-world example would be ARM. The ARM Cortex big Cores use Out Of Order Execution whereas ARM's LITTLE cores use In order execution.

[spartaman64]

13 minutes ago, AluminiumTech said:

Actually no.

That's not it either.

 

In order execution is when a CPU executes instructions in exactly the order it receives them. This is generally not preferable for situations where performance matters as it is not super fast. It does use less power than out of order execution which is the only thing it has going for it.

 

Out of order execution is when a CPU is able to use instruction cycles which would otherwise be wasted to execute other instructions which are able to be run immediately.

 

Thie Wikipedia article is wayyy more detailed as it's a complicated subject. https://en.wikipedia.org/wiki/Out-of-order_execution

 

The only inherent issues with out of order execution are that it costs more to implement and it uses more power. A real-world example would be ARM. The ARM Cortex big Cores use Out Of Order Execution whereas ARM's LITTLE cores use In order execution.

asynchronous compute?

[mr moose]

They're a bit slow, I've already got one of those.  I keep it the display cabinet next to my holy grail and bottle of unicorn tears.

[Arika]

Quote

Start up company

Quote

claims

Quote

a processor that can replace cpus and gpus

 

[Christophe Corazza]

"Tachyum’s architecture overcomes the limitations of semiconductor device physics, which were thought to be insolvable."

 

About as likely as me manufacturing a non flammable Lithium-ion battery in my garage.

 

Smells like pump and dump investment to me!

[Bananasplit_00]

I hope its true at least? I'm willing to belive that they have a working chip at least. Maybe they mean it's 10x the performance of the Intel Itanium chip because its going for the same thing? 

[Guest]

48 minutes ago, Froody129 said:

Lmao I wanna know how powerful the human brain is...

 

"Out of order" processing sounds like it's processing in parallel... Sooo it's a multicore chip? How would they deal with different workloads where parallel (graphics) or sequential (traditional CPU tasks) performance is preferred? 

 

 

 

Until we can accurately model the brain you can't really quantify it. It's not the same processing as a computer chip (and will never be really). AI will always be behind the human brain. 

[mr moose]

51 minutes ago, RorzNZ said:

Until we can accurately model the brain you can't really quantify it. It's not the same processing as a computer chip (and will never be really). AI will always be behind the human brain. 

We have over 100trillion neural connections, that's like a self learning FPGA with 100trillion gates.  Easy to imagine hard to make. 

[Guest]

1 hour ago, mr moose said:

We have over 100trillion neural connections, that's like a self learning FPGA with 100trillion gates.  Easy to imagine hard to make. 

A neural connection isn't really an on/off affair. We don't even know how the brain works fully, so how can we model it. I'm not really an expert on neural networks however. 

[mr moose]

1 minute ago, RorzNZ said:

A neural connection isn't really an on/off affair. We don't even know how the brain works fully, so how can we model it. I'm not really an expert on neural networks however. 

Quite a bit is known actually, the problem is it is so large that mapping it and correlating one action to another (I.E memories to events or chemicals to feelings) is very hard.

 

http://www.human-memory.net/brain_neurons.html

[Guest]

2 minutes ago, mr moose said:

Quite a bit is known actually, the problem is it is so large that mapping it and correlating one action to another (I.E memories to events or chemicals to feelings) is very hard.

 

http://www.human-memory.net/brain_neurons.html

Although a neuron is an electric signal, a lot is facilitated. Largely the cortex contains the higher functions really needed to improve neural networks, and works in companion with the lower centres. They form very complex relationships, and we don't know much especially regarding thinking, for instance thinking can occur in the lower centres of the brain (mid-brain mostly) but complexity of thought occurs in the cortex. Glial cells also contribute much more than just nutritional roles to the neurones and their function isn't entirely understood (for instance inhibition and myelin synthesis). Mapping in terms of electrical signals isn't an accurate way to think about it. 

[mr moose]

9 minutes ago, RorzNZ said:

Although a neuron is an electric signal,

No, a neuron is a nerve cell. 

9 minutes ago, RorzNZ said:

a lot is facilitated. Largely the cortex contains the higher functions really needed to improve neural networks, and works in companion with the lower centres. They form very complex relationships, and we don't know much especially regarding thinking, for instance thinking can occur in the lower centres of the brain (mid-brain mostly) but complexity of thought occurs in the cortex. Glial cells also contribute much more than just nutritional roles to the neurones and their function isn't entirely understood (for instance inhibition and myelin synthesis). Mapping in terms of electrical signals isn't an accurate way to think about it. 

I'm not sure that makes a lot of sense. But I'll leave it at that because My experience with neurons and development stems mainly from working with pediatricians in early childhood development and studying the roles of environment versus genetics in brain development of children with mental disorders.

[Guest]

6 minutes ago, mr moose said:

No, a neuron is a nerve cell. 

A neuron transmits an electric signal is what I meant to say (Bit sleepy but need to stay up). It's much more convenient to think of it in pathways rather than a simple cell to cell connection as well, but it's rather hard to find those for certain functions. There's people walking around with half a brain at the moment who function normally, or even hardly any cortex. 

6 minutes ago, mr moose said:

I'm not sure that makes a lot of sense. But I'll leave it at that because My experience with neurons and development stems mainly from working with pediatricians in early childhood development and studying the roles of environment versus genetics in brain development of children with mental disorders.

I've studied neurophysiology at 300-level in college. Not an expert by any means but I know the basics. 

[Dash Lambda]

Yes, modern microarchitectures are getting a bit bloated and have a lot of inefficiencies grandfathered in, but they're saying they'll have a chip that outperforms Volta by ten times and has the single-threaded muscle to replace a CPU. That's like saying you'll make a car that goes 350mph and gets 100mpg.

 

3 hours ago, RorzNZ said:

AI will always be behind the human brain. 

How so?

[NumLock21]

Getting a cpu from sand is really difficult and the machines to do that probably cost in the billions. Not to mention the cost of R&D. If they got the machine to make a prototype cpu in the first place, they won't need any crowd funding. As of now they're just pulling claims out of their asses.