Huawei to enter the server GPU business

[igormp]

Looks like Huawei wants to build some desktop-class GPUs in order to face nvidia and AMD head on.

 

Quote

However, there's at least one reason why Huawei might want to expand its horizons - the company has already built several AI accelerator solutions that are able to compete with those of Nvidia. For instance, Huawei's Ascend 910 chip is able to deliver 256 teraflops of half-precision performance, which is two times faster than Nvidia's Tesla V100 solution in certain scenarios.

----

It would make sense for Huawei to move into GPUs for high-performance computing for scenarios that require general-purpose capabilities as opposed to AI chips which are mostly tailored for training neural networks in deep learning applications.

Source

 

Quote

With Huawei's issues with the US government and its ability to sell smartphones outside of China, the company is forced to look for other revenue streams to compensate for the lost opportunities. So the latest rumor coming from South Korea is that the company is setting up a research center in the country with some new recruits coming from Nvidia.

 

It's important to note that these are not mainstream GPUs we are talking about here. It's about server-grade GPUs used for cloud computing and AI - a market which Nvidia is currently dominating. According to industry insiders, the tech giant has been developing its own GPU chip since 2012 and will present it to the world when the time is right.

It's also possible that the Huawei is working on a GPU to satisfy its own needs as it's currently running some cloud computing services.

Source

 

I wonder if we'll finally have a nice market when it comes to the server GPU market. Nvidia currently dominates this market, with Intel having the handle on some niche markets with their Xeon Phi product and AMD selling some GPUs to Google for Stadia.

Edited April 14, 2020 by igormp
New title to better suit the news

[Crunchy Dragon]

That's interesting. Wonder how it'll play out in years to come.

[TempestCatto]

So now they can spy on my shite tier gaming abilities? Sweet. It's like streaming without the work!

[G0dSpeed]

I get enough crap for supporting AMD GPUs so....no thanks...

[Takumidesh]

That's interesting. It would be nice to see someone try to make some waves in the GPU world and get Nvidia to wake up a bit. 

 

On the other hand, is Huawei really going to be able to compete? I wonder if they actually would be able to bring something to the table.

[givingtnt]

Man I can't wait to play mobile games on my pc!

[Pickles von Brine]

@igormpYou may want to change the title. This is not desktop GPUs but the server market. High performance compute. 

 

[justpoet]

…and now we know what the latest companies they stole IP and trade secrets from are.

[igormp]

2 hours ago, Pickles - Lord of the Jar said:

@igormpYou may want to change the title. This is not desktop GPUs but the server market. High performance compute. 

Well, it's probably going to be a PCIe card with a regular desktop gpu form factor, that's why I said so, but sure.

 

2 hours ago, Takumidesh said:

That's interesting. It would be nice to see someone try to make some waves in the GPU world and get Nvidia to wake up a bit. 

 

On the other hand, is Huawei really going to be able to compete? I wonder if they actually would be able to bring something to the table.

If they ignore patents and make their GPUs compatible with CUDA, or try to build a nice API that's not as painful to write as OpenCL or Vulkan Compute, then I can see some future in their efforts.

[Pickles von Brine]

11 minutes ago, igormp said:

Well, it's probably going to be a PCIe card with a regular desktop gpu form factor, that's why I said so, but sure.

 

If they ignore patents and make their GPUs compatible with CUDA, or try to build a nice API that's not as painful to write as OpenCL or Vulkan Compute, then I can see some future in their efforts.

Server GPUs all look similar to desktop GPUs. It is a standardization that is in the industry. 

[BuckGup]

10 minutes ago, Pickles - Lord of the Jar said:

Server GPUs all look similar to desktop GPUs. It is a standardization that is in the industry. 

No they don't. GPUs used in the server market don't have active fans, display outputs, and now have custom connectors like NVLink on them

[Pickles von Brine]

3 minutes ago, BuckGup said:

No they don't. GPUs used in the server market don't have active fans, display outputs, and now have custom connectors like NVLink on them

The overall form factor and shape is what I am getting at. There are going to be differences with NVLink and such. Hell, Tesla V100 use a mezzanine socket. However, in general, a lot of server GPUs will usually look similar. They will look like a desktop GPU. Even if they are cooled with airflow via the chassis fan. 





I think that is what you are talking about with NVLink. But most are like the first picture. 

[BuckGup]

19 minutes ago, Pickles - Lord of the Jar said:

The overall form factor and shape is what I am getting at. There are going to be differences with NVLink and such. Hell, Tesla V100 use a mezzanine socket. However, in general, a lot of server GPUs will usually look similar. They will look like a desktop GPU. Even if they are cooled with airflow via the chassis fan. 
I think that is what you are talking about with NVLink. But most are like the first picture. 

I guess I don't really get the point of what your saying then. I mean there's a reason GPUs don't look like donuts and triangles.

[williamcll]

What architecture is it based off though? RISC-V

 

Probably manufactured TSMC, SMIC is still a bit behind.

[igormp]

25 minutes ago, williamcll said:

What architecture is it based off though? RISC-V

I'd love for it to be the case, but I really doubt it, they're probably going to roll their own arch.

Quote

Probably manufactured TSMC, SMIC is still a bit behind.

Likely, since their current chips are already being manufactured by TSMC.

[xAcid9]

earlier 
AMD: our graphic IP got stolen.

now
Huawei: we want to enter mainstream GPU market.

 

soon
WCCF: Huawei stole AMD's graphic IP and entering the GPU business. 

[Salv8 (sam)]

7 hours ago, igormp said:

Looks like Huawei wants to build some desktop-class GPUs in order to face nvidia and AMD head on.

i'm sorr..i...

 

 

may..maybe they need to...

ta..take on intel before AMD...

*beath*

ok...

like, good for them, but the entirety of the IT industry doesn't like nor trust them, like hell they are even gonna get anyone to buy their crap.

it probably won't be even able to beat an intel chip from 1992!

plus most of the people that could take advantage of this use linux as the server OS!

and you know china and their relationship with open source software! (spoilers: they don't like it)

 

i'm just gonna sit here and watch this shit show unfold like the turd paper plane it is!!!!

[thorhammerz]

4 hours ago, xAcid9 said:

WCCF: Huawei stole AMD's graphic IP and entering the GPU business. 

Introducing "Radeon Graphics solutions with Chinese characteristics"!

[porina]

I'll follow with indirect interest, since AI/cloud isn't something that interests me. Compute oriented GPUs are going to be easier than gaming ones as the operations tend to be more natively parallel and scaleable in that way, like Intel are doing.

 

If they really did use stolen AMD IP, it would set them back years...

 

5 hours ago, williamcll said:

What architecture is it based off though? RISC-V

That's a CPU architecture though, but I suppose the question then becomes if you make it wide enough, does it start to look like a GPU? Especially if you add/remove function blocks to optimise for the use case.

 

Actually, I'd go one step further. Is it even fair to call what they're making a GPU? It'll be a something-PU for compute uses. GPUs also have that capability, but if they don't need the graphical side so much they don't need to focus on it.

[S w a t s o n]

This is for national interests more than v nvidia or amd. Also, price

[williamcll]

7 hours ago, Salv8 (sam) said:

and you know china and their relationship with open source software! (spoilers: they don't like it)

 

i'm just gonna sit here and watch this shit show unfold like the turd paper plane it is!!!!

Where on earth have you heard that before? The majority of Chinese tech survives off open source because other nations are reluctant to sell for national security.

[igormp]

7 hours ago, Salv8 (sam) said:

like, good for them, but the entirety of the IT industry doesn't like nor trust them, like hell they are even gonna get anyone to buy their crap.

it probably won't be even able to beat an intel chip from 1992!

plus most of the people that could take advantage of this use linux as the server OS!

and you know china and their relationship with open source software! (spoilers: they don't like it)

 

Many companies only care about good prices and still buy from Huawei

Huawei already has many competitive IP cores in many other areas (from mobile CPUs and GPUs to cloud AI accelerators), so I guess than can truly come out with a nice product this time.

China actually has a nice track with open source, idk where you got that idea from.

 

4 hours ago, porina said:

That's a CPU architecture though, but I suppose the question then becomes if you make it wide enough, does it start to look like a GPU? Especially if you add/remove function blocks to optimise for the use case.

 

Actually, I'd go one step further. Is it even fair to call what they're making a GPU? It'll be a something-PU for compute uses. GPUs also have that capability, but if they don't need the graphical side so much they don't need to focus on it.

You can totally build a GPU based on RISC-V, it has been made before (look at NEOX). Simply taken, throw many minion cores at your project and abuse the SIMD/Vector instructions and you basically have a GPU.

 

Yes, it's a GPU, which is being used for GPGPU purposes, much like a Tesla card is used (even though it isn't used for graphics per se, we still call them GPUs, right?).

[porina]

1 hour ago, igormp said:

You can totally build a GPU based on RISC-V, it has been made before (look at NEOX). Simply taken, throw many minion cores at your project and abuse the SIMD/Vector instructions and you basically have a GPU.

To me that sounds more like a Xeon Phi (Knights... series), if you're focusing on FP units than pushing pixels.

 

1 hour ago, igormp said:

Yes, it's a GPU, which is being used for GPGPU purposes, much like a Tesla card is used (even though it isn't used for graphics per se, we still call them GPUs, right?).

My thinking is more along the lines if you don't actually use pixel functions, there's not much point in adding them in the first place unless it is essentially "for free" or needed for some other compatibility use. Without it, can you really call it a GPU at that point?

[igormp]

1 minute ago, porina said:

To me that sounds more like a Xeon Phi (Knights... series), if you're focusing on FP units than pushing pixels.

IMO a Xeon Phi is more suited to a MIMD case, where you can do proper branching and whatnot across your many core with those fat cores. What Huawer is building sounds like a regular GPU that's only able to work in a SIMD scenario. An implementation of that would be a main Risc-V core dispatching the instructions across the many minion cores to perform the same instruction on different data.

 

Quote

My thinking is more along the lines if you don't actually use pixel functions, there's not much point in adding them in the first place unless it is essentially "for free" or needed for some other compatibility use. Without it, can you really call it a GPU at that point?

Well, you're still using the shading units to perform compute (hence compute shaders that most graphical APIs have). The hardware is still the same, and the "GPU" name had already stuck before we starting using the hardware to do compute anyway.

I guess a better name would be Shader Processing Unit? Too bad Sony and IBM already claimed the SPU acronym haha

[Doobeedoo]

At first felt like a April fools.