• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

Article Imagination B-Series GPU - first MCM Multi-GPU acting as single GPU

Krteq

Senior member
May 22, 2015
922
525
136
AnandTech - Imagination Announces B-Series GPU IP: Scaling up with Multi-GPU
The most importantly, this new multi-GPU system that Imagination introduces is completely transparent to the higher-level APIs as well as software workloads, which means that a system running a multi-GPU configuration just sees one single large GPU from a software perspective. This is a big contrast to current discrete multi-GPU implementations, and why Imagination’s multi-GPU technology is a lot more interesting...

The most interesting aspect of the multi-GPU approach is found in the BXE series, which is Imagination’s smallest GPU IP that purely focuses on getting to the best possible area efficiency. Whilst the BXT and BXM series GPUs until now are delivered as “primary” cores, the BXE is being offered in the form of both a primary as well as a secondary GPU implementation. The differences here is that the secondary variant of the IP lacks a firmware processor as well as a geometry processing, instead fully relying on the primary GPU’s geometry throughput. Imagination says that this configuration would be able to offer quite high compute and fillrate capabilities in extremely minuscule area usage...

...Imagination’s current highest-end hardware implementation in the BXT series is the BXT 32-1024, and putting four of these together creates an MC4 GPU. In a high-performance implementation reaching up to 1.5GHz clock speeds, this configuration would offer up to 6TFLOPs of FP32 computing power.


This seems to be very interesting twist in GPU development... waiting eagerly for AMD's/NVIDIA's MCM GPU implementations :)

//Updated - Thx to @NTMBK

First B-Series architecture/IP will be implemented in Fantasy series GPUs by Chinese manufacturer Innosilicon
The Innosilicon’s “Fantasy” series, soon globally available, powers the 5G infrastructure deployment with high performance, high security, and high reliability. Its built-in physically unclonable iUnique Security PUF information security encryption technology that improves data security and anti-attack ability, along with a series of global leading technologies like 16Gbps GDDR6 high-speed video memory, HDMI2.1 8K display and Cache-consistent Innolink Chiplet, will make their debut in the Fantasy GPU.

“Fantasy” GPU series, featuring floating point and intelligent 3D graphics processing, fully customized multi-level pipeline computing core, together with high-performance rendering and intelligent AI computing, can be interconnected with multiple chips for processing and flexibility enhancement. As a result, it is ideally suited for the 1080P/4K/8K high quality display for desktop applications, supporting VR/AR/AI, multi-server cloud desktop, cloud games, cloud office and other big data graphics application scenarios in the 5G era.
 
Last edited:

NTMBK

Diamond Member
Nov 14, 2011
8,841
1,916
136
Imagination subforum plz

In all seriousness- it's pretty interesting. TBDR is naturally a better fit to a multi-chip GPU, but I don't know if Imagination really has the chops to compete in their target markets. Their currently unlaunched top end product only has 6TFLOPs, while NVidia's already shipping 3090 has 36TFLOPs. Outside of the Chinese market for non-US IP, I can't see this being a big success.
 

guidryp

Senior member
Apr 3, 2006
444
292
136
This doesn't actually seem to be MCM. But multiple GPU blocks on the same SoC:

We’ll get into the details of the scaling in the next page, but currently the B-Series multi-GPU support scales up to 4 GPUs. The other interesting aspect of laying down multiple GPUs on an SoC, in contrast to one larger GPU, is that they do not have to be adjacent or even near each other. As they’re independent design blocks, one could do weird stuff such as putting a GPU in each corner of an SoC design.
 

guidryp

Senior member
Apr 3, 2006
444
292
136
I expect we will see ZERO MCM implementations.

The devil is in the details. Off and on chip connectivity tend to be significantly different. This seems more aimed at on chip scaling their designs which seems to be moribund and lacking in native scaling.
 

soresu

Golden Member
Dec 19, 2014
1,491
703
136
No surprise that RT acceleration isn't in this release either, delayed until C series generation.

Albeit I don't expect to see much in the mobile arena to take advantage of it until at least the first C series implementations land.

Having said that I have yet to hear of a single implementation using the A series, has anyone hear heard of one?
 

NTMBK

Diamond Member
Nov 14, 2011
8,841
1,916
136
I expect we will see ZERO MCM implementations.

The devil is in the details. Off and on chip connectivity tend to be significantly different. This seems more aimed at on chip scaling their designs which seems to be moribund and lacking in native scaling.
The first implementation is the Innosilicon Fantasy, and it mentions chiplets in the press release:

The Innosilicon’s “Fantasy” series, soon globally available, powers the 5G infrastructure deployment with high performance, high security, and high reliability. Its built-in physically unclonable iUnique Security PUF information security encryption technology that improves data security and anti-attack ability, along with a series of global leading technologies like 16Gbps GDDR6 high-speed video memory, HDMI2.1 8K display and Cache-consistent Innolink Chiplet, will make their debut in the Fantasy GPU.

“Fantasy” GPU series, featuring floating point and intelligent 3D graphics processing, fully customized multi-level pipeline computing core, together with high-performance rendering and intelligent AI computing, can be interconnected with multiple chips for processing and flexibility enhancement. As a result, it is ideally suited for the 1080P/4K/8K high quality display for desktop applications, supporting VR/AR/AI, multi-server cloud desktop, cloud games, cloud office and other big data graphics application scenarios in the 5G era.

 
Last edited:
  • Like
Reactions: Krteq and Mopetar

Mopetar

Diamond Member
Jan 31, 2011
4,850
1,141
136
I can't wait to see benches on scaling with the MCM designs.
It's going to be the compute or professional workloads that scale best. Just as with Threadripper, I don't see this changing much in terms of gaming workloads. If anything it just makes it less expensive to deliver a high-end gaming GPU because it no longer needs to be a monolithic die.
 
  • Like
Reactions: Tlh97

lobz

Golden Member
Feb 10, 2017
1,401
1,673
106
Also, thanks for the massive amount of R&D that the companies pour into interconnect tech, multi GPU is already very good for general computing and hosting VMs - therefore for cloud too - and no doubt we will see the first MCM solutions in the next 1-2 generations from AMD and NVIDIA in the professional market.

When will it viable for games and such workloads, that's another question. I don't think it's within 5 years, but we'll see.
 

ThatBuzzkiller

Senior member
Nov 14, 2014
988
129
106
It's probably using SMIC's N+1 process ...

It's not meant for the global market anyways. It's just a bone to throw for Chinese market to the likes of Tencent or or the other Chinese developers ...
 

NTMBK

Diamond Member
Nov 14, 2011
8,841
1,916
136
It's probably using SMIC's N+1 process ...

It's not meant for the global market anyways. It's just a bone to throw for Chinese market to the likes of Tencent or or the other Chinese developers ...
I still hope Anandtech can get it in the Bench!
 

ThatBuzzkiller

Senior member
Nov 14, 2014
988
129
106
I still hope Anandtech can get it in the Bench!
You'll be disappointed since it can't run D3D or even a good portion of AAA Vulkan games. The only benchmarks Anandtech will be doing when they get their hands on the hardware will be Android games or future/updated Chinese games with Vulkan or some other in-house proprietary low level gfx API ...

It would not surprise me that the CCP are thinking about coming up with a CUDA alternative either just like the other vendors ... (AMD/ROCm & Intel/oneAPI)
 

NTMBK

Diamond Member
Nov 14, 2011
8,841
1,916
136
Imagination have a blog post:

Since then, however, everything has changed and A-Series higher-end capabilities have proven ground-breaking for Imagination. We have now launched the IMG B-Series GPU family, the latest generation of PowerVR (our 11th!) which has 70% higher compute density compared to existing desktop GPUs, and advanced multi-core features.

And I’m excited to say that we have five customers targeting the desktop, performance laptop and cloud spaces using the PowerVR GPU architecture.

One of them, Innosilicon, has announced their plans today alongside the B-Series launch. Innosilicon has integrated the newly launched IMG B-Series BXT multi-core GPU IP in PCI-E form-factor GPU SoCs for desktop and data centre applications. Our companies are also exploring a long-term strategic partnership to bring even more powerful GPU SoCs to the market.

For all the hype about the desktop, disappointing to see no support for Windows or DirectX listed- or even full OpenGL, instead of just ES.
 

ThatBuzzkiller

Senior member
Nov 14, 2014
988
129
106
For all the hype about the desktop, disappointing to see no support for Windows or DirectX listed- or even full OpenGL, instead of just ES.
Told ya so but even if it doesn't have D3D, Chinese ISVs will have no choice but to either use Vulkan for portability or potentially a proprietary low level gfx API for maximum performance because the CCP stopped assuming that Windows will keep playing a role in their future ...

OpenGL is a dumb API anyways since just about every Khronos member wants to kill it off and replace it with Vulkan ...
 

Headfoot

Diamond Member
Feb 28, 2008
4,444
636
126
For desktop/server tbh I expect it will end up being a glorified encoder/decoder and multi desktop VDI for cloud. Just shippin' virtual desktops to as many users as possible for cheap in some *nix flavor, or running decode/encode en masse for cloud services. I can't imagine any serious gaming or ML/compute use given the lack of any history of drivers or software integration for Imagination GPUs for this. For desktop, I imagine it won't even be able to successfully run games at all. The only thing I expect this chip to actually do well is compete in Android phones.
 

PhoBoChai

Member
Oct 10, 2017
63
157
76
Ppl should not rush to judge based on "6 TFlops". TBDR and Imgtech's rendering methodology is extremely optimized for graphics. Would not rule them out competing against much higher paper Tflop specs from AMD & NV.

Only major issue is brand recognition, availability and definitely drivers considering no major game engine caters to Imgtech natively.
 

ASK THE COMMUNITY