Open-Source GPU Used For Research (binghamton.edu)

← Back to Stories (view on slashdot.org)

Open-Source GPU Used For Research (binghamton.edu)

Posted by Soulskill on Wednesday January 20, 2016 @06:19AM from the strong-progress dept.

Theovon writes: For quite some time now, "open hardware" enthusiasts have had access to a number of open source CPUs, including OpenRISC. However, it wasn't until recently that there has been any kind of open source GPU. In 2014, the Vertical Research Group at the University of Wisconsin-Madison announced MIAOW. MIAOW is in many ways a clone of the AMD Southern Islands architecture and can even run some of the same binary code. Unfortunately, MIAOW is missing some key components such as video and memory systems, making it not currently possible to implement fully in hardware. For this, Nyuzi comes to the rescue. Nyuzi (formerly Nyami) has been in development since 2010 and is a fully functional open source GPU inspired by Larrabee. Although architecturally different from the SIMT architectures from AMD and Nvidia, researchers at Binghamton University and several other places have already used it to conduct research on GPUs. A paper (PDF) was published in March 2015 about this processor (one of the authors was the original founder of the Open Graphics Project), and Nyuzi (homepage) can be downloaded from GitHub.

28 comments

Min score:

Reason:

Sort:

MIAOW? by Anonymous Coward · 2016-01-20 06:24 · Score: 0

No, the audio hardware is supposed to MIAOW. How am I going to get a proper Catputer if the component developers keep assigning the names wrong?
1. Re:MIAOW? by Anonymous Coward · 2016-01-20 06:30 · Score: 0
  
  No idea. Someone should get them on the horn right MIAOW and tear them a new asshole.
Wrong audience by 110010001000 · 2016-01-20 06:41 · Score: 0

This story is way too technical to the Slashdot readers of today. We would rather talk about CS education, global warming and women in computing. Also, where has Ethan been? I miss his posts on astronomy. I can understand them real good. It is like he wrote it just for normal people like me!
1. Re:Wrong audience by Aighearach · 2016-01-20 07:14 · Score: 1
  
  I don't know where they got you new kids, but I don't remember changing any paradigms. Some of us are backwards compatible and still doing things the old way, still thinking about tech, still running weird software on breadboards and wishing we had breadboard GPUs.
  Kids these days, and their astronomology obsessions.
2. Re:Wrong audience by 110010001000 · 2016-01-20 07:27 · Score: 1
  
  You sound like a terrirst. Why do you hate our freedoms?
3. Re:Wrong audience by amigabill · 2016-01-20 07:47 · Score: 1
  
  I think it's interesting. Does that mean you are calling me an old fart, you whiny little punk? Get off my lawn!
4. Re:Wrong audience by Anonymous Coward · 2016-01-20 08:00 · Score: 0
  
  It's fine, as long as it doesn't have any 7-segments on it it can't possibly be a bomb.
5. Re:Wrong audience by Anonymous Coward · 2016-01-20 08:07 · Score: 0
  
  Don't tell the bad guys national security secrets! Who do you think you are? Hillary Clinton?
6. Re:Wrong audience by arglebargle_xiv · 2016-01-20 11:02 · Score: 1
  
  It also seems a bit sad that "Open-source CPU/GPU actually used for something!" qualifies as newsworthy...
Re:A most unfortunate name by Anonymous Coward · 2016-01-20 07:37 · Score: 0

Did the person responsible for the name actually come and personally downvote me? Or was it a kneejerk trademark slashdot "anti-sjw" downvote?
Because frankly the name sucks.
GPU by Anonymous Coward · 2016-01-20 07:46 · Score: 0

But does it run Crysis?
Bit sketchy paper... by Anonymous Coward · 2016-01-20 07:55 · Score: 0

The Nyami paper seems, after a skechy read, a bit on the hand wavy side. E.g., they claim nvidia use switch on stall, but it is pretty well known their gpus switch thread after each cycle to avoid pipeline stalls.
What a misleading article. by BarbaraHudson · 2016-01-20 07:57 · Score: 2, Informative

If you go to the web site and actually read the article, you'll see that there's no hardware - it's just a "synthesizable gpu architectural model that you can download from github". This is just about modeling a gpu architecture, and seeing how the simulated model can be improved. Considering the cost of fabbing a one-off piece of silicon, ...

The paper, "Nyami: A Synthesizable GPU Architectural Model for General-Purpose and Graphics-Specific Workloads" appeared in International Symposium on Performance Analysis of Systems and Software
Now you know why there are no pictures. "Pics or it didn't happen" :-)

--
"Transparent" is a shit show that trades on every stereotype going. A man in drag is NOT a transsexual.
1. Re:What a misleading article. by Caelius · 2016-01-20 08:09 · Score: 2
  
  Huh? They've done an incredible job here. Just because you didn't give them $100k for them to do a tape-out doesn't mean they didn't create the first open-source GPU.
  
  This is just about modeling a gpu architecture...
  There's a HUGE difference between a simulator like GPGPU-SIM and something like this, which is doing actual RTL and gate-level simulation using the actual Verilog RTL.
2. Re:What a misleading article. by BarbaraHudson · 2016-01-20 08:27 · Score: 1
  
  What part of "There is still no hardware" don't you get? So there is still no open-source GPU here, contrary to the headline "Open-Source GPU Used For Research". It should have read Simulation of open-source gpu design used for research.
  
  --
  "Transparent" is a shit show that trades on every stereotype going. A man in drag is NOT a transsexual.
3. Re:What a misleading article. by Caelius · 2016-01-20 08:36 · Score: 3, Informative
  
  They ran this on an FPGA, does that not count as hardware? Or do you want to go to all of the open-source hardware communities and tell them what they do is not actually open-source hardware?
4. Re:What a misleading article. by Anonymous Coward · 2016-01-20 09:50 · Score: 1
  
  FPGAs don't simply run "binary programs" like a CPU. You dynamically alter the physical properties of the gates to produce functional if temporary versions ASICs.
  A chair built out of Lego is still a chair.
5. Re:What a misleading article. by Anonymous Coward · 2016-01-20 10:07 · Score: 1
  
  The HDL behavioral models are synthesized and then routed (just like a PCB) by the FPGA vendor software for your specific FPGA chip. What's running on your FPGA are hardware blocks (DDR controllers, multipliers, PLLs, block RAMs, LUTs), interconnected between each other by real wires. Nothing else can be further from a browser and an OS.
6. Re:What a misleading article. by Anonymous Coward · 2016-01-20 11:39 · Score: 0
  
  It really boils down to what you think "a GPU" is.
  I'm guessing that most people on Slashdot think of "a GPU" as a physical thing of plastic and silicon which they can physically install on a computer. Under that definition, this isn't really a GPU, it's more "a design for a GPU".
  On the other hand, if you think that the physicalness is a minor detail, and that "the GPU" is an abstract entity - the realizable conceptualization of the circuits that implement it - then yes, this would be a GPU.
  It all depends on your perspective, and how you personally define what constitutes "GPUness".
7. Re:What a misleading article. by Anonymous Coward · 2016-01-20 20:42 · Score: 0
  
  As an electronics engineer who has designed VLSI ASICs, this is very much a GPU. The design can be fed into something like Synopsys along with the technology description from someone like TSMC, and a mask set for manufacture generated. Of course it's only worth doing this if you have an order for a few thousand, but that could be Kickstarted.
  The disappointment is that it is not a SIMT architecture. Larabee failed as a GPU, and the lack of SIMT is a significant reason. The performance NVidia and AMD GPUs achieve is due to the SIMT architecture, and that is what really defines a modern GPU. So any performance work done with this won't really be relevant.
8. Re:What a misleading article. by Anonymous Coward · 2016-01-21 11:01 · Score: 0
  
  Well, yes and no.
  On FPGA you don't alter gates' physical properties. When booting, the FPGA chip loads a binary blob called bitstream, which just configures if/how the FPGA resources are working. These resources are: LUTs (look-up-tables), wires, and macro-blocks. The macro-blocks are fixed-function HW blocks implemented with ASIC technology and performance, and can be connected together with the generic LUTs for more flexibility. FPGAs are also used for low-cost ASIC prototyping, where several large FPGAs are mounted on a single board and interconnected, so the ASIC design can undergo formal verification.
More info, pics, youtube, about Nyami/Nyuzi by Anonymous Coward · 2016-01-20 08:46 · Score: 2, Informative

I googled this and found this from an OGML discussion going on about this GPU. There are some screenshots and even a youtube video.
Since 2010, Jeff Bush (github, blog) has been working on an Apache-licensed open source GPU (github, home page, wiki), and he has a few other interesting github projects as well (link, link, link). The Nyuzi Processor is a fully functional GPU. It is written in synthesizable Verilog, has a functional compiler toolchain, and comes with test suites, benchmarks, the software component of 3D rendering engine, and more. Its development has been gaining momentum in discussions (link, link, Google Group) and coding projects (gsoc). It has been implemented on an Altera FPGA, and there are some videos online of it animating a rotating teapot and a Phong-shaded torus, along with the results of recently-added mipmap support. Recently, Jeff Bush got together with the founder of the Open Graphics Project, and they co-wrote a peer-reviewed publication about this GPU and some experiments they did, which was recently presented at a well-respected academic CS conference (ISPASS). Although its developer and other hobbyists are doing this for fun, academics and engineers who specialize in GPU architecture are already showing interest in using Nyuzi for their own research (e.g. link, link), which gives them finally an open platform to estimate not just cycle count but also clock frequency, energy, and circuit area effects of GPU design experiments.
1. Re:More info, pics, youtube, about Nyami/Nyuzi by Anonymous Coward · 2016-01-20 17:53 · Score: 0
  
  Wasn't Vincent van Gogh a GPU?
Re: A most unfortunate name by Type44Q · 2016-01-20 09:28 · Score: 1

Perhaps the individual is a Super Troopers fan?
Too different things by DrYak · 2016-01-20 11:05 · Score: 1

they claim nvidia use switch on stall, but it is pretty well known their gpus switch thread after each cycle to avoid pipeline stalls.
"claim switch on stall" = refers to some of the top level documentation that explains that having as much (wraps) threads as possible in flight will hide any latency due to memory access. If one wraps stalls another will hide the latency.
"Cycle threads after each cycle" = you're refering to a different section of Nvidia documentation (I think it was the annexes about fine/tuning optimisation) where the whole thing about "warps/half-warps" and the gory details there of are explain. (If I remember correctly - I'm too lazy to whip out my printouts - each instruction takes 4 cycle to work, and basically, that's why for each single decoded instruction, threads are organised as a serie of 4 waves doing the same instructions. I think having read that it's 32 threads organised as 4 groups of 8. Except some memory access had 2 cycles latency, thus half-wraps are 16 thread: the first 2 groups of 8 then the next 2. Was it shared memory that is access by half wraps ? Memory is a bit blurry).

--
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
A good progression by Hells+Ranger · 2016-01-21 04:13 · Score: 1

I remember a few year back i was a member of the OGP group.
The group was able to do a development platform for the hardware. Also a demo code was made to support VGA and a first iteration of a fix GPU was made.
At the end we were a few working on a spec for the Programmable GPU, before starting to code it to test.
What was released was another project in collaboration with the leader of the OGP. For those who don't get what the news mean :
-It's a programmable GPU
-It can be turned into hardware
-People can thinker with it and improve it
-It's a working base for hardware test in FPGA and software simulation
-The available and open code for a GPU.
-It's not commercial, it's academic
-Graduate and undergraduate student across the globe now have access to it for student project and thesis work
What it's not is:
-A graphic card you can buy
It is a pretty great news for those into the open hardware community. Albeit a latter one than most wished. But it is a game changing moment.