What’s Better Than 40 GPU-Based Computers? A Computer with 40 GPUs
spectrum.ieee.org
spectrum.ieee.org
I've always been wary of the idea of building extra ordinarily complicated hardware and rely on the software people to figure it out. It doesn't seem to work. It's too many degrees of freedom. Compilers are great at taking complex problems and fitting them to fixed solutions. Anything more is like handing off an NP-hard problem.
Your argument fails because it assumes everything is automate-able. It's not. Programming is not automate-able, at least not now. If you wait your C compiler to generate code that will run in L computers with N to M threading model you're gonna wait too long. This is what we software engineers are paid for. And it does seem to work, as explained above.
No, the history of parallel computing is a lot more chaotic than that!
Parallel, distributed or vector computing graphics have been around since the early days and there are many fancy parallel processors architectures that never did well exactly because no ones knew how to program them. The Tera architecture (128 threads) comes to mind[1]. The Cell that was used in the playstation 3. It has 1 general purpose core and 8 accelerators that were vastly underexploited by game developers. The greatly anticipated Intel Larrabee (32 cores) never made it to the market. And even today the fate of the Xeon Phi (about 60 cores) remains very uncertain.
In 2019 parallel computing is reserved to the same restricted set of easy to parallelize applications than 40 years ago (numerical simulations, 3D rendering, database servers and maybe few others) + deep learning. But most software is sequential, and most people would not notice if they had 1 or 2 cores instead of 8 or 16. That's because writing parallel programs is inherently and debugging is hell.
The drive to multi-core design wasn't an epiphany that it was better, it was forced by the fact that we'd really reached the end of the road when it came to single core performance. To get performance we've been forced in that direction and it's still difficult. If you want to get performance on an Intel multi-core CPU you need to actually worry about concurrency which is already a serious complication versus writing software for a single threaded design. If you want to write software for a GPU not only are you going to need to worry about concurrency, you're probably going to have to actively think seriously about memory management as well as how to express MIMD constructs and often you're even talking about needing to go specifically to OpenCL to get that performance.
Every time we add a hardware feature or concept that is not automatically handled by a compiler we create a new difficult problem for software engineers. It makes software harder to write, and very often those features go completely unused. Which is why it's going to be so important for benchmarks to really get into the meat and potatoes of what performance you will really get.
Let me give you a concrete example - an Intel Stratix 10 can claim 10 TFLOPS of compute power on their website. The way they come up with that number is taking the number of DSP blocks on the FPGA, timesing that by 2 because they're Multiply-Accumulates and then timesing that by the maximum frequency the silicon can support which they (used to) claim is 1GHz. For the largest part that's 5760 x 2 x 800MHz = 9.2TFLOPS. (The reason for the discrepancy is they've found out the DSP block won't run at 1GHz but they haven't updated their headline marketing).
Now, I've got friends who work on Intel FPGAs. Firstly, it'd be close to a miracle if you got 80% utilization. Secondly, frankly the chip literally won't work at 800MHz when full, it'll most probably run at 400MHz. So before you've even considered your algorithm, you've lost 60% of your FLOPS. Now onto the meat and potatoes, to compile a program for that device takes atleast 24 hours. So even accessing 40% of the peak performance requires such a huge amount of effort there will be a handful of companies that ever acheive that.
That is why real benchmarks are important. It's not about peak performance, it's about what you can actually achieve with your design.
Edit: Nevermind, fixing that is the point of the article; I should stop skimming these things
>"One solution would be to use standard chip-making techniques to build all 40 GPUs on the same silicon wafer and add interconnects between them. But that’s the philosophy that killed Amdahl’s attempt in the 1980s. There is always the chance of a defect when you’re making a chip, and the likelihood of there being a defect increases with the size of the chip. If your chip is the size of a dinner plate, you’re almost guaranteed to have a system-killing flaw somewhere on it.
So it makes more sense to start with normal-sized GPU chips that have already passed quality tests and find a technology to better connect them. The team believes they have that in a technology called silicon interconnect fabric (SiIF). SiIF replaces the circuit board with silicon, so there is no mechanical mismatch between the chip and the board and therefore no need for a chip package."
The problem they now have is the heat that is emitted from such a beastly contraption.
> Edit: Nevermind, fixing that is the point of the article; I should stop skimming these things
The only reason why I'm in a better boat than you is because I had followed a discussion over at Reddit on this very story already. Mistakes when reading articles are pretty common, and don't worry about it. Anyone who gives you much trouble over that is just asking for Karma to smite them. Most of us are here to help each other understand the story.
I wonder if this is something like Intel's Foveros 3D chip stacking, where the wafer is treated more like a "PCB" (with the traces and such for signals/power) to the CPU chips bonded on top?
Basically, a "PCB" made from silicon, where the components mounted on it are also bare CPU dies?