Traditional compute? The cores were too weak.
Number crunching? Okay-ish but gpus were better.
Useless stuff.
I'd love to hear about what didn't work. OpenMP support seemed ok maybe but OpenMP is just a platform, figuring out software architectures that's mechanistically sympathetic to the system is hard. It would be so interesting to see what Xeon Phi might have been if we had Calcite or Velox or OpenXLA or other execution engine/optimizers that can orchestrate usage. The possibility of something like Phi seems so much higher now.
There's such a consensus around Phi tanking, and yes, some people came and tried and failed. But most of those lessons, of why it wasn't working (or was!) never survived the era, never were turned into stories & research that illuminates what Phi really was. My feeling is that most people were staying the course on GPU stuff, and that there weren't that many people trying Phi. I'd like more than the heresay heaped at Phi's feed to judge by.
- Very bad performance at existing x86 workloads, so a major selling point was basically not there in practice, because extracting any meaningful performance required a software rewrite anyway. This was an important adoption criteria; if they outright said "All your existing workloads are compatible, but will perform like complete dogshit", why would anyone bother? Compatibility was a big selling point that ended up meaning little in practice, unfortunately.
- Not actually what x86 users wanted. This was at the height of "Intel stagnation" and while I think they were experimenting with lots of stuff, well, in this case, they were serving a market that didn't really want what they had (or at least wasn't convinced they wanted it).
- GPU creators weren't sitting idle and twiddling their thumbs. Nvidia was continuously improving performance and programmability of their GPUs across all segments (gaming, HPC, datacenters, scientific workloads) while this was all happening. They improved their compilers, programming models, and microarchitecture. They did not sit by on any of these fronts.
Ironically the main living legacy of Phi is AVX-512, which people did and still do want. But that kind of gives it all away, doesn't it? People didn't want a new massively multicore microarchitecture. They wanted new vector instructions that were flexible and easier to program than what they had -- and AVX-512 is really much better. They wanted the things they were already doing to get better, not things that were like, effectively a different market.
Anyway, the most important point is probably the last one, honestly. Like we could talk a lot about compiler optimizations or autovectorization. But really, the market that Phi was trying to occupy just wasn't actually that big, and in the end, GPUs got better at things they were bad at, quicker than Phi got better at things it was bad at. It's not dissimilar to Optane. Technically interesting, and I mourn its death, but the competition simply improved faster than the adoption rate of the new thing, and so flash is what we have.
Once you factor in that you have to rewrite software to get meaningful performance uplift, the rest sort of falls into place. Keep in mind that if you have a $10,000 chip and you can only extract 50% of the performance, you more or less have just $5,000 on fire for nothing in return. You might as well go all the way and use a GPU because at least then you're getting more ops/mm^2 of silicon.
I tend to think there was tons of untapped potential still on the table. And that a failure to adopt potential isn't purely Intel alone's fault. The story we are commenting on is about the rest-of-industry trying to figure out enduring joint strategies, and much of this is chipmaker provided, but it is also informed and helped by plenty of consumers also pouring energy in to figure out what's working and not, trying to push the bounds.
Agreed that anyone going in thinking Xeon Phi would be viable for running a boring everyday x86 workload was going to be sad. To me the promise seemed clear that existing toolchains & code would work, but it was always clear to me there were a bunch of little punycores & massive SIMD units and that doing anything not SIMD intensive wasn't going to go well at all. But what's the current trend? Intel and AMD are both actively building not punycores but smaller cores, with Sierra Forest and Bergamo. E-cores are the grown up Atom we saw here.
Yes the GPGPU folks were winning. They had a huge head start, were the default option. And Intel was having trouble delivering nodes. So yes, Xeon Phi was getting trounced for real reasons. But they weren't architectural issues! It just means the Xeon Phi premise was becoming increasingly handicapped.
As I said I broadly agree everywhere. Your core point about giving the market more of what it already does is well taken, is a river of wisdom we see again and again. But I do think conservative thinking, iterating along, is dangerous thinking that obstructs us from seeing real value & possibility before us. Maybe Intel could have made a better ML chip than the GPGPU market has gotten for years, had things gone differently; I think the industry could perhaps have been glad they had veered onto a new course, but the barriers to that happening & the slow down in Intel delivery & the difficulty bootstrapping new software were all horrible encumberances which were rightly more than was worth bearing together.
Most experimenters saw it as a way to have something GPU-like in terms of raw power but with no limitations charateristic of SIMT's. Like, slightly different code paths for threads doing number crunching or something.
But it turns out that it's easier to force everything into a matrix. Or a very big matrix. Or a very-very-very big matrix.
And then see what sticks.
When CPUs actually come with bandwidth and a decent vector unit, such as the A64FX, lo and behold, they lead the Top500 supercomputer list, also beating out GPUs of the day.
Why have we not been getting bandwidth in CPUs? Is it because SPECint benchmarks do not use much? Or because there is too much branch-heavy code, so we think hundreds of cores are helpful?
Existing machines are ridiculously imbalanced, hundreds of times more compute vs bandwidth than the 1:1 still seen in the 90s. Hence matmul as a way of using/wasting the extra compute.
The AMD MI300a looks like a very interesting development: >5 TB/s shared by 24 cores plus GPUs.
Math guys came up with a list of algorithms to try for a search engine backend.
What we needed was matrix multiplication and maybe some decision tree walking (that was some time ago, trees were still big back then, NNs were seen as too compute-intensive for no clear benefits). So we thought that it might be cool to have a tool that would support both. Phi sounded just right for both.
And things written to AVX-512 did work. Software surpisingly easy to port.
But then comes the usual SIMD/CPU trouble: every SIMD generation wants a little software rewrite. So for both Phi generations we had to update our code. For things not compatible with the SIMD approach (think tree-walking) it is just a weak x86.
In theory Phi's were universal, in practice what we got was: okay number crunching, bad generic compute.
GPU was somewhat similar: the software stack was unstable, CUDA just did not materialize as a standard yet. But every generation introduced a massive increase in compute available. And boy did NVIDIA move fast...
So GPU situation was: amazing number crunching, no generic compute.
And then there were a few ML breakthroughs results which rendered everything that did not look like a matrix multiplication obsolete.
PS I wouldn't take this story too seriously, details may vary.
If you don't want that, program the GPU directly in assembly or C++ or whatever. A kernel is a thread - program counter, register file, independent execution from the other threads.
There isn't a Linux kernel equivalent sitting between you and the hardware so it's very like bare metal x64 programming, but you could put a kernel abstraction on it if you wanted.
Core isn't very well defined, but if we go with "number of independent program counters live at the same time" it's a few thousand.
X64 cores are vaguely equivalent to GCN compute units, 100 or so if either in a 300W envelope. X64 has two threads and a load of branch prediction / speculation hardware. GCN has 80 threads and swaps between them each cycle. Same sort of idea, different allocation of silicon.
But what you mention is important, and also a reason for the ultimate demise of e.g. Xeon Phi. Intel surely realized they could just scale their existing Xeon designs up-and-out further than expected. Like from a product/SKU standpoint, what is the point of having a 300 core Phi where every core is slow as shit, when you have a 100 core 4-socket Xeon design on the horizon, using an existing battle-tested design that you ship billions of dollars worth every year? Especially when the 300 core Xeon fails completely against the competition. By the time Phi died, they were already doing 100-cores-per-socket systems. They essentially realized any market they could have had would be served better by the existing Xeon line and by playing to their existing strengths.
This made me wonder a couple of things-
What kind of workloads and problems is that best suited for? It’s a lot of cores for a CPU, but for pure math/compute, like with AI training and inference and with graphics, 288 cores is like ~1.5% of the number of threads of a modern GPU, right? Doesn’t it take particular kinds of problems to make a 288 core CPU attractive?
I also wondered if the ratio of the highest core count CPU to GPU has been relatively flat for a while? Which way is it trending- which of CPUs or GPUs are getting more cores faster?
Here is a relevant article: https://www.kdnuggets.com/2020/03/deep-learning-breakthrough...
And what you suggest could be done, but it would likely flop commercially if you made it today, which is why they aren’t doing it. SIMD machines are faster on homogenous workloads, by a lot. It would be a bummer to develop a CPU with thousands of cores that is still tens or hundreds of times slower than a comparably priced GPU.
SIMD isn’t going away anytime soon, or maybe ever. When the workload is embarrassingly parallel, it’s cheaper and more efficient to use SIMD over general purpose cores. Specialized chiplets and co-processors are on the rise too, co-inciding with the wane of Moore’s law; specialization is often the lowest hanging fruit for improving efficiency now.
There’s going to be plenty of demand for general programmers but maybe worth keeping in mind the kinds of opportunities that are opening up for people who can learn and develop special purpose hardware and software.
There's no intrinsic reason why multiplying matrices requires massive parallelism, in principle it could be done on few cores plus good management of ALUs/memory bandwidth/caches.
In practice, people weren’t afraid to roll up their sleeves and write CUDA code. If you wanted good performance you had to think about data parallelism anyways, and at that point you’re not benefiting from x86 backwards compatibility. It was a fascinating dream while it lasted though.
> setting up the graphics pipe
I’ve picked D3D11, as opposed to D3D12 or Vulkan. The 11 is significantly higher level, and much easier to use.
> compiling, binding
The compiler is design-time, I ship them compiled, and integrated into the IDE. I solved the bindings with a simple code generation tool, which parses HLSL and generates C#.
> No proper debugging
I partially agree but still, we have renderdoc.
Indeed, in D3D they are called “wave intrinsics” and require D3D12. But that’s IMO a reasonable price to pay for hardware compatibility.
> no cooperative matrix multiplication
Matrix multiplication compute shader which uses group shared memory for cooperative loads: https://github.com/Const-me/Cgml/blob/master/Mistral/Mistral...
> tensor cores
When running inference on end-user computers, for many practical applications users don’t care about throughput. They only have a single audio stream / chat / picture being generated, their batch size is a small number often just 1, and they mostly care about latency, not throughput. Under these conditions inference is guaranteed to bottleneck on memory bandwidth, as opposed to compute. For use cases like that, tensor cores are useless.
> there's no way D3D11 can compete with either CUDA
My D3D11 port of Whisper outperformed original CUDA-based implementation running on the same GPU: https://github.com/Const-me/Whisper/
Most ML inferencing is throttled with memory, not compute. This certainly applies to both Whisper and Mistral models.
> it certainly doesn't spell the end of CUDA
No, because traditional HPC. Some people in the industry spent many man-years developing very complicated compute kernels, which are very expensive to port.
AI is another story. Not too hard to port from CUDA to compute shaders, because the GPU-running code is rather simple.
Moreover, it can help with performance just by removing abstraction layers. I think the reason why compute shaders-based Whisper outperformed CUDA-based version on the same GPU, these implementations do slightly different things. Unlike Python and Torch, compute shaders actually program GPUs as opposed to calling libraries with tons of abstractions layers inside them. This saves memory bandwidth storing and then loading temporary tensors.
[1]: https://github.com/bevyengine/bevy/blob/main/examples/shader...
One of the features that would get compute shaders far ahead compared to now would be pointers and pointer casting - Just let me have a byte buffer and easily cast the bytes to whatever I want. Another would be function pointers. These two are pretty much the main reason I had to stop doing a project in OpenGL/Vulkan, and start using CUDA. There are so many more, however, that make life easier like cooperative groups with device-wide sync, being able to allocate a single buffer with all the GPU memory, recursion, etc.
Khronos should start supporting C++20 for shaders (basically what CUDA is) and stop the glsl or spirv nonsense.
So yet another thing where Khronos APIs are dependent on DirectX evolution.
It used to be that AMD and NVidia would first implement new stuff on DirectX in collaboration with Microsoft, have them as extensions in OpenGL, and eventually as standard features.
Now even the shading language is part of it.
Modern Vulkan is looking pretty good now. Cooperative matrix multiplication has also landed (as a widely supported extension), and I think it's fair to say it's gone past OpenCL.
Whether we get significant adoption of all this I think is too early to say, but I think it's a plausible foundation for real stuff. It's no longer just a toy.
[1] https://community.arm.com/arm-community-blogs/b/graphics-gam...
It's been awesome seeing folks like Keras 3.0 kicking out broad Intercompatibility across JAX, TF, Pytorch, powered by flexible executuon engines. Looking forward to seeing more Vulkan based runs getting socialized benchmarked & compared. https://news.ycombinator.com/item?id=38446353
[1]: https://kompute.cc/
In researching this answer, I came across a really interesting thread[1] on diagnosing performance problems with USM in SYCL (running on AMD HIP in this case). It's a good tour of why this is hard, and why for the vast majority of users it's far better to just use CUDA and not have to deal with any of this bullshit - things pretty much just work.
When targeting compute shaders, you pretty much have to manage buffers manually, and also do copying between host and device memory explicitly (when needed - on hardware such as Apple Silicon, you prefer to not copy). I personally don't have a problem with this, as I like things being explicit, but it is definitely one of the ergonomic advantages of modern CUDA, and one of the reasons why fully automated conversion to other runtimes is not going to work well.
[1]: https://stackoverflow.com/questions/76700305/4000-performanc...
SYCL builds on top of OpenCL so you need to know the history of OpenCL. OpenCL 2.0 introduced shared virtual memory, which is basically the most insane way of doing it. Even with coarse grained shared virtual memory, memory pages can transparently migrate from host to device on access. This is difficult to implement in hardware. The only good implementations were on iGPUs simply because the memory is already shared. No vendor, not even AMD could implement this demanding feature. You would need full cache coherence from the processor to the GPU, something that is only possible with something like CXL and that one isn't ready even to this day.
So OpenCL 2.x was basically dead. It has unimplementable mandatory features so nobody wrote software for OpenCL 2.x.
Khronos then decided to make OpenCL 3.0, which gets rid of all these difficult to implement features so vendors can finally move on.
So, Intel is building their Arc GPUs and they decided to create a variant of shared virtual memory that is actually implementable called unified shared memory.
The idea is the following: All USM buffers are accessible by CPU and GPU, but the location is defined by the developer. Host memory stays on the host and the GPU must access it over PCIe. Device memory stays on the GPU and the host must access it over PCIe. These types of memory already cover the vast majority of use cases and can be implemented by anyone. Then finally, there is "shared" memory, which can migrate between CPU and GPU in a coarse grained matter. This isn't page level. The entire buffer gets moved as far as I am aware. This allows you to do CPU work then GPU work and then CPU work. What doesn't exist is a fully cache coherent form of shared memory.
https://registry.khronos.org/OpenCL/extensions/intel/cl_inte...
> [1] https://community.arm.com/arm-community-blogs/b/graphics-gam...
"Using a pointer in a shader - In Vulkan GLSL, there is the GL_EXT_buffer_reference extension "
That extension is utter garbage. I tried it. It was the last thing I tried before giving up on GLSL/Vulkan and switching to CUDA. It was the nail in the coffin that made me go "okay, if that's the best Vulkan can do, then I need to switch to CUDA". It's incredibly cumbersome, confusing and verbose.
What's needed are regular, simple, C-like pointers.
Until then, it's a bit funny claim, especially considering what a failure OpenCL was (programmer's experience and fading support). Or trying to do GPGPU with compute shaders in DX/GL/Vulkan. Are they really "motivated"? Because they had so many years and the results are miserable... And I don't think they invested even a fraction of what got invested into CUDA. Put your money where your mouth is.