5 ms·
CPU cores are pretty much a done deal IMO (unless Nuvia's hype is real. I guess we'll see about that soon). For the last 10, maybe 20 years, CPUs have simply gr
by dragontamer 6y ago
CPU cores are pretty much a done deal IMO (unless Nuvia's hype is real. I guess we'll see about that soon). For the last 10, maybe 20 years, CPUs have simply grown bigger and have more or less done the same features.
CPUs simply push wider SIMD (SSE -> AVX -> AVX512), bigger out-of-execution buffers, larger registers, deeper branch prediction, more cache. The biggest advancement seemed to be Intel's uop-cache in Sandy Bridge (a concept that was first tried in the Pentium4, so not really a "new" innovation).
Even ARM chips are more of the same. Wider SIMD (see A64fx: #1 supercomputer in the world for 512-bit SVE), deeper out-of-order execution, better branch prediction. Its rare to see something new
----------
In contrast, whenever I see GPU-architectures, everything is so hugely different. True, GPUs all have a "SIMD focus", but there's so many exciting things about SIMD that just haven't been explored yet.
NVidia has Tensor-cores, Raytracing cores. It was already crazy that NVidia had a full matrix-multiplication implemented in its FP16 processor in Volta, but Ampere is now pushing sparse matrix multiplication at the chip level (!!).
AMD GCN and RDNA push GPU/SIMD architecture in a different direction. While a chunk of RDNA can be described as "more of the same", the "SubVector execution" feature grossly changes how Wave64 wavefronts are executed to save on register space and potentially have a better memory access pattern. Innovation continues in the AMD space for sure.
Intel finally enters with their GPU / SIMD architecture, and they start of with... SIMD8 + SIMD2 (instead of the Wave32 of RDNA/NVidia or Wave64 for GCN). Finally, Intel is minimizing hardware scheduling (saving on die area), and focusing on software to schedule the instructions.
Three very, very different architectures. Three very different ideologies for how SIMD compute / GPUs should be built. The SIMD / GPU architecture space remains more innovative than the CPU-architecture space by a longshot. Its exciting to see.
--------
I'm not saying that CPU-advancements are dead per se. AMD did make chiplets a hot commodity (but chiplets were already experimented with Power5, and chiplets seem like a natural evolution to dual-socket or NUMA architectures... just cheap enough for a consumer now). Intel's new Tremont Atom core has this cool "dual-decoder" design (where one decoder works on one-branch, and the 2nd decoder works on a 2nd branch in parallel). There are certainly new ideas being tested in the CPU world. But nothing as majorly game changing as what Intel / AMD / NVidia are doing to the GPU space. (Or Google if we include TPUs in the discussion. A TPU isn't SIMD, but it shares some degree of similarity, especially to NVidia's Tensor cores)
I guess big.LITTLE (and Intel's Lakefield platform) was a big change in the last 10 years. But that hasn't really seen much success in the Desktop/Server markets.
- Veedrac 6y agoApple's going to be entering the mid to high range soon, and they use Tile-Based Deferred Rendering. Very different indeed.
- SahAssar 6y agoFrom what I've heard in online discussions (admittedly a bad source, but since it is not released we don't have a lot) and based on https://en.wikipedia.org/wiki/Tiled_rendering https://en.wikipedia.org/wiki/Tiled_rendering it does not seem that different from what we have now.
- Veedrac 6y agoIt's a pretty big architectural difference, and the effect travels all the way through the API. Apple had a few talks on this in their recent WWDC. https://developer.apple.com/videos/play/wwdc2020/10602 https://developer.apple.com/videos/play/wwdc2020/10602 https://developer.apple.com/videos/play/wwdc2020/10631 https://developer.apple.com/videos/play/wwdc2020/10631
- awalton 6y agoDavid Kanter makes an excellent point about why CPUs still aren't going away any time soon - nobody's building GPUs with arbitrarily upgradeable memory, and there are still more larger-than-memory problems you want these systems to be able to tackle. As soon as your data set is too big for memory, you're basically forced to head to a CPU to coordinate things. And as big as some GPU memories have grown, the ML programming models are vastly bigger still, so there's still a lot of reasons you want a CPU in these systems. So naturally the market is currently trying to find a middle, which is exactly where Intel's been heading if you've been paying any attention. That's the total explanation on why CPUs have been getting wider and wider - it just makes sense for them to, given the transistor budgets and the fact that just stamping more CPUs on a core is already showing signs of diminishing returns (to the point where you have to use an approach like chiplets to manage thermals appropriately, even). The laptop market being repressed by Microsoft's stone-aged kernel CPU core management holding back designs like Lakefield (which are now completely old hat for phone-class processors) for a decade is a completely different discussion and really goes out of band with what we're talking about here though.