Intel Xe-LP GPU Architecture Deep Dive | Hacker News Reader

Intel Xe-LP GPU Architecture Deep Dive

anandtech.com

50 pointsm3at17 comments

> The downside is that it also means that Intel is the only hardware vendor launching a new GPU/architecture in 2020 without support for the next generation of features, which Microsoft & co are codifying as DirectX 12 Ultimate. The consumer-facing trade name for feature level 12_2, DirectX Ultimate incorporates support for variable rate shading tier 2, along with ray tracing, mesh shaders, and sampler feedback. And to be fair to Intel, expecting ray tracing in an integrated part in 2020 was always a bit too much of an ask. But some additional progress would always be nice to see. Plus it puts DG1 in a bit of an odd spot, since it’s a discrete GPU without 12_2 functionality.

Not really on a good starting foot.

If it supports GVT-g like their integrated graphics then it may still be useful as a cheaper alternative for virtualized systems. Otherwise you either have to buy multiple GPUs or the professional lines of AMD/Nvidia

Not many people are going to be playing high-end games on the integrated graphics anyways, even with 2x performance improvement over the previous generation Intel graphics. At this point, the Xe graphics are going to be just slightly better than the recently released AMD APU integrated graphics. If this limitation is still true on their discrete card targeted for gamers (not the DG1), then that would be a poor showing.

Those buying P5 and XBox Series X games will.

Huh? I wouldn’t call AMDs architecture integrated. It’s a UMA with a discrete GPU.

The most analogous thing is a laptop with a very nice GPU.

meh. Ray tracing is hype. I don't think it's omission is that big of a deal.

Hollywood is of a different point of view.

https://home.otoy.com/

https://www.disneyanimation.com/technology/innovations/hyper...

Also a standard feature of PS 5 and XBox Series X.

The price premiums for hardware with ray tracing hardware is still high and only when it becomes common place so as to not command a higher price will your average gamer care. There’s also always the next gen for Intel to add it.

Neither of those are for realtime though. Ray tracing have been used in rendering for animations and movies a lot.

Real-time ray tracing is hype.

And I agree - any arguments that "the next gen consoles have it" is specious at best; they always front load next gen consoles because there's usually a very, very long tail where they're essentially obsolete PCs but companies still have to churn out games for them. The PS4 was released in early 2013 - that's seven years, or almost 5 Moore generations old.

The games released with RTRT right now are not going to be great, but in three years? Probably showing up in AAA titles, and they'll all be designed around the PS5's level of capabilities. But by then Intel will have revved their graphics card to the next generation (as will everyone else), and all of the "hardcore" gamers will be picking up new graphics cards anyways.

That said, there are plenty of reasons to be skeptical of Intel's hardware's capabilities and priorities given how much they're trying to spackle the gap between being a server ML/NN play and a gaming device, their complete and utter management failures towards non-core market hardware (Intel's history is paved in abandoned hardware, up to and including the Xeon Phi) and them having little to no decent experience delivering SPMD programming models for their hardware as is now 100% bog standard in GPU land...

It's in all the next-gen consoles and in most of nvidia's current parts and (afaik) in all of amd and nvidia's future parts. At this point instead of a hyped up vendor-specific feature, it's looking like a feature many games will be using which will leave intel users out in the cold.

On the other hand, nobody presently expects to run modern games on integrated GPUs, so...

Some other 12_2 features are pretty important for VR.

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.

David 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.

To clarify: I don't mean to say that GPUs will become superior to CPUs.

Just that CPUs are a mature technology that seem unlikely to have any dramatic evolution any time soon. GPUs in contrast, are a lesser understood architecture with significant amounts of experimentation still being done on them. Its a more exciting field to keep an eye on.

After all, CPUs are latency optimized, GPUs are bandwidth optimized. They're made for different kinds of problems.

Apple's going to be entering the mid to high range soon, and they use Tile-Based Deferred Rendering. Very different indeed.

From 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 it does not seem that different from what we have now.

It'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/wwdc2020/10631

Reply on news.ycombinator.com