Comment by drbawb | Hacker News Reader

Comment by drbawb | Hacker News Reader

I watched this the other day, and I've been mulling over the observation that onboard graphics are getting more and more die-space.

On mobile, where discrete graphics are a rarity and power is a scarce resource, it makes sense to offload as much work onto the GPU as possible. Light it up, do the work as quickly as possible, then go back to sleep.

However I started wondering about what can be done for high-end desktops. I just bought an i7-6700K, and of course it's paired with a very powerful discrete GPU.

So as I understand it, effectively this means about 1/3 of that 6700K is just dark all the time. If there's any GPU work to be done it will all be offloaded to the PCIe bus, not the onboard graphics.

How can we make use of that silicon in scenarios like that? Is there any advantages an onboard GPU has that can be leveraged over a discrete GPU? (Apart from power efficiency; which in my opinion is a non-issue when thinking about PCs with 800W+ power supplies.)

---

P.S: Of course all this would be moot if I had just bought the 5820K. Live and learn.

> How can we make use of that silicon in scenarios like that?

Well, in theory you could run them simultaneously. Use one of them for compute and one of them for display. Of course, this may well be quite a stretch in practice; GPU compute in any consumer software is completely nonexistent, to a first approximation. But it's worth thinking about, as there has been a steady stream of research coming out for years about various useful things that you can do with GPU compute.

Using two GPUs for rendering (or rendering + compute) is a lot easier to do with DX12 and Vulkan - there are some benchmarks/game engines out there that can do AFR across paired GeForce and Radeon cards, for example, and it actually performs really well. So with Vulkan you could probably utilize both an Intel onboard GPU and a discrete GPU simultaneously. The question is how much leverage you can get out of that, but I bet people could come up with good uses (rasterizing font glyphs or doing some computational work, perhaps)

Exactly. As an example: when using OpenCL, both the IGPU and the discrete GPU will show up as compute devices. Currently if an application uses the GPU at all, it would often pick the one it considers the most powerful. But even today you could theoretically run benchmarks and then use the IGPU if that is sufficient, or use both at the same time.

When the iGPU is not in use, the CPU is able to stay in its thermal budget and sustain turbo speed for longer.

For some workloads though you're not maxing out the CPU or its thermal budget, so mixed GPU/CPU could still be a win.

There is a piece of software called VirtuMVP that attempts to use both the iGPU and the discrete one.

>which in my opinion is a non-issue when thinking about PCs with 800W+ power supplies

My electricity bill isn't a non-issue ;-) Also power efficiency is good from an ecological standpoint.

Texture upload to an onboard GPU are faster, iirc.

Especially when it's already in right format in the same address space. Then upload is not needed at all.

Except that GPUs don't usually read from linear "raster order" images or are very slow doing so. So even in a shared memory CPU/GPU setup, there is a texture "upload" which organizes the pixel data to a GPU cache-friendly block format.

What is die-space? / What does die-space mean?

It's covered in the video too (with visuals that illustrate what your parent is talking about), but the "die" is the slab of silicon that the CPU itself is on. https://en.m.wikipedia.org/wiki/Die_(integrated_circuit)

Reply on news.ycombinator.com