Adrian Cochrane
To generate an entire screenful of image data live, we design special hardware that can process all vertices, triangles, & "fragment" pixels simultaneously. Or all output pixels simultaneously.
In contrast CPUs as we stuffed transistors into our microchips GPUs got conceptually simpler, so how does a modern GPU work?
In short: We stuff it full of "compute units" cycling between processing multiple such items as it fetches data via a cache hierarchy, or other causes of latency.
1/4?
Adrian Cochrane
Each "compute unit" cycles between processing multiple vertices/triangles/pixels/etc. There's several compute-units per L1data cache. There's several of those per control-unit (& L2 cache). And maybe a handful of those in the same GPU.
The cache primarily serves to hold textures & shaders, though it can serve other purposes.
Caveat here is the control flow is less efficient. The Control Unit must emit a branch if it is taken by *any* of its compute units. And needs lots of data to process.
2/
Adrian Cochrane
The interesting question is: How do we transition between the different steps of the Depth Buffering pipeline?
I'd start by parallelizing over triangles. The cache can be used to avoid duplicate compute effort as we iterate over its vertices.
Then I'd have circuitry to fork each of these threads during the interpolation step.
To output/blend results I could use atomic writes via the cache hierarchy. And use an alternate sorting control-unit to optimize this write-order.
3/3 Fin for today!
Adrian Cochrane
P.S. From all my reading on the subject, the answer to the question asked in the last toot is not publicly known. So that's all educated guesswork.
P.P.S. I'll discuss older GPU designs tomorrow! From the 1990s & 20'00's.
3.1/3 Truly fin!
Adrian Cochrane
Oh, I forgot: What would those Compute Units do?
Basically everything in 3D computer graphics boils down to a floating point multiply-sum operation. Including matrix transforms, normalizing vectors, combining lighting, interpolation, or computing texture indices & subsampling results.
There's also occasional square-roots, logarithm/exponentiations, trigonometry, etc. Thanks to the magic of calculus (Taylor Series) those can be done via a handful of multiply-adds.
4/3 Fin, I mean it!
Adrian Cochrane
@lain_7 I'll queue that topic up!
David Mankins
@alcinnz I’m kind of hoping you’ll take a look at the Connection Machine( late 80s): 64000 (2^16, actually) 1-bit processors connected in a hypercube network; each processor has an instruction set that was actually just an 8-bit truth table with result determined by the contents of three one-bit registers. Basically a GPU, only five feet on a side.
Only… it wasn’t a GPU, it was for AI.
They decided there was a big market for supercomputing, and added 2000 floating-point chips (in those days, 32-bit floating point). Now, 32 one-bit processors were responsible for the bits in a floating-point word (each processor was responsible for one bit in the word). The processors and network were a routing mechanism for moving data to the floating point chips.
Then, as now, a supercomputer was a device for turning a computation-bound problem into an I/O-bound one.
０７６萌ＳＮＳ is a social network, courtesy of ０７６. It runs on GNU social, version 2.0.2-beta0, available under the GNU Affero General Public License.
All ０７６萌ＳＮＳ content and data are available under the Creative Commons Attribution 3.0 license.