Intel Publishes Icelake "Gen 11" Graphics Architecture Overview

[phoronix]

Phoronix: Intel Publishes Icelake "Gen 11" Graphics Architecture Overview

While Intel was building up anticipation for their Xe Graphics dedicated hardware last night in San Francisco with their inaugural "Odyssey" event, quietly hitting the Intel servers is an architecture overview for their Icelake "Gen 11" graphics...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-Gen11-Graphics-Overview

[andrei_me]

About the TeraFLOP thing, where does the Raven Ridge get? Surely this will be used for Intel marketing

[bridgman]

About the TeraFLOP thing, where does the Raven Ridge get? Surely this will be used for Intel marketing

I think the corresponding number for Raven Ridge (2400G) is 1.76 TF, but it's not a competition

[andrei_me]

And more or less the same for the mobile parts, 1.66 for Ryzen 2700U and 1.83 for Ryzen 2700H according to wikipedia

[nranger]

I'm generally excited for more AMD based laptops and good integrated graphics, and as I have funds available I'll probably be upgrading to one.

That being said these look like really good improvements on the Intel side. Their iGPUs in the Core iSeries era seem to have been computationally efficient, good at filling their pipelines, and decent at video streams. In other words the silicon appears to be "smart" but relatively narrow. These changes for Gen11 add more raw horsepower and allow more of the ALUs to keep busy with better caches, better access to data, etc.

Don't be surprised if framerates in games double generation over generation from Intel. Still nowhere near dGPU territory, but impressive nonetheless.

[esmth]

Interesting how they're changing the way the DRAM interface is stated. 4x32bit is looking a lot more GDDR like...

[coder]

these look like really good improvements on the Intel side. Their iGPUs in the Core iSeries era seem to have been computationally efficient, good at filling their pipelines, and decent at video streams. In other words the silicon appears to be "smart" but relatively narrow. These changes for Gen11 add more raw horsepower and allow more of the ALUs to keep busy with better caches, better access to data, etc.

What strikes me about Intel's approach is that it's much more thread-intensive. IMO, this should make Intel GPUs more applicable to general-purpose compute workloads - not only those that vectorize well.

The current doc doesn't specify this about Gen11, but recent generations were dual-issue, 7-way SMT. So, a 64-EU processor can simultaneously execute instructions from 128 of 448 total threads. And Intel's GPU should have far higher throughput of general-purpose integer instructions. All of this is to say that you could realistically anticipate decent throughput from general-purpose workloads on it, so long as they have enough concurrency.