http://www.tomshardware.com/news/amd-vega-workstations-wx9100-pro-ssg,35128.html
AMD claims the Radeon Pro WX 9100 will double the performance of its previous-generation W9100, as well as fold in new functionality associated with the Vega architecture. For example, HBCC allows Vega 10 to fall back to main memory when on-package HBM2 is exhausted, enabling much larger data sets than were previously possible (even if this means incurring a performance hit from essentially paging data in and out of main memory).
AMD demonstrated loading a city model from Baahubali: The Beginning consisting of billions of polygons and rendering it using AMD's ProRender GPU renderer. Other GPUs would simply get an out-of-memory error when even attempting to load this, according to company representatives, but the WX 9100 managed to do so with a modicum of interactivity.
Building on the SSG (Solid State Graphics) technology announced at Siggraph last year, AMD will soon introduce the Radeon Pro SSG, a card that unites the SSG and Vega architecture. Performance-wise, the Radeon Pro SSG is said to be identical to the WX 9100. The difference is its on-board non-volatile storage.
The Vega architecture's HBCC means the SSG memory can function more seamlessly, allowing the solid state storage to be seen and used as local memory for loading extremely large data sets. The version of the card that AMD announced has 2TB of SSG memory on it, and that extra space is capable of up to 8 GBps reads and 6 GBps writes. While that's much slower than the on-package HBM2 the Vega 10 uses as cache, it's still faster than going across PCI Express to system memory and back.
(Score: 0, Disagree) by Anonymous Coward on Wednesday August 02 2017, @10:38PM (4 children)
If a "GPU" is used to process large dataset, it's not GPU, it's DPU, eh?
Back in the days, we called such things "math co-processor". Now, DPU, ehem, is bigger, fatter, power-hungrier, and even more costly than CPU.
(Score: 2) by takyon on Wednesday August 02 2017, @10:53PM (1 child)
It's an SSG!
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: -1, Offtopic) by Anonymous Coward on Wednesday August 02 2017, @11:17PM
This reminds me of the 90s, when electronic music keeps on inventing "new genres."
"Acid jazz" was the worst - have not heard of a more pompous, hipster douchebaggery.
(Score: 3, Informative) by FatPhil on Wednesday August 02 2017, @10:58PM
And most GPUs are based on DSP architectures, rather than math coprocessor architectures (thinking x87 or weitek). In fact the first accelerated graphics card I had back in the 90s had precisely a TI DSP chip (34020 IIRC) doing its acceleration.
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 2) by kaszz on Wednesday August 02 2017, @11:44PM
I think the difference to a Floating Point Unit (FPU) or Digital Signal Processor (DSP) is that these GPUs are optimized for matrix and vector data with specific optimizations for graphics presentation.
To make them useful however a standardized and open API is needed.
OpenCL maybe works..