Thursday, December 14th 2017
Mystery AMD APU with 1,792 SP Shows Up on SANDRA Database
A mysterious AMD APU showed up on SiSoft SANDRA online database, featuring a massive integrated graphics. The chip reports itself to SANDRA as "AMD Fenghuang Raven," and is likely a semi-custom chip being tested by an AMD engineer in the course of its development. SANDRA reports the integrated graphics component as "AMD 15FF Graphics," featuring 1,792 stream processors across 28 compute units, 555 MHz engine clock, and 2 GB of video memory with 182.15 GB/s memory bandwidth. The result doesn't put out too many details about the CPU component, except its 2.40 GHz clock speed. The iGPU scored 98 points on SANDRA graphics tests with Direct3D 11 API, and 39.99 GB/s observed score.
27 Comments on Mystery AMD APU with 1,792 SP Shows Up on SANDRA Database
As for 'just add a gpu': I've worked in a datacenter managing large scale compute cluster systems. You do not just 'add a Gpu' to these systems. Downtime to do so costs money, and risk of hardware failures due to power cycles is very real, and the calculations running on the cluster might run for months.
Instead what they do with clusters is spec a cluster to meet expected tasks for a period of several years going forward. Then after several years they look at whether a major overhaul ie 'just add a gpu' to all or some nodes is needed. Most of the time they will be replacing cpu and gpu in such a major overhaul, so why not replace an apu?
As for Apus specifically: they are SOCs nowadays. That means less external circuitry which means cheaper to make each node and at least somewhat cheaper to operate for the same FLOPS bang. For a medium size cluster if you can knock 10% off your electricity bill or cooling bill you will save 100's of thousands of dollars. If AMD does produce an HPC APU you can be sure it will get buy in, you will see it used in some production clusters.
And why else might anyone be interested in APUs over CPU+PCie+GPU? One good reason is shared registers between cpu and gpu. Bandwidth between cpu & gpu for data sitting in shared registers is absolutely astronomical: far far higher than when the PCIe bus is in the way.