AMD Reveals More Vega GPU Architecture Details
AMD's embargo on the Vega architecture expired this morning. Unfortunately didn't get much information in advance and thus just a short article this morning. Some of the highlights from the Vega architecture details sent to the press include:
- HBM2 backs the "High Bandwidth Cache" (HBC; what is basically the vRAM) and delivers 256GB/s per stack, twice that of HBM1.
- Vega comes with a Draw-Stream Binning Rasterizer as similar to a tile-based approach to rendering, leading to more efficient shading of pixels.
- The GPU has 512 TB of virtual address space.
- Vega will be able to interface with NAND Flash or 3D X-Point SSDs over PCI-E.
- Render back-ends are now clients of the L2 cache.
- Significantly improved compute units.
- Rapid Packed Math is a new feature for clumping multiple 16-bit operations between 32-bit registers.
- Vega when running DOOM on Windows is estimated to perform between a GTX 1070 and GTX 1080.
Vega is still on track for release later in H1'2017. The many changes with the Vega GPU architecture are exciting as finally having a big step forward since GCN, though all of the changes due raise some concerns. Vega is launching within a few months while AMD has yet to publish any of their AMDGPU or RadeonSI Gallium3D changes for bringing up Vega. Hopefully they have a lot of it queued up internally and just waiting on legal review / clearance to publish; just really hoping it won't be a bumpy road for Vega on Linux when the launch happens. Given no code is out yet, you'll almost certainly need to use third-party repos and unofficial kernels if you aren't on a rolling-release distribution, but hopefully the code will be in good shape and that AMDGPU-PRO will also be in good standing similar to NVIDIA's binary Linux driver on launch days. Of course, once Vega GPUs are available there will definitely be Linux benchmarks on Phoronix.
Ve.ga is also now the promotional site for the new architecture.
28 Comments
