Monday, January 2nd 2017
Ominous ve.ga Website Surfaces with a Word Cloud
An ominously named website ve.ga surfaced on the web, pointing to the placeholder of a website AMD dedicated to its upcoming "Vega" GPU architecture. This placeholder has a word-cloud (or tag cloud), which is a 3D spacial clump of keywords/tags sorted by popularity. The words in this word-cloud could spell out key features of the "Vega" architecture, which AMD is expected to unveil later this week.
Here's what we make of these keywords:
Source:
3DCenter.org
- 8x Capacity/stack: This probably refers to the fact that HBM2 memory has 8 times the capacity per DRAM stack, as the first-generation HBM. While HBM1 maxed out at 4 GB over a 4096-bit interface (1 GB per stack), on the Radeon R9 Fury series; HBM2 allows you to cram up to 32 GB of memory (or 8 GB per stack), over a 4096-bit interface. This also means that whatever Vega-based AMD is going to launch, will come with HBM2 memory.
- 2x Bandwidth Per Pin: Refers to the increase in bandwidth per pin of HBM2. With a memory clock of 500 MHz, 4096-bit HBM2 could reach 1 TB/s. This gives AMD the cost-saving flexibility of deploying a 2048-bit memory bus with 8 GB of memory, and yet ending up with 512 GB/s bandwidth (more than that of the TITAN X Pascal).
- High Bandwidth Cache: This is a very curious phrase. Perhaps it means that AMD managed to increase the data throughput of the GPU's on-die caches, or, given the way this is worded, it could be a larger cache faster than the video-memory, which the GPU can use as a large scratchpad. The fact that another keyword in this cloud refers to a "high-bandwidth cache controller" adds weight to this theory.
- 512 TB Virtual Address Space: This adds even more weight to the theory that the HBM2 video memory, and the on-die L2 cache may not be the only physical memory on "Vega" graphics cards. The GPU probably needs a vast address space of up to 512 TB, to juggle data between the VRAM, caches, and the host.
- 4x Power Efficiency: This is probably just a 4x performance/Watt increase over the "Fiji" silicon, and shouldn't come as a surprise, given the new 14 nm process, and the energy-efficiency improvements AMD already made with "Polaris."
- Next Generation Compute[r] Engine: This probably refers to an upgraded compute unit (CU) design, which is at the heart of the 4x performance/Watt increases
- Primitive Shaders: If this is what we think it is, it's big. Most modern shader operations can be simplified into smaller, simpler operations.
43 Comments on Ominous ve.ga Website Surfaces with a Word Cloud
It is the same nvme hybrid feature as in firepro cards imho.
Disclaimer: I'm sure there is worse, I just haven't seen one personally.
On topic: horrible advertising, BUT ... at least they are advertising this time. Always remember: bad press is better than no press. And them throwing the gauntlet not at nVidia's current gen, but its next gen GPU is quite ballsy, although unlikely to pan out.
I like how AMD is playing this, they are drawing attention and playing out their underdog role pretty well. Nothing but pure benchmarks and price/performance ratios will convince me, but this will help their position with the uncouth masses.
EDIT: So bored at work ...
@Xzibit by "first dibs" on HBM 2, what that means is they get first pick at HBM 2 production chips.
consumer gradeGeForce series besides profit margins.HBM2 has been in production for a while. In that sense AMD has been stockpiling them because Nvidia products were released half a year ago. All 3 variants have been on the market long enough they are 40% cheaper now.Where is it? I tried Lenovo and a few other places to "buy" them, but there is no option for it... so......
They sold a few for HPC and that's it, I'm assuming. I sure haven't seen anyone bragging about a farm. There's been zero talk of public release that I've seen.
NVIDIA 900-2H400-0003-000 Tesla P100 12GB CoWoS HBM2 PCIe 3.0 -- Passive Cooling - $4,599.00
Supermicro GPU-NVTP100-16 Tesla P100 16GB CoWoS HBM2 PCIe 3.0 -- Passive Cooling - $6,259.99
NVIDIA 900-2H400-0000-000 Tesla P100 16GB CoWoS HBM2 PCIe 3.0 -- Passive Cooling - $5,599.00
Supermicro GPU-NVTP100-SXM Tesla P100 16GB SXM2 16GB CoWoS HBM2, NVLink - $7,999.99
NVIDIA 900-2H403-0000-000 Tesla P100 16GB SXM2 16GB CoWoS HBM2, NVLink - $6,999.00
If you want bulk try the Nvidia program
Pre-Order is in stock at call for availability now today sometime in the future.....
Point is the HBM2 exclusivity is bollocks and equally a low yield/high return strategy makes sense however you want to slice it.