AMD has announced some significant changes and adjustments to its future GPU lineup, including the launch of a new Radeon Pro graphics card. Going forward, the FirePro lineup is being retired altogether in favor of new, Radeon Pro branding. AMD has announced a handful of Radeon Pro WX (Workstation eXperience) cards already — the W4100, W5100, and W7100, likely based on derivatives of the company’s RX 480 Radeon GPU.
The interesting card, however, is the Radeon Pro SSG, a new Radeon Pro GPU with integrated M.2 slots for adding PCI Express-based NAND storage. According to Raja Koduri, the GPU can add up to 1TB of SSD storage connected via a PEX8747 bridge chip. According to AMD, adding a 1TB switch to the GPU is a simple way to bypass a common problem: the GPU memory limit.
Here’s the problem in a nutshell: Current desktops can scale up to 64GB, while high-end workstations can theoretically address up to 1,536GB of memory. Graphics cards, by contrast, are limited to a fraction of that — 32GB, as of this writing. Worse, the GPU can’t leverage system RAM. If you want to perform a workload on the GPU, you either have to pull it all into GPU memory or rely on the comparatively high latency PCI Express bus.
Stacked DRAM cuts power and boosts density, but it’s not cheap yet.
Moving workloads into on-card NAND flash solves the latency problem — AMD claims that it can access local memory over the M.2 interface at much lower latencies than it can pull data from the PCI Express bus. Based on what we know of the GPU’s path to RAM, that’s probably true. Less clear is whether or not there’s any kind of bandwidth advantage to this kind of access — the prototype uses a pair of Samsung 950 512GB drives in a RAID 0, giving them theoretical access to eight lanes of PCI Express connectivity. That’s still just half of a standard x16 PCI Express 3.0 slot, so latency rather than bandwidth may be the distinguishing factor here.
Developers will have to code for the SSG in order to enable support for the 1TB memory pool. In theory, it’s a really interesting idea for keeping workloads local and we could see more solutions like this — HBM2 may improve power efficiency a great deal and allow for long-term higher-memory solutions, but in the short term it’s going to be expensive and consequently limited.
Performance of the SSG when scrubbing 8K video topped 4GB/s while running the same workload with an attached 950 Pro was only capable of around 900MB/s, according to Anandtech. The SSG shown off today is just a prototype — it’s not clear if AMD will bring this specific card to market — but we think we could easily see solutions like this for professional graphics from both AMD and Nvidia as a way to boost performance in specific cases.
The one thing we don’t expect to see is a consumer equivalent. Consumer GPUs are well-provisioned as far as overall memory and don’t see the bunker-busting memory usage that workstation and HPC or professional applications can require. The majority of games still fit comfortably into 4GB buffers at 1080p, and 1080p is still the most common resolution.
Given that beta developer kits are going on sale for a cool $10,000, we don’t expect to see many of these units ship, period — but if the technology proves as useful as AMD’s demo implies, we may see Nvidia move towards this concept as well. Faster PCI Express storage and higher-end GPUs may make the pairing more attractive in the future once Vega arrives.