AMD Vega: Founders Frontier Edition
AMD’s first Vega video card bears the “FE” badge that nVidia recently stuck on its reference cards, but changes the definition of “FE” from “Founders Edition” to “Frontier Edition.” This follows a trend of one-upmanship from AMD, where the company has also iterated on Intel’s chipset branding (B250 becomes B350, Z270 becomes X370) and prime-number labeling of CPUs. The Vega FE card is aimed at data scientists and is not a gaming-targeted device, claiming 13 TFLOPs of single-precision COMPUTE power, with roughly a 2:1 ratio of FP16 (25TFLOPs). There was no discussion as to the FP64 double-precision capabilities of Vega at time of the presentation.
AMD claims that Vega will outpace nVidia’s 2016 P100 accelerator GPU when using a Baidu benchmark for machine learning, positioning Vega at 88ms completion versus 133ms completion for the P100. This is the only test of its kind that AMD showcased, so we can’t draw definitive conclusions from the results, but it is the first look at real performance we’ve been given. The V100 isn’t present, as it won’t ship until quarter 3. Note that, of this DeepBench test result, Raja Koduri said: “I am not declaring victory here over nVidia or anyone else. This is basically stating that AMD is finally on the chart. A year ago, we would have taken just being on the previous chart. We are on the chart with a fairly impressive number.”
Initial Vega Benchmarks (AMD Internally Conducted)
As for gaming performance, AMD demonstrated two different gaming demos with Vega: Sniper Elite 4, which generally favors AMD, and High-Bandwidth Cache VRAM thrashing simulations with Rise of the Tomb Raider.
In the first test, Sniper Elite, AMD showed the game running 4K with undisclosed graphics settings and operating a framerate in the range of 60-63FPS from the on-screen monitor. (with one or two spikes over 70FPS during the ladder scene). Without knowing the graphics quality settings, we can’t actually make any comparisons to other products in the stack. It’s sort of a test in a vacuum, in this regard, but AMD has previously shown its early Vega cards as being on-par with the GTX 1080 at CES. We’d imagine that this has not changed.
The second test was a simulation of VRAM thrashing by limiting Vega to just 2GB of its high-bandwidth cache, which was done because games don’t use anything close to the 16GB total allowance on Vega.
AMD toggled HBC on and off in A/B test fashion with the 2GB limit, showing that performance bumped from roughly 56FPS AVG to roughly 72FPS AVG in Rise of the Tomb Raider. This was a simulation, isn’t real world, and is a chosen slide for an analyst meeting, but gives some idea as to what the company is hoping HBC will do. The theory is that HBC helps in VRAM-limited scenarios, should one present itself. 4K60 is the target that was set two years ago.
Vega: Probably July or Later for Gaming
Most of the presentation was focused on professional applications for Vega, though, with discussion heavily centering around deep learning capabilities and around a new introduction to AMD’s SSG line.
For gamers, this means Vega for the enthusiast market probably won’t ship until a bit later. Vega is targeted for delivery in June for professional users, so we’d anticipate a July or August timeline for the gaming market. This is speculation; it could be the case that AMD begins shipping plans for Vega gaming cards sooner than this, but based on industry trends, a post-June launch seems most plausible.
With regard to the Vega SSG, this new version of the CAD-targeted GPU comes with a 2TB SSD strapped to it, which AMD claims will better enable real-time ray tracing and 8K video workflows without timeline lag in Premiere. The demonstration was impressive, in this regard, and AMD hopes to show more in the future. We previously talked about Radeon SSG with Raja Koduri at AMD HQ, if you’re curious to learn about the original model.
AMD Ryzen + Vega Notebooks
AMD also discussed mobile deployments of their Ryzen CPUs, which will be accompanied by Vega integrated graphics chips in the near future. AMD says that these integrated Vega graphics cores will run a generational improvement of about 40% over the previous IGP solution, and claims a 50% power reduction. Ryzen mobile will include 2-in-1s, for folks interested in that market.
AMD Threadripper, AMD Epyc
As for AMD’s Threadripper CPU, that’s been officially shown and will ship in 16C/32T configurations, targeted at HEDT platforms. This would likely be competing with Intel’s newly leaked i9 lineup, and will have its first live demos at Computex this year. We’ll be there to check it out and can talk more at that time. For now, all we know is that AMD says Threadripper will expand memory bandwidth, I/O bandwidth, and focus on production-level performance and professional users.
There was also a large CPU called “Epyc” spoken of today. This is what became of Naples (32C/64T), and will allow AMD to compete in the x86 server market where Intel currently holds about 95% marketshare. Initial demonstrations included a Linux code compile at 15.7 seconds to complete on 2x Epyc CPUs, which AMD matched against 2x Intel’s E5-2699A v4 at 22.5 seconds. Both of these tests were done in dual-socket configurations. A later test matched Epyc in a single socket configuration versus Intel’s E5-2650 v4 in a dual-socket configuration, showing a complete compile time of 33.7 seconds on Epyc and 37.2 seconds on the E5 CPU.
More to come on all of this. June is looking like a busy month for AMD professional audiences.
Editorial: Steve Burke
Video: Andrew Coleman
Additional reporting: Big thanks to GN's Discord members for helping with screenshots and live discussion (support us on Patreon to participate).