TECH TALK Radeon VII Is the First 7nm Consumer GPU
THERE HAS BEEN a lot of speculation over when we’d see the first consumer graphics cards built on a 7nm process. AMD launched an updated version of its Vega architecture late last year, shrunk down to 7nm, as the Radeon Instinct MI60 and MI50.
These are professional products built for machine learning and other computational workloads, but would we see anything for gamers built off the chip? It turns out the answer is yes. And soon.
AMD revealed the Radeon VII at CES, with the cards going on sale February 7 at $699. There are multiple tie-ins to the name, it seems: “VII” for “Vega 2” is one, and the Roman numeral for seven hints at the 7nm lithography. It’s probably also why AMD selected February 7 for its launch. The Radeon VII is destined to go up against Nvidia’s RTX 2080, and AMD claims it will offer comparable performance. That would require a decent boost compared to the existing Vega 64, as my benchmarks have the RTX 2080 leading by 45 percent on average in 1440p and 4K workloads.
AMD improves performance with Radeon VII in several ways. First, it will pack 3,840 stream processing cores, fewer than Vega 64’s 4,096, but they’ll be clocked at up to 1.75GHz—about 200MHz higher than Vega 64. That’s a net improvement in computational performance of about 10 percent, which doesn’t seem like much, but there are some architectural improvements as well. Vega 7nm has 13.5 billion transistors, compared to 12.5 on Vega 64, and some of the additional transistors help improve latencies and shader performance.
AMD also doubles down on memory bandwidth, by including four stacks of HBM2, giving it 16GB of VRAM and a massive 1TB/s of bandwidth. We already know Vega scales well with memory bandwidth, and the Radeon VII looks promising for workloads that stress that aspect. Overall, AMD claims it will be about 30 percent better performance on average compared to Vega 64. There are some concerns, however. Even if Radeon VII does deliver “comparable performance” to the RTX 2080, that doesn’t include ray-tracing or deep-learning workloads. AMD’s stance is that those are both early and expensive features, and it will offer hardware with ray-tracing support when it makes sense to do so. In the meantime, like all other non-RTX GPUs, the Radeon VII will be far slower in DXR games.
Radeon VII also hints at what we can expect from TSMC’s 7nm process. It’s mostly die shrink (plus about 10 percent more transistors for the above enhancements), and the resulting GPU core is 331mm2 compared to 495mm2 for the 12nm version. So, it’s not 40 percent smaller, while perfect scaling from 12nm to 7nm would have yielded something closer to 65 percent smaller. The bigger problem remains efficiency, however.
The Radeon VII is a 300W part; basically the same as the Vega 64, and 80W more than the RTX 2080. If we’re talking about price and performance parity, that’s not good, especially since the RTX includes hardware ray-tracing and deep-learning features. I haven’t had a chance to test the Radeon VII yet, but on paper it looks to be a repeat of AMD’s last two highend GPU launches: slower than Nvidia’s best, uses more power, lacking some key features, all while being months late to the party. It seems more reasonable to peg the Radeon VII as the AMD alternative to the GTX 1080 Ti, which is nearly two years old.
I suspect the Radeon VII will end up being of only marginal interest to gamers, who will continue to be better served by either Nvidia GPUs or lower priced AMD and Nvidia offerings. However, content creators may find the gobs of memory and bandwidth useful, and it’s good AMD at least has an alternative to RTX. But the true test will be Navi, which will likely feature more architectural changes and bring power use in line with Nvidia’s GPUs. Unfortunately, with Radeon VII launching in February, it’s a stretch for AMD to release a second high-end GPU in 2019. Let’s hope I’m wrong, because in 2020, it will likely be going up against a 7nm die shrink and enhancement of Nvidia’s Turing architecture.