Review: AMD Radeon VII

Introduction

Nvidia and AMD share what amounts to a duopoly for discrete graphics cards. Depending on which market-research bulletin you care to follow, Nvidia outsells AMD by a factor of 2:1, though with the latter having gained arguably more share through the cryptocurrency-led boom that peaked just over a year ago.

However, post-boom and things are not so rosy. Both companies have excess inventory in the channel while Nvidia's all-new RTX series has got off to a faltering start, handily exemplifying a case of potent hardware pre-empting widely-available software.

What such commentary disguises is that AMD doesn't have a play in the genuine premium end of the graphics-card market. The >$500 space, though admittedly small in volume but great for marketing, is dominated by two generations of Nvidia GPUs.

The red team is hoping the status quo changes today with the release of Radeon VII, representing the first gaming GPU to be built on a leading-edge 7nm process from industry giant TSMC.

Shipping at $699 (£650), let's examine what silicon-level magic AMD has weaved into its most powerful gaming card ever.

Analysis

The most salient point at this juncture is to explain that AMD is not harnessing its next-generation Navi architecture for Radeon VII. That simply is not ready, and won't be anytime soon, so the only recourse has been to look at other, non-gaming GPUs in the stack and choose one that most closely resembles the required credentials of a high-end gaming card.

It is no coincidence, therefore, that Radeon VII uses the latest Vega architecture - GCN 5th Generation - and footprint of the deep-learning-optimised Radeon Instinct MI50 accelerator. Though manifestly powerful on the spec sheet, it is not the finest, fullest interpretation of the Vega 20 die, because that honour belongs to the MI60, which trumps the 60-compute unit (3,840 shaders) Radeon VII and MI50 by housing the full complement of 64 CUs (4,096 shaders).

Okay, so the Radeon VII is the MI50 in gaming livery. That statement largely rings true, but there are important distinctions between GPUs because otherwise the dearer MI50 would be rendered immediately obsolete.

Not quite the MI50

Sure, these two graphics behemoths share a 13.2-billion transistor die, harness 7nm FinFET technology that leads to an impressively small 331mm² surface area, are equipped with a gargantuan 16GB HBM2 framebuffer offering a massive 1TB/s of bandwidth. Heck, they even share the same 300W TDP, so what gives?

Radeon VII has a marginally higher peak core speed of 1,800MHz (vs. 1,746MHz for MI50), though its nominal speeds are 1,400MHz base and 1,750MHz boost. To be fair, there's not much in it with respect to metrics such as single-precision GFLOPS or bandwidth. Where MI50 outshines its gaming cousin is the speed at which it can process double-precision floating-point (64 bit), where MI50 runs at half of SP speed while Radeon VII can only manage 1/4th.

The inherent accuracy of double-precision isn't required in the gaming world, so it's a miss that will not be felt, and this deliberate segmentation enables AMD to properly differentiate the two near-identical GPUs. Still, Radeon VII's otherwise muscular specification plays well with content creators who take liberal advantage of GPU-accelerated compute.

Another minor difference rests with AMD removing the PCIe 4.0 support present in MI50, replacing it with standard PCIe 3.0. Makes sense, as there are no current CPUs that support the faster standard - Ryzen 3000-series waits in the wings - and super-fast throughput is arguably more necessary in the machine-learning space.

And that's your lot from an architecture perspective. Radeon VII is effectively the extant MI50 for desktop gamers, and that means no raytracing support, unlike Nvidia's latest RTX cohort. The next question is one of how it compares against the competition.

Radeon VII vs. Radeon Vega 64

AMD's previous gaming champ, RX Vega 64, uses the same GCN 5th architecture and packs in a greater number of shaders. Yet its instruction-process throughput is actually about 10 per cent lower due to a slower peak core speed. Makes sense when you consider the manufacturing process improvements for Number Seven. The key difference, however, relates to memory. Radeon VII has twice as much and it is more than twice as fast, all accomplished by only inching up the overall board power by just five watts.

Do we really need 16GB of memory for today's games? Good question, and one that AMD answers in the affirmative for a handful of titles run at highest in-game settings at 4K. What that also means is that the vast majority of games are cosy enough within the standard 8GB present on most other high-end cards. The truth is that AMD is having to make a case for 16GB; it has no other choice if going down the doppelganger MI50 route.

The sum of the changes results in, AMD says, a 29 per cent performance uptick over Radeon RX Vega 64. On an unrelated note, Radeon VII supports multi-GPU in DX12 and Vulkan through specific game implementation by the developer, meaning traditional CrossFire is no longer supported.

Radeon VII vs. GeForce RTX/GTX

RX Vega 64 isn't the benchmark that we need concern ourselves with. Priced at $699, the real competition comes from a pair of suitably high-end GeForce cards: the last-gen GTX 1080 Ti and present-gen RTX 2080, which starts at the same price. It is always difficult to cross-compare different architectures, of course, and though Radeon VII thumps either Nvidia GPU in the spec shootout - it has way more throughput and bandwidth than either - history suggests the fight will be much closer. We know this because RTX 2080, though ostensibly packing the same kind of potential as RX Vega 64, is in a different performance league.

It's clear that AMD has built the Radeon VII to take it to Nvidia's second-rung RTX GPU in today's games. Let's examine the card and then roll the benchmark dice.