With the new disclosures we’re hearing around the PlayStation 5, it’s a given that we’d also be hearing tips about the GPU that powers it. There have been a few broad categories of leaks around the chip.
Last year, AdoredTV suggested that AMD would aggressively bring a set of GPUs to market to undercut Nvidia’s Turing. Those leaks collectively looked like this:
Now, someone posting at 4channel (the SFW variant of 4chan) is claiming to work at AMD, with additional information on the “Radeon RX 3080.” (The thinking is, AMD will combine its Radeon and Ryzen branding into similar numbers for each family).
According to this leak, Navi has an additional 1MB of L2 cache compared with Polaris (3MB total). The L1 is now 32KB, up from 16KB. The top-end GPU offers 410GB/s of memory bandwidth on a 256-bit bus, compared with 256GB/s of bandwidth on a current RX 590.
None of this is in conflict with the earlier information leak, but the next bit of information suggests that Navi will target Vega 56 / GTX 1080 performance as opposed to the GTX 1080 / RTX 2070 — and that is new. It’s also not great news if you’re hoping AMD will deliver a devastating blow to Nvidia’s RTX family.
Here’s the problem: First, the gap between the Vega 56 and the GTX 1080 is between 10-20 percent, depending on the games you benchmark. They aren’t equivalently positioned; the Vega 64 was the GTX 1080’s closest point of comparison. The RTX 2070, meanwhile, is about 8 percent faster than the GTX 1080. Matching the Vega 56 with Navi is considerably slower, therefore, than matching the RTX 2070.
While a Navi that matches Vega 56 would be underwhelming compared with expectations, matching Vega 56 performance while slashing power consumption and price would be a larger collective set of improvements than the Radeon VII delivered over Vega 64 this year.
The Radeon VII improved performance over Vega 64 significantly with minor changes to power consumption and noise. This hypothetical Navi would be cutting price, boosting performance, and dramatically slashing power consumption (and therefore noise). It would unquestionably deliver more benefits — just not necessarily in the places people want them the most.
If the earlier leak is true, of course, Navi would be truly gunning for the stars. A $ 250 GPU competing against a $ 500 GPU would punch gaping holes in Nvidia’s price brackets and product banding.
The biggest reason to think AMD would take a step like this is the company’s own weak position in the GPU market. AMD may own most of the console space, but it’s been all-but driven out of the high-end desktop and laptop markets. Coming out of the gate swinging could help it win back mind share.
The biggest reason to think they won’t? Let’s be honest. In the nearly six years since Hawaii launched, AMD’s competitive track record in graphics hasn’t been all that great. Hawaii was a great competitor but had volume issues that weren’t resolved until third-party cooler designs were available. Fury X had an issue with glue in its water cooler and couldn’t cleanly beat the GTX 980 Ti.
AMD’s Radeon VII
When the Polaris family dropped, we saw issues and concerns surrounding inappropriate power consumption on the PCIe slot as a result of improper GPU 12V distribution. Vega 56 and Vega 64 were only available in limited quantities, at sky-high pricing (they had the misfortunate to launch as the GPU price boom of 2017 was really getting started), and in two different SKUs with two different cooler z-heights. This might seem like a weird, inside-baseball issue, but I promise you that the last thing AMD wanted to have to do for its niche GPU was design two niche cooling solutions, only one of which could be used depending on which part you had. Radeon VII’s laudable performance improvements were overshadowed by its volume and general lack of new features compared with other cards on the market.
AMD has had a few bright spots since then. The Radeon Nano was a well-received niche product, and the HD 7790 boosted AMD’s low-end performance over the HD 7770 in a cost-effective manner. But AMD needs more than a niche product or a well-received single budget card. It needs a new architecture that demonstrates it can compete with Nvidia from top to bottom. Even if it doesn’t launch new cards into every segment, it needs to demonstrate an architecture that’s capable of scaling. But any honest examination of the past six years reveals that AMD has had trouble doing exactly that. It has earned hundreds of millions of dollars in the console market and it still has a very strong lock on the lower end where the RX 570 plays, but it’s been a long time since AMD delivered a GPU introduction firing on all thrusters with no problem or issue in sight.
This is not to say that I’m preemptively judging Navi to be a failure. A $ 250 Vega 56 or GTX 1080-equivalent would be a huge performance jump over AMD’s current silicon in a fraction of the power. We don’t know what the ray tracing situation is yet, because we don’t know if the Navi silicon Sony is using in 2020 is identical to the Navi silicon AMD is bringing to market in 2019. AMD has, thus far, been very quiet about ray tracing except to say that they didn’t expect to introduce the feature until they could do it, top-to-bottom. This could imply that they are holding back the capability for a future GPU launch, possibly aligned to the PS5’s formal debut. There are also rumors of different flavors of Navi and different performance targets debuting up and down the stack, with some arriving in 2020. Alternately, these could belong to Arcturus, Navi’s supposed successor.
Gamers who have been calling for AMD to deliver a “Ryzen” level leap in gaming performance aren’t wrong; the company needs one. While the company still maintains performance against Nvidia in some price bands, it now draws nearly 2x the power to do it. That’s an ugly, unsustainable position to be in. Will Navi be the architecture to turn it all around? Hopefully, yes — but we’ve been waiting for a proper Hawaii follow-up for six years now. 4chan isn’t exactly a reliable data source. Then again, AMD’s past six years of GPU deliveries haven’t been all that strong, either.