AMD has been quite tight-lipped about Vega, its upcoming consumer GPU meant to compete at the high end of the market, as opposed to the midrange RX 400/500 families. Thanks to some leaks, however, we now know a bit more than we did before about what Vega can do and how it does it.
Longtime AMD fans may remember the company’s use of various X-derived GPU names, including XL, XT, and XTX. AMD is reviving this branding style for Vega, according to Tom’s Hardware, with plans to offer four different GPU models, as shown below:
We know that the current Vega Frontier Edition has a rated 300W TDP and most of these GPUs end up in similar territory. We don’t know the clock on the consumer memory (Vega Frontier Edition uses 1890MHz), but there are two significant differences between Vega FE and the consumer RX Vega.
First, if THG is right, all of the Vega consumer cards are limited to 8GB of RAM. No exceptions. This implies a few things about AMD’s current positioning and the state of HBM2. While the Vega Frontier Edition has 16GB of VRAM in two stacks of 8GB each, consumer Vegas would be returning to two stacks of 4GB. That’s still 4x the memory density per stack than what the Fury X fielded, but it’s not entirely what people expected. Then again, there could be a small benefit here — less HBM2 means lower power consumption. AMD has talked about Vega’s significantly improved memory efficiency compared with Fury X, which means the company may feel it can get away with using less VRAM, given that it has improved both its color compression (saving bandwidth) and its overall memory utilization.
Second, dropping from 16GB to 8GB should give AMD some breathing room on cost. We’ve increasingly believed that HBM2 is having some serious growing pains, thanks to AMD’s delayed Vega launch, certain high-end product SKUs vanishing off part availability lists from Hynix, and comments that AMD executives like Raja Koduri have made. This could mean AMD has more room to price the RX Vega against whatever GPUs it winds up competing with (we have to admit early figures on the 1080 Ti don’t look great, though it matches fairly well against the GTX 1080).
At the same time, however, we can’t put too much importance on hypothetical higher clock rates. While Raja Koduri has said the consumer variants will be clocked higher than the Vega FE, AMD hasn’t historically gotten a great return here. For example: The RX 580 has a base clock 1.12x higher than the RX 480 and a boost clock 1.06x higher. The bad news is, it consumes 1.29x more power to deliver those benefits. (To be fair, the RX 580 does hold its boost clock better and more consistently than the RX 480.) But that’s not a pretty ratio to be starting from.
So what does that tell us about Vega? It depends on what assumptions you make about the card. If, for example, we assume that Vega’s clock speed and TDP scale the same way Polaris does, then my assumption about a 1.7GHz clock is probably pretty good. If RX Vega scales better than Polaris, a 1.75GHz or even 1.8GHz clock doesn’t seem completely impossible — and an 1800MHz clock would give RX Vega a 12.5 percent higher boost clock than Vega FE. That’s not going to be enough to match the GTX 1080 Ti, but it would give the RX Vega some solid wins over the GTX 1080 and turn a few losses into narrow victories or practical ties.
One other observation. Since GCN debuted, one of the hallmarks of the design was it has almost always scaled better when moving from 1080p to 4K. This has vanished according to tests by PC Perspective (grain of salt, early drivers, not a gaming card, etc, etc). Nvidia cards are now showing superior scaling characteristics from 1080p to 4K than AMD is.
There are a few ways this could play out, based on what we know so far. The 56-CU Vega XL could be positioned between the GTX 1060 and the 1070 with the 64-CU Vega XT (air-cooled) landing between the GTX 1070 and GTX 1080. The air-cooled and water-cooled Vega XTX (OC) would then try to close the gap with the 1080 and approach the 1080 Ti. This makes sense if AMD wants to match the GTX 1080 with its air-cooled XTX, then surpass it by a solid margin with its water-cooled variant.
Alternatively — and this assumes game benchmark scores are low due to driver issues or possibly a need for that water cooler — the Vega XL actually drops in between the 1070 and 1080; the air-cooled Vega XT has a win-some-lose-some battle with the GTX 1080; the air-cooled RX Vega XTX (OC) beats or at least matches the GTX 1080 in most games; and the water-cooled RX Vega XTX (OC) performs well enough to be a slightly slower (but also slightly less expensive) competitor to the GTX 1080 Ti. But based on what we’ve seen to-date from the Vega Frontier Edition, AMD needs to pull two rabbits out of its hat to make that happen. First, it needs to pick up at least 10 percent in average game performance thanks to better drivers. Second, it needs to pick up 8-12 percent of clock speed over the Vega Frontier Edition’s 1.6GHz.
That’s a pretty tall order. With RX Vega launching later this month, AMD doesn’t have much time left. This is going to be an extremely interesting launch.