AMD has quietly launched a new version of the RX 580 with fewer stream processors and slightly lower boost clocks. The new Radeon RX 580 2048SP has, as the name implies, just 2,048 stream processors down from the 2,304 in the full version of the card. It’s not unusual for GPU manufacturers to fill out their product lines with a handful of country-specific or OEM parts created for specific markets, but AMD is splitting a rather thin hair here.
The RX 580 2048SP isn’t really an RX 580 at all. It’s memory clock, GPU clock, and the number of stream processors all mark this as an RX 570, with only the tiny core clock adjustment to bring the chip’s performance up a whisker. A 3 percent clock boost isn’t going to do much for overall performance, particularly considering the reduced number of SPs. As far as we know, this chip is headed for the Chinese markets, not the West, though that could theoretically change.
It’s not unheard of for manufacturers to make subdivisions like this, as mentioned above. Back when Nvidia launched its Tesla line of GPUs (GT200, not its line of HPC accelerators and add-in boards), it quickly realized it had miscalculated. The GTX 280 and GTX 260 both got substantial price cuts just weeks after launch once AMD’s HD 4870 and HD 4850 proved to be faster and far better priced. In addition, Nvidia introduced a “GeForce 260 Core 216” edition of the GeForce GTX 260 (the original flavor had just 192 cores).
And AMD has been down this road before with Polaris. Last year, it permitted OEMs to sell RX 560 cards with just 892 cores, as opposed to the 1,024 standard. AMD’s justification for this has been that offering more flexible solutions helps OEMs sell more graphics cards to their customers at specific price points. Sure. That’s the reason companies subdivide markets in the first place. If the only GPUs you sold were priced at $100 and $1,000, you’d wind up missing an awful lot of buyers.
What people tend to take issue with is the branding, which AMD is presumably trying to head off with its 2048SP. In this case, however, it feels a bit disingenuous. Back when Nvidia built the Core 216, it was upgrading the GPU it had previously offered and introducing price cuts at the same time. The RX 580 2048SP is an RX 570 for all intents and purposes. A 3 percent boost clock improvement isn’t enough to justify a new brand or model number, but if AMD did want to introduce one, a part number like RX 575 would distinguish this card from the RX 580. Best of all, from an accuracy perspective, would be to call this an RX 570+ or something equivalent.
Switching topics slightly, we’ve seen a lot of action in the GPU market of late, with Nvidia’s Turing refresh, but we’ve seen relatively little from Team Red. There are rumors of a Polaris refresh cycle still to come this year (we haven’t heard more on that topic yet) and AMD is on track to launch its 7nm Vega as far as we know. I’ve seen some readers asking if I thought AMD had an opportunity to bring 7nm Vega to the consumer market to compete with Pascal or Turing. Looking at the performance of the RTX 2070, I think the signs still point more towards a “no,” than a yes. While Vega 64 is reasonably good competition for the GTX 1080, a 7nm Vega would need to deliver a 20-25 percent performance improvement to really carve a niche for itself. That kind of jump would put it between the RTX 2070 and RTX 2080 (or, if you prefer, between the GTX 1080 and GTX 1080 Ti). But pushing Vega performance up 20-25 percent would likely take a clock jump of equivalent size, putting the card’s clock speed between 1850 – 1930MHz.
There have been rumors that most of AMD’s engineering staff were pulled off Vega to work on Ryzen and that the design suffered for it, but AMD would have had to truly screw the pooch to leave a problem that big sitting in their GPU design. And such a problem would need to exist for Vega’s clock speed to suddenly skyrocket in such fashion, with a corresponding drop in power consumption at every clock speed. We’re not going to say it’s impossible for AMD to whip a vastly improved Vega out of its back pocket and wow everyone, but the company has given no sign of such a coup. We would expect a 7nm Vega to outperform its 14nm counterpart and to draw less power doing it, but something would have to have gone catastrophically wrong to give AMD a clock jump that huge.
To put this in perspective, when AMD moved from 28nm Fury to 14nm Vega, it was able to increase clocks by 1.18x. Boosting them by 1.2 – 1.25x more would mean AMD found a way to squeeze even more performance out of its GCN design than it got when it moved from 28nm planar to 14nm FinFET transistors. The jump from 14nm to 7nm doesn’t provide as much performance as the 28nm – 14nm shift, making this unlikely.