Well I was definitely wrong on my assumptions. I'm baffled. AMD outright said they kept Vega 64's rendering stuff the same with Vega 20, so where they double ROPs comes from I don't know. But I saw one of the Anandtech guys saying that AMD had told him they double ROPs, but that also makes little sense to me since Vega 20 was supposedly enterprise focused. But then they're apparently using these in some of the game streaming enterprise solutions, so I guess that makes sense (but doesn't make sense that was basically the only way they addressed that). For me it would make more sense to buy up Vega 64 cards that were made for mining.
Granted a lot of my takes (Navi being ready now) were based on rumor (and I tried to make that clear), so I'm not really disappointed that didn't pan out like I thought. But this card seems to deliberately refute things AMD has said, so now even what they say has to be taken with a grain of salt.
The silliest part is that, its not like this card is really upsetting Nvidia's market strategy, so there really was little to no reason to keep it under wraps and say Vega 20 would not be for gamers (which they didn't make this a Frontier Edition card, so either they're taking the total opposite track of Nvidia and claiming that above this is prosumer level, or this is a gaming card). I don't see that they likely would've lost Vega 64 sales (if anything, they should've announced this thing ASAP to steal some thunder from Nvidia's RTX), but then I wonder if the only reason we're seeing this is because the RTX stuff didn't go over as well - largely because of the pricing - which made AMD feel like they could offer this as a competitive product. Well that and either the yields on 7nm are already that good or its flopping hard in enterprise (which I'd think they'd just drop the price, or shift it to prosumer where they could probably sell it for double or at least $1000; maybe offer special pricing packages for Threadripper workstations that use it).
I really hope this isn't an indication that Navi is that far off.
Did other performance figures get leaked? (Just saw that there are some being discussed.) The only thing I've seen touted its performance via AMD compared to the 2080 (which is likely best case scenario), and while its definitely a step up over Vege 64, its honestly not that impressive, and considering the price, and I have serious questions if AMD is making money selling the card at that price (since 7nm and more HBM channels and stacks), I just have to wonder why they're doing this. Its inline with what I was thinking and to me isn't that good. I'm pleasantly surprised that they apparently can get it down to that price, as I was thinking it'd cost them even more, but even for that price I'm not impressed.
Not only that, but I'm not at all impressed with the efficiency. That to me indicates the issues with Vega were not related to GF's process much (that Zen 2 is seeing such a big boost in efficiency tells me that 7nm seems to be living up to its potential quite a bit already). Same with the clock speeds. Vega seems like an inherently flawed architecture, one that seemingly prioritized compute capability. Which that would be fine, but its rendering/rasterizing capability suffered mightily for it, and I'm not sure that compute outstripped just doubling Polaris and adding support for further reduced precision math. We can be fairly certain that gaming performance didn't benefit versus that, and if they would've done 96 ROPs, it'd probably be more balanced. On top of that, just doubling the memory channels would've probably made it cheaper and it'd have had more bandwidth than HBM2 solution had.
Guess we'll see. People have speculated a lot that Vega was bottlenecked in ROPs and/or memory bandwidth. But if things were that severe, I'd expect a bigger improvement when doubling both. Seems like it will still be unbalanced, just in the opposite direction now or in some other manner (unless its still ROP and/or memory bandwidth bottlenecked in which case...WTH).
I really hope AMD was getting access to some of the data that Microsoft used to decide on the One X setup, as that seems like a well balanced and efficient design. Which that would seem to support that AMD's dGPUs tend to be bandwidth limited some, but I believe its actually the TMUs that was boosted relative to Polaris, while the ROPs were the same. And I believe AMD has said they don't believe ROPs is a limiting factor in the past (so I wouldn't be surprised if most of Vega 20's performance increase comes simply from the memory).
Its a shame that I think GPU IP has been pretty staunchly locked up, as I feel like this market could use a boost from companies trying some similar ideas but just trying different ratios or other things. I also feel like right now that AMD and Nvidia are both trying to figure out next steps, where they're trying to determine specialization versus integration. While they're also dealing with major engineering challenges, and trying to balance product portfolios, where they're trying to design GPUs for such a wide breadth of markets that I can't help but feel like we're getting a lot things that aren't well suited for their markets.
Facebook
Twitter