That would still be a >1000mm2 die, so it's not gonna happen Zeppelin-like. MCM, of course, is a possibility, let's see if they're gonna try it before Navi exciting times indeed!
Sent from my VTR-L09 using Tapatalk
That would still be a >1000mm2 die, so it's not gonna happen Zeppelin-like. MCM, of course, is a possibility, let's see if they're gonna try it before Navi exciting times indeed!I think a very real possibility is that dual Vega is 2 complexes connected together similar to Zepplin. We know Vega has Infinity Fabric incorporated and so it makes sense that they are meant to connect to each other.
I understand dual GPU support is falling, but dx12 is still only a handful of games and I'm going on assumption that Vega does not beat 1080 ti.If you mean Dual GPU in the traditional sense, then sounds like a terrible Idea.
The talking point of "we have the fastest GPU" isn't worth much if the rest of that sentence is "but it is 2x the die size of the competition, guzzles power, and in ~50% of titles it only performs half as well as it theoretically can." Which is why a dual GPU card in the traditional sense is a horrible idea for a flagship in 2017. We've seen for years and years how multi-GPU support is hit-or-miss at best, while single GPUs have become powerful enough to deliver all the performance we want (and more). Remember back in the day, when pretty much nobody could play a game at 1024x768 with settings maxed out, or more recently, at 1080p 60fps maxed out? These days, even with far more demanding current titles, a $200 GPU gives you that. 1080p Ultra @60Hz has become the new baseline for low-to-midrange cards (RX 480, 580, 1060). Which is flippin' crazy, but we're there. Heck, most reviewers these days refuse to lower graphics settings, as if that's some sort of sacrilege. Even 4-5 years ago, playing any game maxed out at any respectable resolution was a privilege only a few could afford. As such, we don't need single-card multi-GPU solutions any more. Even $70 motherboards support CF, and multiple cards have the benefits of cooling far better, giving you an upgrade path, and just generally being far more practical. Oh, and not sounding like a jet engine.I understand dual GPU support is falling, but dx12 is still only a handful of games and I'm going on assumption that Vega does not beat 1080 ti.
AMD needs a halo part, and this would give them the fastest card on games that support CF and the talking point "We have the fastest GPU".
The real money for them is going to come from the $500 part, and the halo part will drive sales of that. It would obviously be better if AMD is launching a Volta competitor, but I don't think they have that.
I'm taking a " show me" stance on infinity fabric at this time.
Hmm, JFK's new frontiers... Does anyone have a clue what exactly they meant by this?
The talking point of "we have the fastest GPU" isn't worth much if the rest of that sentence is "but it is 2x the die size of the competition, guzzles power, and in ~50% of titles it only performs half as well as it theoretically can." Which is why a dual GPU card in the traditional sense is a horrible idea for a flagship in 2017. We've seen for years and years how multi-GPU support is hit-or-miss at best, while single GPUs have become powerful enough to deliver all the performance we want (and more). Remember back in the day, when pretty much nobody could play a game at 1024x768 with settings maxed out, or more recently, at 1080p 60fps maxed out? These days, even with far more demanding current titles, a $200 GPU gives you that. 1080p Ultra @60Hz has become the new baseline for low-to-midrange cards (RX 480, 580, 1060). Which is flippin' crazy, but we're there. Heck, most reviewers these days refuse to lower graphics settings, as if that's some sort of sacrilege. Even 4-5 years ago, playing any game maxed out at any respectable resolution was a privilege only a few could afford. As such, we don't need single-card multi-GPU solutions any more. Even $70 motherboards support CF, and multiple cards have the benefits of cooling far better, giving you an upgrade path, and just generally being far more practical. Oh, and not sounding like a jet engine.
1080Ti has 11.3 TFLOPS at 1582Mhz. One can easily get 1950Mhz from 1080Ti custom cards, which translates to around 14TFLOPS, that's higher than stock Vega, and based on past polaris results, theres not much overclocking for Vega probably.If we look at these numbers and the supposed 12.5 TFLOPS; one could expect Vega to land close to the Titan xp. But there's still the chance that even if, given optimal scenario, vega could outperform 1080Ti -we might not see it in practice that often. Perhaps either due to limitations in DX11(majority of game titles) or games being optimized for Nvidia in specific.
For instance lets say the consumer Vega can vastly outperform Pascal in FP16. All Nvidia has to do is to make sure upcoming games don't heavily rely on FP16 and Vega runs underutilized. In fact they don't even have to explicitly take any action since anyone who wants their games to run well on Pascal will have to avoid FP16 anyway. Hopefully Volta will be an improvement in that regard.
All that aside I for one am willing to assume that Vega can theoretically outperform 1080Ti, but at the same time I wouldn't be surprised if what we see in practice is performance that is "only" around 1080 level more often than not. Personally I would be fine with that(if the price is right), but it would also mean they aren't ready to take on Volta and they would end up making fools of themselves with the whole "poor volta" thing..
1080Ti has 11.3 TFLOPS at 1582Mhz. One can easily get 1950Mhz from 1080Ti custom cards, which translates to around 14TFLOPS, that's higher than stock Vega, and based on past polaris results, theres not much overclocking for Vega probably.
While the Steam Hardware Survey is hardly representative of forum users here, the vast majority even on this forum (although probably not the most active users) are using 1080p or 1440p monitors at 60Hz. 4K60 is gaining popularity, albeit slowly (the monitor upgrade cycle is 5-6 years if not more for most users), while high refresh rates at lower resolutions are only slightly ahead of this. And I'd say video cards that cost $400 and up are a niche market, not $800 and up, even if Nvidia probably sells 10x as many 1070s as 1080s (not to mention 1080Tis). And who here is talking about AMD having a "halo product" as fast as the 1080? I certainly don't expect that. Besides, CF and SLI scaling is bad enough that for it to make sense even in games that make use of it, your GPU budget needs to be significantly above the price of a 1080Ti - otherwise, a single card will perform better in the vast majority of cases. This wasn't the case just a few years ago.I wasn't aware we had GPUs that were putting out 100fps at 4K Ultra settings, must have missed that.
In my opinion, 1080p at 60Hz doesn't even bear mentioning on an enthusiast graphics forum, how many people posting here even care about that? And if they do, what's the point of discussing? Buy any $200 card and you're there.
People with 4K , Eyefinity, and 3440 X 1440 monitors with high refresh rates still find use for multi GPU. Video cards that cost $800 and up are a niche market for people with niche market monitors.
While I agree on the "half the games" assessment, what would be better for AMD?
To have a card that is faster than NVIDIA's best card in half the games and costs significantly less?
Or to have a card that is slower than the 1080Ti, the first Titan Xp, and the second Titan Xp?
Having a "halo product" that is as fast as your competitor's third best part sort of re-defines "halo part" doesn't it?
We know from the Vega preview slides that each CU has the same amount of SP's, so we can rule this out.If...
All we know from the drivers is the number of compute cluster or what they are called at 64. And since in previous AMD design this always had 64 SPs we land at 4096. But it's a new uArch. And number of SP per cluster is one thing that changed from Kepler to maxwell for example. We also know that each core/SP was optimized (eg. the image with the large truck and smaller truck). Anyway each SP will be IMHO significantly smaller but also a bit slower than old SPs. Important part being more efficient (power and die size). So each cluster will have more slower SPs. I would guess 96 per cluster at 64 clusters that would be 6144 SP. That would also explain why 2xtimes polaris 10 doesn't add up with die size and transistor count. It then should not be 500mm2+. But if it actually has more SPs but smaller ones, that then totally makes sense.
Really? Can you pls share a slide?We know from the Vega preview slides that each CU has the same amount of SP's, so we can rule this out.
Really? Can you pls share a slide?
Well, but can't current GCN CUs do "only" 64 single precision MADD per clock?
4x 16-lane wide SIMDs per CU - each line can do one 32-bit operation per clock
GCN Whitepaper:
IEEE 754-2008 FLOATING-POINT ARITHMETIC IN AMD'S GRAPHICS CORE NEXT ARCHITECTURE:
So, that means that FP32 throughput is doubled in NCU, correct?
Hmm, JFK's new frontiers... Does anyone have a clue what exactly they meant by this?
TX for reply. Can you please point me to some materials for study?GCN (and VLIW4 for that matter) do FMA, which is counted as two operations in a single cycle.
Assuming linear scaling perf->core clock, dont you need memory to scale accordingly to make that call? Also we know jack right now, speculating in an overclocked ti vs a stock vega that cant overclock cause reasons is... very very crystal ballish
Damn... SK Hynix seems to be the global failuries of the memory industries with countless over the top promises and yet certain let downs. Why can't AMD just steer away of those useless "partners" that make proper competition to be impossible.
Damn... SK Hynix seems to be the global failuries of the memory industries with countless over the top promises and yet certain let downs. Why can't AMD just steer away of those useless "partners" that make proper competition to be impossible.
TX for reply. Can you please point me to some materials for study?
Just a question to CatMerc and antihelten:
In that case, what they meant by this slide? Are they referring to SP ops/sec throughput with or without FMA?
End notes:
Isn't that what I just described?
What's the problem? A Memory catalog means absolutely nothing, especially for customers like AMD and nvidia.Damn... SK Hynix seems to be the global failuries of the memory industries with countless over the top promises and yet certain let downs. Why can't AMD just steer away of those useless "partners" that make proper competition to be impossible.