Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
AMD announces more details about 7 nm Vega
There are a variety of sources on this. For example:
https://techreport.com/news/34243/amd-radeon-instinct-mi50-and-mi60-bring-7-nm-gpus-to-the-data-center
AMD says that the GPU will be rated at 14.7 TFLOPS. For comparison, a Radeon RX Vega 64 is 12.5 TFLOPS. That's faster, but not hugely so. AMD also says that it will have 4 stacks of HBM2 rather than 2, and clocked higher, which will more than double the memory bandwidth.
Today's presentation was about data centers, not gaming, so AMD didn't say whether the GPU will come to Radeon cards. While I don't know, I'd lean toward an answer of "no", as while it could be a new top end card for AMD, it would probably still be slower than a GeForce RTX 2080 (non-Ti), so what's the point? If they do make Radeon cards out of it, or perhaps out of the salvage parts, it still won't be competitive with Nvidia's top end, which limits how AMD can price it.
But that's not to say that all that AMD got out of the die shrink was another 20% performance. It's a compute card, and they put a lot of compute stuff in that Vega 10 doesn't have. For starters, it has half speed double precision compute, as compared to 1/16 in a Vega 64. It also has full ECC memory. Four stacks of HBM2 probably means 32 GB of memory, which is a lot more than you need in a consumer card. PCI Express 4.0 means double the bandwidth to get data to and from the GPU. Infinity Fabric allows a GPU to GPU connection akin to NVlink. They've added 8-bit and 4-bit packed integer instructions like what Nvidia added to Turing, though without doing a full matrix multiply-add with them. On paper, it looks competitive with a Tesla V100, and in about 40% of the die space. That's what AMD got out of the die shrink.
But all that compute stuff costs money, and if you're going to disable it in consumer GPUs, you can't charge extra for it there. It's highly probable that using the same die size and power for a Navi card on the same architecture would offer a far superior gaming card, so those wanting a new, high end gaming GPU from AMD will probably have to wait for Navi.
https://techreport.com/news/34243/amd-radeon-instinct-mi50-and-mi60-bring-7-nm-gpus-to-the-data-center
AMD says that the GPU will be rated at 14.7 TFLOPS. For comparison, a Radeon RX Vega 64 is 12.5 TFLOPS. That's faster, but not hugely so. AMD also says that it will have 4 stacks of HBM2 rather than 2, and clocked higher, which will more than double the memory bandwidth.
Today's presentation was about data centers, not gaming, so AMD didn't say whether the GPU will come to Radeon cards. While I don't know, I'd lean toward an answer of "no", as while it could be a new top end card for AMD, it would probably still be slower than a GeForce RTX 2080 (non-Ti), so what's the point? If they do make Radeon cards out of it, or perhaps out of the salvage parts, it still won't be competitive with Nvidia's top end, which limits how AMD can price it.
But that's not to say that all that AMD got out of the die shrink was another 20% performance. It's a compute card, and they put a lot of compute stuff in that Vega 10 doesn't have. For starters, it has half speed double precision compute, as compared to 1/16 in a Vega 64. It also has full ECC memory. Four stacks of HBM2 probably means 32 GB of memory, which is a lot more than you need in a consumer card. PCI Express 4.0 means double the bandwidth to get data to and from the GPU. Infinity Fabric allows a GPU to GPU connection akin to NVlink. They've added 8-bit and 4-bit packed integer instructions like what Nvidia added to Turing, though without doing a full matrix multiply-add with them. On paper, it looks competitive with a Tesla V100, and in about 40% of the die space. That's what AMD got out of the die shrink.
But all that compute stuff costs money, and if you're going to disable it in consumer GPUs, you can't charge extra for it there. It's highly probable that using the same die size and power for a Navi card on the same architecture would offer a far superior gaming card, so those wanting a new, high end gaming GPU from AMD will probably have to wait for Navi.
Comments
Guess we should all get used to $600 "midrange" cards from now on.
"The surest way to corrupt a youth is to instruct him to hold in higher esteem those who think alike than those who think differently."
- Friedrich Nietzsche
A good analysis of why and how imo.
Besides it’s pretty accurate IMO and there is no need to deflect its truthfulness
Well, if history is any indicator, the consumer high end parts based on the same general architecture as the professional/commercial stuff usually have very similar TFLOPS. A node shrink should produce much more than a ~18% performance improvement. Now, yes im aware that tflops isn't the whole picture. However it's generally a good run of the mill comparison. Yes, new faster versions of HBM could definitely help out, but realistically we might, maybe see a 25% gain on Vega 64.
That barely puts it in 1080ti territory, much less 2080 ti.
Now let's fast forward to the 6-9 (or more) months from now when we *might* see consumer variants of this card. By that point, Nvidia will be very close to, probably only a few months out, to releasing their own 7 or 8nm parts. Even if we assume the same 18-25% improvement purely from the process node, that still puts nvidia way WAY above AMD's part.
So again, nvidia would, as they are now, still be in a perfectly "safe" position to price the mid-high/high end however they want. Unless AMD can someone produce this new card at a 300-400 USD price range, the video card market is basically screwed from a consumer stand point.
I personally believe Nvidia 100% acheived their goal, which in my mind (again speculation as i can't prove it) was to get rid of old stock of pascal (they've done that spectacularly), and introduce price stickiness into the equation. By pricing the new cards so astronomically higher, consumers who unfortunately tend to have very short memories, will think that over the next few months when nvidia brings pricing down 10-20%, that they are going to be getting a "deal", all the while forgetting that they are basically going to be paying for the same price/perf as if they had bought a 1080ti (at this point in the future) 2 years prior.
"The surest way to corrupt a youth is to instruct him to hold in higher esteem those who think alike than those who think differently."
- Friedrich Nietzsche
At a 1.25 increase in tflop count only, this card will be slower than a GTX 1080 ti in gaming considering the bottlenecks of the Vega architecture at higher frequencies and would thus lose to a RTX 2080 which got slammed for it's performance not that long ago.
With only this much tflops, it makes absolute zero sense for AMD to release this as a consumer gaming card which it seems people want. You would be looking at speeds at best 20% faster than Vega 64 which put it slight faster than an RTX 2070 but much more difficulty being able price near that region because of the 7nm die and the 4 stacks of HBM2. If there ever will be a consumer variant of Vega at 7nm I wouldn't be buying it either.
Volta's tensor cores might be important for machine learning. They're certainly intended for machine learning, but I haven't studied it enough to know how useful they are there. For pretty much everything else, they're absolutely useless. I don't mean to dismiss them entirely, but what hardware is the best at what you're doing depends tremendously on what you're doing.
It's extremely unlikely that this will be AMD's highest performance GPU at graphics on 7 nm, or even their largest die. If you want to compare it to 14 nm, then comparing it to AMD's first GPU there (Polaris 10) shows a much larger than 18% improvement. Even if you want to compare it to Vega 10, it has a much smaller die size and probably much lower power consumption.
A comparison to a GTX 1080 Ti really isn't that helpful unless AMD decides to offer a Radeon card based on it. The GTX 1080 Ti is focused on graphics, while the new Vega is a compute card that isn't guaranteed to even be able to graphics at all. It probably can, but we don't know that for certain. It will almost certainly blow the GTX 1080 Ti out of the water on compute. Its real competition is the Tesla V100.
And while you focus purely on FLOPS, the new Vega has more than double the memory bandwidth of a Radeon RX Vega 64. It's also possible that they seriously beefed up the L2 cache, which is an area where GCN/Polaris lagged behind Maxwell/Pascal; I'm not sure if Vega 10 is more competitive there, but my guess is "no".
You have 13b transitions consuming 300W on 7nm, vs 21b transistors doing the same thing on 12nm (basically 16nm)
Improved cache latency
IF links have latency of 60-70ns! | Nvlink is 10us....Or... 10.000ns
As for infinity fabric, it sounds like it's the same thing that they use on their CPUs, with about the same latency. Not that the latency matters on a GPU. For that matter, not that infinity fabric or NVlink will matter beyond a handful of oddball cases, as transferring data directly from one GPU to another is a peculiar thing to do and hardly ever useful. Getting data from the CPU to the GPU and back is the big thing, and PCI Express 4.0 will help a lot there.
Or, they could have gotten the same performance at lower power by having more compute units clocked lower, at the expense of a larger die size. Which is basically what a Tesla V100 did.
Still, trying to infer something about energy efficiency of a process node by comparing a compute card on one node to a graphics card on another isn't a clean comparison. It would be like saying that AMD's Zen cores are really energy inefficient because a 8-core EPYC 7261 can burn as much as 170 W with a max turbo of only 2.9 GHz. It's not the CPU cores burning all that power; it's a bunch other stuff moving data around, whether PCI Express, memory controllers, infinity fabric, or whatever.
And it still tells us exactly nothing about Navi, which is what will matter to gamers.
The last time AMD had the crown (and it wasn’t that long ago) - it wasn’t like AMD marketshare all of a sudden spiked and went through the roof.
So AMD not having a card that is direct competition to a 1080ti or 2080 I don’t think really affects that much.... so they don’t have a $600+ GPU. That doesn’t make their sub-$300 market any less significant.
What I think has hurt their marketshare and perception more than anything is the mining craze. Gamers haven’t been able to get their hands on GPUs at what should be competitive prices. You may think it’s all roses and rainbow for AMD because they got the card sales regardless, but now that mining is tapering off, the loss of gaming marketshare, and AMDs perception in the gaming community, really hurts them more than not having a Halo card.
I would expect a few things to occur:
AMD would try to price it “to compete”, and since the competition is at $1,000+, I would expect AMD to be as well.
Folks would still say AMD drivers suck.
nVidia would react - we would see a either a Technical response to reclaim the crown, a price war of sorts, or both.
People will claim that CUDA/hardware PhysX/RTX/Gsync actually matter over performance.
AMD marketshare would not budge significantly
Any proprietary stuff from nVidia doesn't matter. Most of that shit gets dropped after a couple years and only a handful of developers ever support it. Only thing that has stuck is CUDA since nVidia dominates in professional cards.
That doesn't mean that those GPUs will cost $1200, though. Die shrinks allow a given level of performance for much less money. The reason that the GeForce RTX 2080 Ti costs a fortune to buy is that it costs a fortune to build. I've said that so many times that people are probably sick of hearing it from me, but it's true. If a die shrink of Turing on 7 nm gives a 400 mm^2 die in an RTX 3080 that is a little faster than an RTX 2080 Ti, even if AMD still isn't competitive at the high end, Nvidia would probably charge something like $700 for it.
"The surest way to corrupt a youth is to instruct him to hold in higher esteem those who think alike than those who think differently."
- Friedrich Nietzsche