Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
AMD Navi and Rome will launch in Q3 of this year
https://www.anandtech.com/show/14286/amd-7nm-navi-gpu-and-rome-cpu-to-launch-in-q3
For the uninitiated, Navi is AMD's next generation GPU architecture, while Rome is the server version of their next generation CPU architecture. That news is directly from AMD, not just random Internet rumors. Furthermore, it's close enough to launch that the big orders for parts have surely been placed with the foundries so major launch delays at this point are unlikely.
Rumors say that the initial Navi parts aren't going to challenge Nvidia for dominance of the high end of the GPU market. They could, however, crater prices on the mid-range, as this will be the first real die shrink in about three years. If you're looking to spend $200-$500 on a GPU, then Navi might be an interesting product to wait for, depending on how long you're inclined to wait. With the move to 7 nm, it's also likely to be more energy efficient than anything on the market today.
Meanwhile, Rome is a server part, so it's not something for home users. However, AMD's previous CEO had previously said that third generation Ryzen would be their next major launch. It's not clear if that's also going to be in Q3 or coming sooner. The great unknown with Rome is what happens to memory latency, as it's going to take two hops to go from CPU cores to memory rather than one. If that doesn't cause problems, then it's likely that AMD take a large chunk of the server market, as Intel won't have any credible competitor to it until at least next year and possibly much later.
The CPU dies in Rome will be the same as in third generation Ryzen, so it will probably have similar memory latency characteristics. How high the consumer parts will clock is unknown, but it will probably put AMD in a stronger position in the consumer CPU market than they're in today. It's also probable that it will be much more energy efficient than anything you can buy today, at least outside of the low power, low performance parts that aren't particularly interesting for a gaming desktop. As with the server parts, taking two hops to get to memory instead of one as in current parts could cause memory latency problems, which could make third generation Ryzen not that interesting as a desktop part.
Launch day doesn't necessarily mean a hard launch, however. Furthermore, with AMD currently well behind the competition in both the GPU and server markets, AMD has plenty of incentives to rush their launches to get parts out there as soon as possible. Furthermore, the server market is so big and so lucrative that it's possible that it will consume most of the 7 nm fab capacity that AMD can get, forcing consumers to mostly wait weeks or months after a soft launch. The highest margin parts always get priority when there's limited fab space, and that means servers, not desktops.
That said, TSMC 7 nm isn't a new process node, but has been out for quite some time. AMD surely wants to claim massively more capacity on it than the did a few months ago, but it might be possible for them to get all that they need.
For the uninitiated, Navi is AMD's next generation GPU architecture, while Rome is the server version of their next generation CPU architecture. That news is directly from AMD, not just random Internet rumors. Furthermore, it's close enough to launch that the big orders for parts have surely been placed with the foundries so major launch delays at this point are unlikely.
Rumors say that the initial Navi parts aren't going to challenge Nvidia for dominance of the high end of the GPU market. They could, however, crater prices on the mid-range, as this will be the first real die shrink in about three years. If you're looking to spend $200-$500 on a GPU, then Navi might be an interesting product to wait for, depending on how long you're inclined to wait. With the move to 7 nm, it's also likely to be more energy efficient than anything on the market today.
Meanwhile, Rome is a server part, so it's not something for home users. However, AMD's previous CEO had previously said that third generation Ryzen would be their next major launch. It's not clear if that's also going to be in Q3 or coming sooner. The great unknown with Rome is what happens to memory latency, as it's going to take two hops to go from CPU cores to memory rather than one. If that doesn't cause problems, then it's likely that AMD take a large chunk of the server market, as Intel won't have any credible competitor to it until at least next year and possibly much later.
The CPU dies in Rome will be the same as in third generation Ryzen, so it will probably have similar memory latency characteristics. How high the consumer parts will clock is unknown, but it will probably put AMD in a stronger position in the consumer CPU market than they're in today. It's also probable that it will be much more energy efficient than anything you can buy today, at least outside of the low power, low performance parts that aren't particularly interesting for a gaming desktop. As with the server parts, taking two hops to get to memory instead of one as in current parts could cause memory latency problems, which could make third generation Ryzen not that interesting as a desktop part.
Launch day doesn't necessarily mean a hard launch, however. Furthermore, with AMD currently well behind the competition in both the GPU and server markets, AMD has plenty of incentives to rush their launches to get parts out there as soon as possible. Furthermore, the server market is so big and so lucrative that it's possible that it will consume most of the 7 nm fab capacity that AMD can get, forcing consumers to mostly wait weeks or months after a soft launch. The highest margin parts always get priority when there's limited fab space, and that means servers, not desktops.
That said, TSMC 7 nm isn't a new process node, but has been out for quite some time. AMD surely wants to claim massively more capacity on it than the did a few months ago, but it might be possible for them to get all that they need.
Comments
But all signs point that Navi should deliver more performance per dollar than NVidia's current offerings.
VII was on 7nm and didn't exactly break any energy efficiency records - nVidia is still more energy efficient on an older process node for similar performance. Now, VII was older GCN, and Navi is supposed to be a larger evolution of tech, so we should see something, but 7nm in and of itself doesn't automatically mean better energy efficiency. It will be difficult to tell how much any efficiency improvement is because of improved engineering, and how much is from the better production node.
It will be what it will be. I hope it's good - I've liked Polaris a lot and own a 470 and 580 in two different rigs, but I'm looking for something to upgrade my GTX980 to, and not exactly willing to spend four figures to do that. Navi has potential.
Also not to forget, next gen consoles are around the corner, they are all talking AMD custom APUs again, and that will take up a lot of capacity as well. You already have Apple eating a ton of 7nm process space, and the upcoming flagship Androids will probably all be sporting 7nm Snapdragons as well. Now, all of those are exactly identical processes, but it's still something to consider when looking at a fixed production capacity. So I think Quiz is correct that supply could be constrained on Navi (and possibly consumer Zen2)
1) Navi is yet another iteration of GCN, rather than finally making a substantially new architecture for the first time since 2012. Recall that Nvidia has released three major new architectures (Kepler, Maxwell, Volta/Turing) since then.
2) AMD decides to stock clock the GPUs very aggressively, so that they're not very efficient at stock speeds, but could be made a lot more efficient by throttling the clock speeds back. Right now, AMD is doing this with Polaris and Vega because they're behind and want to make it look like they're less far behind. Recall that Nvidia did the same with Fermi when the situation was reversed and for the same reasons.
3) Navi is simply a bad architecture.
Certainly, any of those could happen. I'd regard each of them as unlikely. The Radeon VII does have quite a case of (1) and (2). But with reduced clock speeds and voltages, it could likely be an efficient, ~150 W competitor to an RTX 2070, not a 295 W competitor to an RTX 2080. And that's in spite of being heavily derivative of a very old architecture and being an early part on a new process node, neither of which will be a problem for Navi.
https://www.anandtech.com/show/14352/amd-roadmap-navi-matisse-rome-q3
Matisse, or third generation Ryzen, is also coming in the third quarter of this year. There had been some speculation that it would launch during Computex, but that's apparently not going to happen.
AMD also says that Navi is an "all new GPU architecture". I'd interpret that as meaning that it isn't yet another GCN derivative.
https://wccftech.com/amd-zen-2-5ghz-12-core-4-3ghz-16-core-cpus-leaked-benchmarked/
That isn’t because I think these will be bad chips. I just think AMDs turbo scheme is that good - it will already have eeked out just about as much as your going to get from the silicon.
On Zen+, manual over locks often performed no better (or in cases even slightly worse) that just leaving aggressive turbo on. I expect Zen2 to be the same if not an improvement.
wccftech is one of the worst rumor site out there. They are clearly a “for clicks” model and accuracy be damned.
wtftech.com is often like a trash can of hardware news. Other news websites publish occasional rumors after trying to vet the credible ones, but wtftech.com publishes all of them as long as they make for good headlines.
Brenics ~ Just to point out I do believe Chris Roberts is going down as the man who cheated backers and took down crowdfunding for gaming.
The 16 core they claimed was stable at an OC of 4.3 across all cores. The 12 core would indeed boost to 5.0, but they were only able to get a stable 4.7 across all cores manually.
Brenics ~ Just to point out I do believe Chris Roberts is going down as the man who cheated backers and took down crowdfunding for gaming.
Brenics ~ Just to point out I do believe Chris Roberts is going down as the man who cheated backers and took down crowdfunding for gaming.
If my next CPU runs at 2.2Mhz but it faster/more efficient than what I have now, I’ll be perfectly happy with that.
Brenics ~ Just to point out I do believe Chris Roberts is going down as the man who cheated backers and took down crowdfunding for gaming.
Brenics ~ Just to point out I do believe Chris Roberts is going down as the man who cheated backers and took down crowdfunding for gaming.
Brenics ~ Just to point out I do believe Chris Roberts is going down as the man who cheated backers and took down crowdfunding for gaming.
Clock speed really isn’t relevant. A given architecture is going to consume a certain amount of energy to perform work. The faster the frequency - yes, the faster it will perform
work, but more so the more energy it will consume. The reason you can’t just keep cranking the clock up to keep going faster is that you hit an energy wall.
Turbo manages energy. It does so by manipulating voltage and frequency. What it aims to do is give you the most work in the most efficient manner.
if you had 2 cores at 5.0, that means there isn’t any work for the other cores to do and their speed is inconsequential. Ideally they will be in sleep waiting until there is work to do. The turbo algorithm is routing as much energy as it can to where it’s needed.
if you did have work for all 12 cores, turbo would crank them all up to maximize efficiency - I think the rumors I saw were around 4.6-4.7. Which is probably about the same as you’d get out of a manual overclock.
Except the manual overclock won’t get you that 5.0 when you only need a couple of cores...
Turbo (or whatever AMD is calling theirs) is basically a very good automatic and dynamic overclock algorithm, which is why in the last couple of generations it’s been able to meet, and in many cases exceed, manual overclock real world performance.
clock speed is just a marketing number. What matters is what the architecture does with that, and the proof of that is in the benchmarks.
Brenics ~ Just to point out I do believe Chris Roberts is going down as the man who cheated backers and took down crowdfunding for gaming.
It used to be that turbo wasn't that aggressive. A Core i7-920 had a nominal stock speed of 2.66 GHz, and a max single-core turbo of up to 2.93 MHz. People who turned off turbo could sometimes overclock it to 4 GHz. If that's your target hardware, then sure, overclocking makes sense.
Today, for the higher end parts, the single-core turbo is sometimes higher than you can get an all-core stable overclock to on the stock cooler. Sometimes it's higher than you can get an all-core stable overclock to even in a pretty good air cooler. You can still beat it with exotic cooling, of course.
But let's suppose that you have a choice of:
1) One core at 5.0 GHz and the rest clocked low or off
2) Two cores at 4.9 GHz and the rest clocked low or off
3) Four cores at 4.8 GHz and the rest clocked low or off
4) All cores at 4.7 GHz
Which do you prefer? The optimal answer is (1) when you have only one thread that is heavily used, (2) when you have two threads, (3) when you have three or four threads, and (4) when you have more than four threads. Turbo can do something like that for your automatically, while leaving the CPU at stock speeds.
Now, with a manual overclock, maybe you can get all cores running at 4.8 GHz. That can be a little better than automatic turbo in situations where you're actually pushing all of the cores. But in a lot of workloads, it will be worse than stock speeds, as turbo state (1) or (2) will be better when you have few threads. If that's all that you can get out of overclocking, then why bother?
One answer is that you didn't get the top bin. You bought a cheaper CPU with a less aggressive stock clock, so its single-core turbo tops out at 4.0 GHz. If you can overclock that to 4.8 GHz, then sure, go ahead.
Turbo isn't perfect yet. But it is getting pretty good. Take about 0.7 MHz off of all of the clock speed numbers above and it's a decent description of a Ryzen 7 2700X. With the numbers as given, it's not that bad of a description of a Core i9-9900K, especially if you didn't spend a lot of money on a cooler.