Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

AMD launches everything at once; review sites overwhelmed

QuizzicalQuizzical Member LegendaryPosts: 25,483
Pity the poor reviewers.  Ideally, you'd like to have one product in for review at a time.  Test it for a while, write up a review, and then post it at launch day.  Then get the next product to review, which doesn't launch until sometime later.  AMD pretty much nixed that approach with today's launch of not one, not two, but three desktop product lines all on the same day.  Sites that just run some canned benchmarks and post the results have got reviews up, but some of the sites that do higher quality reviews with more in-depth analysis aren't done with everything yet.

Let's start with the one that wasn't even rumored:  desktop Picasso.  That is, the desktop version of the laptop chips that launched several months ago.  This is just a refresh of Raven Ridge.  Basically, take the Ryzen 3 2200G and Ryzen 5 2400G and add a few hundred MHz.  That gives you the Ryzen 3 3200G and Ryzen 5 3400G, now available for sale on New Egg.  There is technically more to it than that, as it's Zen+ cores on a GF 12 nm process node, but that's the gist of it.  That's not the most exciting news, so good luck finding a review of it with everything else launching today.

Next is third generation Ryzen, in the desktop parts without an integrated GPU.  This is a move to a Zen 2 architecture and a TSMC 7 nm process node, both of which are big news.  AMD has caught or perhaps even slightly beat Sky Lake (and hence Sky Lake Refresh Refresh Refresh) in IPC.  In other words, if you take an AMD CPU and an Intel CPU and cripple them to disable all but one core and run them at the same clock speed, the AMD CPU will commonly beat the Intel one outright.

The problem is that AMD hasn't caught Intel on clock speeds.  The Ryzen 9 3900X that launched today has a max turbo of 4.6 GHz, while the Core i9-9900K will turbo up to 5.0 GHz.  That tends to make the latter a little faster in single-threaded performance.  Still, even if you compare it to the Ryzen 7 3700X that AMD distributed for review samples so that they're both 8 core, 16 thread CPUs, AMD tends to win at multi-threaded benchmarks and often by a lot.  And while using less power, too.

One reason for this is that it takes the Core i9-9900K a whole lot of power to run that one core at 5.0 GHz.  If you want to push all of the cores, it has to clock them much lower.  A Ryzen 7 3700X isn't going to run all of its cores at 4.4 GHz at once, either, but it can come a whole lot closer to it than the Core i9-9900K.  If you're commonly doing stuff that pushes a lot of CPU cores, Intel's mainstream consumer CPU line is now pointless.  The upcoming launch of third generation Threadripper will probably make Intel's Sky Lake-X line completely pointless no matter what you're doing.

But what about gaming?  Here, if you make things severely CPU bound, the top end Intel CPUs do beat the new AMD CPUs substantially more often than not.  AMD does manage to win outright at some games, though.  Even when Intel still wins, it's a whole lot closer than it was with the previous generation AMD CPUs.  Furthermore, making the game CPU bound often pushes frame rates well into the hundreds.  If you're heavily into e-sports and running a 240 Hz monitor, then yes, you still want a Core i9-9900K.

The problem is that Intel's clock speed advantage at the top end doesn't benefit you unless you actually grab that top end processor.  The Core i9-9900K and Core i7-9700K both cost over $400 right now.  If you need a CPU for under $370, Intel doesn't offer anything with a max turbo above 4.7 GHz.  Under $230, make that 4.1 GHz.  And then Intel's clock speed advantage disappears.  (Prices are based on New Egg right now.)  Intel's single-threaded advantage at the $500 price point doesn't translate to a comparable advantage at the $100 or $200 or $300 price points that most CPU buyers are more interested in.

The are some important architectural differences between third generation Ryzen and Sky Lake Refresh Refresh Refresh that I'd like to highlight.  Performance of L1 and L2 caches look similar, though AMD has a larger L2 cache per core.  It's when you go to L3 cache or memory that they're very different.  Intel has a single, unified 16 MB L3 cache on the Core i9-9900K.  AMD has two separate 16 MB L3 caches on their new die, one for each of the 4-core core complexes.  For the Ryzen 9 3900X, it's four such L3 caches.  For single-threaded performance, that doesn't matter, but if you're using both core complexes, that means that AMD has far more L3 cache bandwidth available than Intel.  If the core complexes are doing completely independent things (so that it won't tend to be the same data cached in both L3 caches), that also effectively means it can cache far more in L3 cache.  That plays a huge role in AMD tending to win at highly threaded benchmarks.

There's also a big difference between the memory controllers.  Intel mostly sells large, monolithic chips.  That means that when you miss L3 cache and have to go to DDR4 memory, it has to move data around within a chip to get from the CPU core to the memory controller, but the memory controller is on the same chip.  Third generation Ryzen doesn't do that.  The I/O die is a physically separate die from the CPU core die, so it has to hop to a different chip to get to the memory controller.  That hop is over some sort of interposer inside of the socket, so it can offer a lot of bandwidth and not add that much latency.  But it does add latency.

As compared to second generation Ryzen, Sky Lake and its many refreshes already had lower memory latency.  As compared to third generation Ryzen, that advantage widens.  Still, it's only adding another 10 ms or so, not the 70 ms that first and second generation Threadripper added when you had to connect to a memory controller based on a different die.  We knew that hopping to a different die was going to add latency, and it was only a question of how much.  I'd regard this as a relatively favorable result for AMD.  But it means that Intel does have a substantial advantage over AMD in memory latency.

AMD tried to counter this by being able to have far more memory reads in flight than Intel.  That will benefit AMD considerably in highly threaded workloads, but not so much in single-threaded ones.  If you make a simple benchmark where you read a random memory address, and that tells you where to go for the next read, then read the next memory address, and so forth, then Intel will win, and by a lot.  Or at least, that will happen if you make it over a large enough table to rarely hit any on-die caches.  But if you do 50 of those at once, then AMD will win, as even if the latency for each read is higher, they can have a lot more in flight at once, so every nanosecond of waiting has the clock ticking for perhaps twice as many memory reads in flight.
Gdemami

Comments

  • QuizzicalQuizzical Member LegendaryPosts: 25,483
    From an investor perspective, the real question is, what does this tell us about EPYC Rome?  The desktop market matters some, but the x86 server market matters a whole lot more.  And there, the news is pretty favorable.  the Ryzen 9 3900X has two different CPU dies as well as an I/O die, but AMD is still able to keep memory latency reasonable.  Yes, it's higher than Intel's consumer line, but it's much, much better than previous EPYC CPUs.  That's not a complete solution on memory latency, but if the architecture's one weak point is not that much worse than Intel, that points toward a killer chip.

    People are debating whether AMD can get 10% or 15% or whatever of the x86 server market.  I say that's a failure of imagination.  For the last 13 years or so, AMD wasn't able to dominate the server market because their chips weren't good enough.  In the early days of Opteron, they weren't able to dominate it because they couldn't produce enough chips.  EPYC Rome will suffer neither of those problems, and Intel doesn't have a credible counter on the horizon.

    Never mind whether AMD can get to 10% or 15% of the x86 server market.  Can they get an outright majority of it?  If customers try AWS instances based on Intel and on AMD and find that the AMD one offers more performance for less money, things could shift very, very fast.  Today, Intel has a market capitalization of around $200 billion, while AMD is around $30 billion.  If AMD takes half of the x86 server market, should Intel still be worth more than AMD?  It's going to take a bit of time for AMD to be able to ship EPYC Rome in full volume with plenty of partners able to offer full servers, but this time next year, AMD should be practically printing money.

    And then there is Navi, the new GPU architecture.  I don't want to say too much about that yet, as I'm waiting on reviews from some additional sites.  Early reviews say it typically runs about even with a GeForce RTX 2070, and handily beats a Vega 64.  It costs less than the RTX 2070, but does use a little more power.  The die size is 251 mm^2, which is significantly smaller than was rumored.

    If one were to completely ignore all of the rumors and leads leading up to the launch, I'd regard the stats of Navi based on early reviews as being okay to mildly unfavorable.  Overall, it looks like AMD has gained enormously on Nvidia in architectural efficiency, but not caught up.  Nvidia can do a die shrink to 7 nm, too, after all, and will then be ahead in a lot of efficiency metrics, though probably not performance per mm^2 if they insist on keeping the tensor core junk around.  On the other hand, this is a substantially more favorable result than one would have expected from rumors a week ago.
    Gdemami
  • CleffyCleffy Member RarePosts: 6,413
    There are several things with the server market that gives Intel time if they want to keep the majority. First any company adopting AMDs server CPUs will need time to thoroughly test them with the intended software they will be using. Second almost everything in the server market is written for Intel CPUs so they will naturally have an advantage in execution outside of a few niche cases were even bulldozer was a better option. Third companies tend to not switch CPUs for a very long time compared to the consumer market. I remember working for a company that still ran their server on Windows XP with Intel Pentium processors in 2014.
    I think there will be some big companies that adopt the CPU outright and rewrite their software to work with the chip, but I don't see a majority shift anytime soon.
  • RidelynnRidelynn Member EpicPosts: 7,383
    Quizzical said:

     The die size is 251 mm^2, which is significantly smaller than was rumored.

    Reddit had it pegged pretty closely.

    https://forums.mmorpg.com/discussion/481576/amd-says-navi-is-all-new-rdna-architecture-launching-in-july-as-radeon-rx-5000-series#latest

    Not sure what rumors had it being significantly larger than that.
  • CleffyCleffy Member RarePosts: 6,413
    edited July 2019
    I think they chose a bad strategy with the 5700 cards. They clocked them a lot higher to compete against NVidia instead of as a replacement to the RX 580. This in turn makes them a lot more inefficient than they should be for a mid-range 7nm card. I think you mentioned before that in order for AMD to compete against NVidia, they purposely run their cards hot instead of running them efficiently. As we saw with Vega cards, they are using a lot more energy for a small 10% change in performance. 
    As a replacement for the RX580, the 5700 XT is performing about 50%~100% better. That's about inline with what one would expect out of a die shrink. If AMD instead went for that 10% performance hit for better thermals, I think it would make more sense for a mid-range part. Leave competing with NVidia up to time and drivers. As we know when games embrace more DX12 and open standards, the AMD parts tend to see tremendous gains.
  • QuizzicalQuizzical Member LegendaryPosts: 25,483
    Ridelynn said:
    Quizzical said:

     The die size is 251 mm^2, which is significantly smaller than was rumored.

    Reddit had it pegged pretty closely.

    https://forums.mmorpg.com/discussion/481576/amd-says-navi-is-all-new-rdna-architecture-launching-in-july-as-radeon-rx-5000-series#latest

    Not sure what rumors had it being significantly larger than that.
    Slightly further up on that thread estimates 275 mm^2.  251 is pretty close to 255, but not as close to 275.
  • QuizzicalQuizzical Member LegendaryPosts: 25,483
    Cleffy said:
    I think they chose a bad strategy with the 5700 cards. They clocked them a lot higher to compete against NVidia instead of as a replacement to the RX 580. This in turn makes them a lot more inefficient than they should be for a mid-range 7nm card. I think you mentioned before that in order for AMD to compete against NVidia, they purposely run their cards hot instead of running them efficiently. As we saw with Vega cards, they are using a lot more energy for a small 10% change in performance. 
    As a replacement for the RX580, the 5700 XT is performing about 50%~100% better. That's about inline with what one would expect out of a die shrink. If AMD instead went for that 10% performance hit for better thermals, I think it would make more sense for a mid-range part. Leave competing with NVidia up to time and drivers. As we know when games embrace more DX12 and open standards, the AMD parts tend to see tremendous gains.
    If you'd like for the cards to be more efficient, it's easy to throttle them back in Radeon WattMan.  That's what I did with my Vega 64.
    Gdemami
  • RidelynnRidelynn Member EpicPosts: 7,383
    Undervolting was a very popular strategy with Vegas. Not sure how it will play out with 5x00's yet.

    RDNA isn't horrible with power. It's a lot better than GCN, that's for sure.

    You can tell they did squeeze the XT a bit for additional headroom, but look at the 5700, it's more power efficient than anything nVidia has in that performance bracket. Granted, it has a process node advantage on nVidia, but hey, whatever it takes. 

    I can't say I'm disappointed. We are looking at mid-range parts. They are giving pretty good performance. I'm still worried that all our price points have shifted upwards significantly from what they had historically been, but maybe that's just me with my rose colored glasses and Mining broke the old price structure forever.
    [Deleted User]
  • Asm0deusAsm0deus Member EpicPosts: 4,600
    Ridelynn said:
    ...snip...

    We are looking at mid-range parts. They are giving pretty good performance. I'm still worried that all our price points have shifted upwards significantly from what they had historically been, but maybe that's just me with my rose colored glasses and Mining broke the old price structure forever.
    Oddly enough I have argued the same thing before.
    ;)

    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.





  • OzmodanOzmodan Member EpicPosts: 9,726
    Well seems that Intel is preparing a new set of processors aptly named Panic Lake:

    https://www.forbes.com/sites/antonyleather/2019/07/10/intel-plans-to-end-amds-ryzen-rampage-13-hyper-threaded-processors-including-5-2ghz-10-core-beast/#3955c73c1f2a

    Of course you will have to buy a new motherboard with it, unlike the AMD processors.
  • VrikaVrika Member LegendaryPosts: 7,973
    Ozmodan said:
    Well seems that Intel is preparing a new set of processors aptly named Panic Lake:

    https://www.forbes.com/sites/antonyleather/2019/07/10/intel-plans-to-end-amds-ryzen-rampage-13-hyper-threaded-processors-including-5-2ghz-10-core-beast/#3955c73c1f2a

    Of course you will have to buy a new motherboard with it, unlike the AMD processors.
    It's Comet Lake, not Panic Lake.

    But Forbes already edited their article to reflect that it looks like it's a fake.
     
  • OzmodanOzmodan Member EpicPosts: 9,726
    edited July 2019
    Vrika said:
    Ozmodan said:
    Well seems that Intel is preparing a new set of processors aptly named Panic Lake:

    https://www.forbes.com/sites/antonyleather/2019/07/10/intel-plans-to-end-amds-ryzen-rampage-13-hyper-threaded-processors-including-5-2ghz-10-core-beast/#3955c73c1f2a

    Of course you will have to buy a new motherboard with it, unlike the AMD processors.
    It's Comet Lake, not Panic Lake.

    But Forbes already edited their article to reflect that it looks like it's a fake.
    Aw...you did not like my attempt at humor....
    Ridelynn
Sign In or Register to comment.