AMD Announces Ryzen 7 5800X3D Shipping On 4/20, New Mainstream CPUs

[drakonas777]

Yup. The data fully supports this. Not only did it increase perf/area and perf/$, but also perf/W.

It's not easy to see the latter, because the damn P-cores boost so high. However, if they'd replaced those two quad-Gracemont clusters with two more Golden Cove P-cores, I'm sure the net result would be not only worse perf, but also worse perf/W.

It's mostly multithreaded perf/area, which directly translates into multithreaded perf/$. Perf/W gains on the desktop as is are pretty insignificant actually, unless as you mentioned we are talking some more theoretical situation where only a lot of E cores are used for extremely well parallelized workload most of the time. But this would be a niche scenario for the mainstream desktop, since a lot of applications still benefit from a few strong cores rather many weak ones, especially games, which I think is still the main application for the mainstream desktop.

All this hybrid stuff is actually nothing more than manufacturing cost and multithreaded benchmarks optimization for the Intel. You could say that the result of that is cheaper MT performance for the user and I'd agree, but in reality this is relevant for quite a small group of mainstream desktop customers who often use CPU based rendering/en/decoding. And even in this case the real benefit will show only with upcoming Intel products where E core count will be increased dramatically, because for now the amount of them is far too low to get even this benefit. Let's take an i7 for example, 8 Ps + 4 Es. Those 4 Es are basically useless. Adding more L3 cache (or implementing extra sutff in big cores) instead them would boost overall performance more that few relatively weak cores.

I guess my main point is hybrid architecture on the desktop is mostly beneficial specifically for MT-heavy load. It's not adding much value for all universally. However, if we going to get say 8C + 32Es, OK, then I can see at least practical benefit for the some customers.

[Jabberwocky]

Would you tell us why?

More testing and research needed on my part so take it with a grain of salt. ( Teggs You might want to take a look at this too.)

The 5800 and 5900 (non-X) with stepping B2 are rated for 65W. In other words lower temps and less fan noise. As per the references below it appears that some of the other Ryzen 5000 parts also have improved power-usage and memory-overclocking. I don't know why AMD did not do another Ryzen XT release.

https://www.guru3d.com/news-story/ryzen-5000-b2-stepping-has-lower-consumption-and-lower-temperatures.html

https://skatterbencher.com/2021/09/19/skatterbencher-29-amd-ryzen-9-5900-b2-overclocked-to-5152-mhz/

https://www.youtube.com/watch?v=i2OLNAlMOoY - AMD Ryzen 9 5950X (B2 Stepping) On $45 A320 Mainboard

I tried to find the serial or manufacture-interval references but not 100% sure on that either. If you want to go down the rabbit hole: https://www.overclock.net/threads/ry...dates.1796602/

I don't like it and I don't expect it'll work.

I can't say that I like it either. I was suprised to not see anything about it here, seems news worthy IMO.

Stacking cache dies on top of your cores isn't exactly great for keeping them cool.

True. I was wondering if you used LN2 if you could in theory still overclock it (might want to attempt a record etc). Turns out LN2 probably won't solve the problem.

It took a while to find an response, but here it is from an AMD employee: https://youtu.be/qKQAY0k8f1E?t=2726

[coder]

More testing and research needed on my part so take it with a grain of salt. ( Teggs You might want to take a look at this too.)

The 5800 and 5900 (non-X) with stepping B2 are rated for 65W. In other words lower temps and less fan noise. As per the references below it appears that some of the other Ryzen 5000 parts also have improved power-usage and memory-overclocking. I don't know why AMD did not do another Ryzen XT release.

https://www.guru3d.com/news-story/ryzen-5000-b2-stepping-has-lower-consumption-and-lower-temperatures.html

https://skatterbencher.com/2021/09/19/skatterbencher-29-amd-ryzen-9-5900-b2-overclocked-to-5152-mhz/

https://www.youtube.com/watch?v=i2OLNAlMOoY - AMD Ryzen 9 5950X (B2 Stepping) On $45 A320 Mainboard

I tried to find the serial or manufacture-interval references but not 100% sure on that either. If you want to go down the rabbit hole: https://www.overclock.net/threads/ry...dates.1796602/

Thanks for the info.

Back when new CPU launches were further apart, I used to benefit from things like this by waiting until at least mid-cycle to buy new machines. By then, most of the bugs had been shaken loose and both boards & CPUs had at least a few major steppings. I lucked into just such a scenario with my Phenom II, which was a stepping that significantly improved its temperatures, etc. So, in a way, this is sort of like a blast from the past.

As for why no XT? Well, Zen4 is coming soon and AMD's product stack is now pretty well fleshed-out. Maybe they thought it was too many SKUs or would force down prices on the non-XT models more than they'd make up for with higher XT pricing. Alder Lake now puts a ceiling on their pricing, so the main price movement from adding SKUs is probably downward.

[rclark]

You could instead buy a board that has a BMC chip, but those are very expensive and typical BMC performance is horrendously bad, if you actually need to do a significant amount of GUI stuff, locally. I have an ASRock Rack board with an ASpeed 2500. Let me tell you: "speed" doesn't belong anywhere in the name.

... For my use, I don't think I need something like that. ECC would be nice, but.... My server motherboard is a Gigabyte B450 Auros M with the 2200G onboard with 16G of memory. It only manages about 1.8TB of data on two 2TB SSDs, and OS installed on another 500GB SSD. No monitor/keyboard/mouse attached to the mid-tower case... unless I need to update the BIOS and then I pull out the spares for that. I can remote desktop into it if necessary if needed which I rarely do. Most everything can be done with SSH. That's why I don't even want a 'cheap' card installed in the box. What I have is really quite fast enough for just serving files, but there is something in me that would like to install a 'later' version of the same class processor on the MB because I can. But no such processor available... Silly I know.

[coder]

No monitor/keyboard/mouse attached to the mid-tower case... unless I need to update the BIOS

That's one of the neat tricks about my ASRock Rack board. You can update its BIOS using the BMC's web interface, even when the machine's main CPU won't POST!

I had a perfect case for this, because it's a Haswell board and I had a Haswell-Refresh CPU that wasn't supported by the original BIOS. Unfortunately, I didn't know about that capability and managed instead to borrow a 1st gen Haswell CPU to do the BIOS update the standard way.

[billyswong]

It's mostly multithreaded perf/area, which directly translates into multithreaded perf/$. Perf/W gains on the desktop as is are pretty insignificant actually, unless as you mentioned we are talking some more theoretical situation where only a lot of E cores are used for extremely well parallelized workload most of the time. But this would be a niche scenario for the mainstream desktop, since a lot of applications still benefit from a few strong cores rather many weak ones, especially games, which I think is still the main application for the mainstream desktop.

All this hybrid stuff is actually nothing more than manufacturing cost and multithreaded benchmarks optimization for the Intel. You could say that the result of that is cheaper MT performance for the user and I'd agree, but in reality this is relevant for quite a small group of mainstream desktop customers who often use CPU based rendering/en/decoding. And even in this case the real benefit will show only with upcoming Intel products where E core count will be increased dramatically, because for now the amount of them is far too low to get even this benefit. Let's take an i7 for example, 8 Ps + 4 Es. Those 4 Es are basically useless. Adding more L3 cache (or implementing extra sutff in big cores) instead them would boost overall performance more that few relatively weak cores.

I guess my main point is hybrid architecture on the desktop is mostly beneficial specifically for MT-heavy load. It's not adding much value for all universally. However, if we going to get say 8C + 32Es, OK, then I can see at least practical benefit for the some customers.

And then there is also the bottleneck caused by managing too many parallel threads. Sometimes threads / tasks need to be synchronized or resources need to be locked / unlocked. Therefore, the maximum number of cores that can utilized efficiently are also limited by the speed of P-cores / main memory / IO latency. When that core count is reached, E-cores lost their appeal again in deskstop / workstation / server use.

Hybrid CPUs are welcomed in mobile in part for their idle power saving. Those "8Ps + 4Es" Intel CPUs are probably for the same purpose, pointless for desktop but useful for laptop.

[coder]

And then there is also the bottleneck caused by managing too many parallel threads. Sometimes threads / tasks need to be synchronized or resources need to be locked / unlocked.

That depends on what you're doing. There are lots of parallelizable tasks that aren't bound by lock-contention. However, memory bottlenecks can definitely start to limit scaling.

Therefore, the maximum number of cores that can utilized efficiently are also limited by the speed of P-cores / main memory / IO latency. When that core count is reached, E-cores lost their appeal again in deskstop / workstation / server use.

Hybrid CPUs are welcomed in mobile in part for their idle power saving. Those "8Ps + 4Es" Intel CPUs are probably for the same purpose, pointless for desktop but useful for laptop.

Using the heuristic that their E-cores are roughly half as fast as their P-cores, it comes down to a question of whether you'd rather have an 8-core CPU or a 10-core CPU. Most people wouldn't regard that distinction as meaningless.

Source: https://www.anandtech.com/show/17047...d-complexity/7

In actual fact, single-thread performance of their E-cores is 64.5% or 54.1% of a P-core, for SPECint2017 or SPECfp2017, respectively. So, a 8+4 core combo is more like 10.6 or 10.2 P-cores, for int or fp-dominant workloads.

You guys seem to regard these E-cores as garbage, but they're really not bad. Intel claims they have similar per-clock performance as Skylake cores.

[drakonas777]

You guys seem to regard these E-cores as garbage, but they're really not bad. Intel claims they have similar per-clock performance as Skylake cores.

E cores are certainly not a garbage. Personally I think they are pretty cool, and assuming the rumors regarding 8E cores-only mobile Pentium series are true - they potentially may be amazing price/performance for budget notebooks, miniPCS, SBC's and whatnot. Also Intel most likely will develop them further, so instead that weak old facion atom-like architecture line we will get more like a small and quite full feature "core series" version of the core. That's nice.

My main criticism is for the AL desktop SKUs specifically, and partially it's not even that technical, but matter of use case and personal taste. It's obvious that hybrid architecture has some advantages. Nobody is denying that. But for me, AL desktop, as it is right now, well, it's just not a clean and elegant solution. I get the reasoning why it's made the way it is, but still. The fact that you like have a full AVX512 support in P core (which by the way costs not a trivial amount to transistors to have), which is artificially disabled is somewhat workaroundish solution. Yeah I get that you do not need those often but it's just not elegant. I hope Intel will implement full features ISA for E's in the future, so no such workaroundish solutions will be necessary and we'll have truly big.Little like ARM does.

BTW, AFAIK AMD will use hybrid architecture for future APU line only, ZEN5 + ZEN 4. At least all rumors I saw were related to APUs. Yea, it's kind of hilarious that ZEN4 will be "the small one", so it's says something about the size of ZEN5 I guess . Anyway currently AMD's philosophy (as per some interview, now forgot exactly which) is to use like "medium" size, but similar cores. So it would not supersize me if ZEN4 will be a little slower than Raptor Lake's P core. Anyway, we'll see.

[coder]

But for me, AL desktop, as it is right now, well, it's just not a clean and elegant solution.

Yup, it lacks simplicity and symmetry. Worse, there are obvious pitfalls.

The fact that you like have a full AVX512 support in P core (which by the way costs not a trivial amount to transistors to have), which is artificially disabled is somewhat workaroundish solution.

True. I'm sure there were heated battles fought at Intel over this decision. After years of singing the virtues of AVX-512 and optimizing lots of software for it, the last thing some of them wanted to do was surely to step back from it.

I hope Intel will implement full features ISA for E's in the future, so no such workaroundish solutions will be necessary and we'll have truly big.Little like ARM does.

ADL is like what ARM does. It's just that they had to cut down the Big core to get there, instead of building up the small core (which they did somewhat, with the addition of AVX2).

I expect future generations of E-cores will have AVX-512. Maybe not in Raptor Lake, which I think is made on the same manufacturing node and possibly even uses the same E-cores, but probably Meteor Lake (which is slated to use the "Intel 4" manufacturing process).

it would not supersize me if ZEN4 will be a little slower than Raptor Lake's P core. Anyway, we'll see.

As long as AMD lacks a proper E-core, they'll have to make more compromises in their micro-architecture than Intel does. I think Zen 4 will offer better perf/W, but maybe not the absolute best performance against Raptor Cove.

[Duve]

since gaming is what they're really targeting with these chips (non-gaming purposes are still likely going to be dominated by Alder Lake)

You would be wrong, on top of gaming applications... be V-cache also seems to be focused on cloud deployments. The Milan-X is an Epyc chip after all and released first.