The Zen Thread
- Thread starter w00key
- Start date
Is there really much of a market for AM5 packed with (entirely with) small cores?
Two identical dies only one of which gets stacked L3 always struck me as a strangely cheap hacky solution, but two different dies, only one of which gets stacked L3, while the other is dense cores, strikes me as optimized.
That's exactly what I meant.
I am not sure if I am whale enough. But I agree that this would be a very interesting CPU model.
A single CPU tile of x8 5c cores, binned for lowpower use, would be a very interesting CPU for all sorts of interesting homelab use cases.
A quick google suggests there is some evidence for the "C" cores being more efficient, but I hadn't really thought of them as being particularly efficient.
Zen cores are already pretty efficient - I'm thinking that if you bin for power efficiencies at middling speeds, rather than ability to reach higher clocks, you might be onto something.
"C" cores are dense, for Cloud compute usage. Can only stuff so much power into a socket, so they have to be somewhat more efficient, or clocked lower, than the regular cores.
I think 8x 5c in AM5 would only make an appearance if the defect rate for CCX's was fairly high or they do what they did with the 5600X3D and release a very late stage product on harvested dies. I fully expect some embedded designs will have a smaller core arrangement though.
But pure 5c packages in a consumer socket would be interesting to to the low(VPS's on LET) to mid(OVH/Hetzner/etc resellers) range hosters, where scale means total BoM of the system is a meaningful percentage of the operational cost. Maybe they can also pick up a bit of the edge compute market, although soldered packages are preferred there for ruggedisation.
I'm still pessimistic on the whole P+E arrangement on desktop. It has potential but OS schedulers still don't seem to make that much good use of them. Benchmarks show a difference sometimes, but my experiments with disabling e-cores on my 14900K didn't reveal any directly observable impact.
Depending on CCD distance with Zen5, AMD will have a better solution thermally though so maybe it'll work a little better without the hot neighbour issue Intel has.
But pure 5c packages in a consumer socket would be interesting to to the low(VPS's on LET) to mid(OVH/Hetzner/etc resellers) range hosters, where scale means total BoM of the system is a meaningful percentage of the operational cost. Maybe they can also pick up a bit of the edge compute market, although soldered packages are preferred there for ruggedisation.
I'm still pessimistic on the whole P+E arrangement on desktop. It has potential but OS schedulers still don't seem to make that much good use of them. Benchmarks show a difference sometimes, but my experiments with disabling e-cores on my 14900K didn't reveal any directly observable impact.
Depending on CCD distance with Zen5, AMD will have a better solution thermally though so maybe it'll work a little better without the hot neighbour issue Intel has.
Is that going to have enough memory bandwidth/core to be worthwhile for plausible workloads?
[referring to a 32 dense core AM5 EPYC model]
For things like photorealistic 3D rendering, bandwidth requirements are very low. For a few types of simulations with lots of computation per data point (I think weather forecasts can fall into that category), bandwidth isn't critical either.
On the other hand, typical high performance computing workloads do have high bandwidth requirements (everything that we map to working with very large matrices). Most perf optimization in this field is about partitioning and re-ordering operations to make best use of caches.
And then there is the large grey area inbetween.
Conclusion: such a machine would have a few very clear wins, a few very clear losses, and an "interesting" middle range where it depends on available manpower to tune software. Maybe some of the perf tuning effort can piggyback off of work that has already been done for the big EPYCs; but don't rely on promises of that. Look at independent benchmarks.
Its hard to get excited about a mixed-bag product that is sold right next to a serviceable, flexible, existing desktop product.
I can see the 2nd die packed with "C" cores while the first die is regular, I can see a mini-APU with everything minimized for price and power, its harder to see the attraction of going all compact cores in full AM5.
I can see the 2nd die packed with "C" cores while the first die is regular, I can see a mini-APU with everything minimized for price and power, its harder to see the attraction of going all compact cores in full AM5.
If that one does get released, it will not be a mainstream product, even if it fits in a mainstream platform. But I wouldn't be surprised if the niche for that thing turns out to be reasonably large.
Smaller scale, lower budget clusters come to mind. A small VFX shop, a tiny research group at some lower profile university, might buy such boxes piecemeal year after year, instead of having a hefty budget for a Threadripper or big Epyc machine.
On another note. There is this rumored Strix Halo, i.e. the big APU with a 256 bit wide bus and a very respectable GPU portion. As the rumors seem to be solidifying, I am wondering if it will be possible to build machines based on that APU where the RAM isn't soldered, but instead attached with the new LPCAMM2 connector standard.
Such a machine could be a beast for GPU workloads with huge VRAM requirements. Edit: depending on availability of high capacity LPCAMM2 modules, of course.
Such a machine could be a beast for GPU workloads with huge VRAM requirements. Edit: depending on availability of high capacity LPCAMM2 modules, of course.
That's an interesting thought. That's he pretty niche, I thought lpcamm was more for laptops. Although obviously if a desktop chip/socket supported it it's be possible. Definitely an interesting thought, especially if they add some extra NPU hardware. Hit some specialized stuff plus the wide iGPU stuff with decent bandwidth and huge amounts of memory for ML model training.
Oh wow that's going to be quite a jump!5700X3D is down to $229 at Amazon, Walmart, Newegg, and B&H. Gonna upgrade the 1800X/Vega 56 and see if it helps.
Jeeze, yeah, even without a GPU upgrade, that's going to be HUGE gains in a number of games like Stellaris or Factorio or the like.
If you want to see the impact that RAM still has on Zen 4, Ancient Gameplays just ran a 5700X3D vs 7500F and 7600X review:
View: https://youtu.be/aGnri9yB5iQ
Stock, a 5700X3D with DDR4-3200 CL16 memory is typically as fast, if not faster, than stock Zen 4 with DDR5-6000 CL36 memory. The only game that is a blowout in favor of Zen 4 is... League of Legends. Everything else, you can just drop a 5700X3D (which is now down to $209) into your existing Zen 3 CPU socket, keep your DDR4-3200 CL16 memory, and get the same or better than Zen 4 performance at the same price point in most cases.
AMD really needs to fix this for Zen 5, and reviewers need to hold AMD's feet to the fire over this abysmal RAM timing behavior of motherboard makers.
View: https://youtu.be/aGnri9yB5iQ
Stock, a 5700X3D with DDR4-3200 CL16 memory is typically as fast, if not faster, than stock Zen 4 with DDR5-6000 CL36 memory. The only game that is a blowout in favor of Zen 4 is... League of Legends. Everything else, you can just drop a 5700X3D (which is now down to $209) into your existing Zen 3 CPU socket, keep your DDR4-3200 CL16 memory, and get the same or better than Zen 4 performance at the same price point in most cases.
AMD really needs to fix this for Zen 5, and reviewers need to hold AMD's feet to the fire over this abysmal RAM timing behavior of motherboard makers.
I'm not sure what the memory angle is here? It's always been the case that the extra cache on the Zen 3 X3D models more than made up the difference to non-cache-equipped Zen 4. But that's L3 cache. What does memory have to do with it? Or are you suggesting that with the right memory Zen 4 could overcome a huge L3? That seems unreasonable.
Over here the cheapest 5700X3D is still £50 more expensive than a 7500F. I mean... Shock. Upgrading an old system to the best (ok, second-best) currently available processor is more effective than buying a new system with the slowest available processor. Given how these things usually go, it's astounding that the cheapest AM5 CPU can keep up.
Over here the cheapest 5700X3D is still £50 more expensive than a 7500F. I mean... Shock. Upgrading an old system to the best (ok, second-best) currently available processor is more effective than buying a new system with the slowest available processor. Given how these things usually go, it's astounding that the cheapest AM5 CPU can keep up.
AMD's fix for "X3D" is called "X3D". It accelerates Ryzen 5000 in games (and a few other things), and it accelerates Ryzen 7000 in games (and a few other things).
For a cost, $ and lower MT performance.
Zen 5 really can use a better memory controller and/or better cache - bigger or more efficient.
The point was that it's odd to compare a Zen3 8-core CPU with 3D cache, which is known to provide a substantial performance uplift in most games, to a Zen4 6-core CPU without 3D cache, and conclude that the memory subsystem of Zen4 is reason why it performs worse in games.
It is also worth noting that 3200 CL16 is 10ns, while 6000 CL36 is 12ns, so it is literally slower memory. DDR5 6000 CL30 is the equivalent grade of memory (which is not faster in latency, but can transfer more in a given time... something which rarely matters much for games compared to latency).
I do not find it odd that the price is lower at many performance points to go with the last generation. This is common, you can get a much better deal on older generations of Intel processors as well. That is especially true if you buy a prebuilt system, but it is also true for components. Cost optimization rarely says to get the latest model of anything.
I do not find it odd that the price is lower at many performance points to go with the last generation. This is common, you can get a much better deal on older generations of Intel processors as well. That is especially true if you buy a prebuilt system, but it is also true for components. Cost optimization rarely says to get the latest model of anything.
Last edited:
The primary timings aren't the problem with DDR5. The problem is the secondary timings:
View: https://youtu.be/dlYxmRcdLVw
Maybe look at the actual graphs and see what happens when a 7500F or 7600X uses "tweaked" memory timings, instead of just assuming it's vcache that's doing all the heavy lifting.
As for the core count difference, it's not relevant for gaming. Hardware Unboxed just re-covered this.
View: https://youtu.be/dlYxmRcdLVw
Maybe look at the actual graphs and see what happens when a 7500F or 7600X uses "tweaked" memory timings, instead of just assuming it's vcache that's doing all the heavy lifting.
As for the core count difference, it's not relevant for gaming. Hardware Unboxed just re-covered this.
Last edited:
I am not going to disagree that subtimings can have an impact, but finding to what extent you can overclock a given memory chip is not exactly a mass market concern. It is probably the one area of manual overclocking which can have an impact without serious stability questions, but it is manual for good reasons. This is not easy to test quickly.
They need timings which work for everything they sell, not just 80% of it or whatever the number is at those timings. If they bin on a dozen metrics, is that even something which can be conveyed to the user? Is it worth that time? XMP or EXPO are not that detailed, if they were would it be easy to determine what to get?
DRAM is also extremely temperature dependent in terms of how often it needs to be refreshed especially, which is another concern when overclocking. The specs are conservative, but usually lower temperatures can be arranged. I also bought expensive memory recently and basically derated it slightly so I could be sure of this. I intentionally let it get hot during testing, and it did not pass. I could almost certainly turn the fan in front of it up a bit and it would run at the advertised speed without problems. Is a relaxed temperature standard fine for high end ram, and how much?
More high end timings in memory standards would be appreciated, but numbers you can hit in volume are mostly what you get. Once you are talking about companies who are binning beyond the standards, it is mostly up to them what they are selling, and what their reputation is.
They need timings which work for everything they sell, not just 80% of it or whatever the number is at those timings. If they bin on a dozen metrics, is that even something which can be conveyed to the user? Is it worth that time? XMP or EXPO are not that detailed, if they were would it be easy to determine what to get?
DRAM is also extremely temperature dependent in terms of how often it needs to be refreshed especially, which is another concern when overclocking. The specs are conservative, but usually lower temperatures can be arranged. I also bought expensive memory recently and basically derated it slightly so I could be sure of this. I intentionally let it get hot during testing, and it did not pass. I could almost certainly turn the fan in front of it up a bit and it would run at the advertised speed without problems. Is a relaxed temperature standard fine for high end ram, and how much?
More high end timings in memory standards would be appreciated, but numbers you can hit in volume are mostly what you get. Once you are talking about companies who are binning beyond the standards, it is mostly up to them what they are selling, and what their reputation is.
< insert lament about externally visible ECC not being a mandatory part of the DDR5 standard >
They gotta keep some kind of simple feature separation between powerful desktops these days and workstations that are much more lucrative.
My issue with how AMD is benchmarked is that most reviewers use a very specific golden sample of RAM for testing. That RAM, out of the box, is on par with "tweaked" timings. You can see what an off the shelf DDR5-6000 CL36 kit does, and that's not the worst outcome. Zen 4 loses around 17% of its performance when using bog standard DDR5-5200, and that's the fastest that AMD specs allows for.
Without that very specific memory kit sent to reviewers, Zen 4 just isn't that good compared to Zen 3 with vcache.
Last edited:
Most of the reviews by sites I would consider reputable are using JEDEC timings for their main reviews, so the specific RAM is not relevant and the motherboard is using timings as calculated from the standard.
Sometimes there is also an overclocked entry for high end processors in reviews which does use overclocked memory and also turns on PBO and maybe undervolts it, but I would think it is a good idea to use a good memory kit in that case as it is trying to show what maximum performance is like.
If the reviews you see are not disclosing overclocking, I would look for better reviews. I am guessing this is a Youtube thing, as if anything I found fewer overclocked reviews than I expected in looking last time (where I basically never use video sites, so would not see those).
The motherboard presets for a given memory module are not guaranteed to work, even if they probably know what they are doing in many cases in terms of what is likely to work for most chips. I do not know how they come up with the lists, but if G.Skill shipped free memory to the vendors along with timing suggestions, that appears to have paid off for them. ASUS at least had a fairly long list of modules by several vendors in their presets when I looked (unfortunately not including the very new at the time 48GB modules I had, so I did need to set everything aside from the XMP values myself.)
That "with vcache" is a fairly important qualifier. You are talking about a much higher end processor from the previous generation, with a specific feature which accelerates some of the workloads you care about. It should win that, it is a better processor, which is better suited to the task.
The DDR4 memory in that comparison is also 10ns compared to the DDR5 at 12. A fair comparison is either CL 19 or 20 on the DDR4, or CL30 on DDR5 (same with other rated timings). Without that change you cannot say much about the memory latency of the CPU anyway, as the DDR4 system is using lower latency memory, and so is expected to win this. It will not always win as by the end of a DRAM burst it will have caught it in wall time, but many tasks do care about first byte latency, and that main timing which it is rated for and advertised at is 20% slower on your DDR5.
Edit: It will not even catch it during a burst looking a bit closer at that, even wall time for the 8th word transferred is slightly lower on that DDR4.
Last edited:
In the actual meaning of the words, a "golden sample" is not something you can buy off the shelf. I am with you on the whole benchmarketing cheat show, but this is a game that a single competitor cannot abandon on their own. As long as one company is being rewarded for cheating, the other competitors will have to cheat likewise.
Of course, if a considerable number of influential reviewers banded together and were to set their own standards based on official specifications[*], then cheating could begin to get called out regularly. And such an agreed standard would also do a lot to transparently distinguish between reviewers, influencers, and freelancer marketeers.
Until then, customers can choose to buy overclocker RAM off the shelf, or put that money to some other use. For example, towards upgrading to an X3D model instead, which renders most DRAM details moot for effective performance.
[*] I know that this is a utopian dream that will never come to pass. But we can decide to spread the suggestion in their comment sections.
Epyc 4000 series announced this morning. Basically the current Ryzen lineup but with ecc support (udimm only) and some added motherboard features. There's a new 4 core/8 thread part but everything else is rebranded Ryzen parts. Clock speeds all feature >5ghz turbo too. Don't think this is going to be the next Opteron 165, but it's AMD going after another market segment.
www.anandtech.com
One More EPYC: AMD Launches Entry-Level Zen 4-based EPYC 4004 Series
The really frustrating part is that even if reviewers and youtubers decided to take the cherry picked parts that are sent out for review, benchmarked them and then did a follow up a few weeks later with retail purchased parts that are actually available to general consumers to see the differences, if any... they risk getting blacklisted by the manufacturers for future pre-release review access and other industry press interaction. It really would have to be a kind of 'union action' style decision for all the press/media types to join together in the interest of accuracy and fairness to consumers to make it really fair.
Fortunately, I think the issue is not super big. Mostly things work more or less as they are claimed to do, even if some of the tippy top of performance is hard to reach without specific parts being used along with occasionally unreliable configurations.
Is this not the distinction between a real review and marketing though?
If you are being paid or deriving benefit such that you do not feel comfortable being entirely honest about the product, it is less a review and more an advertisement, or paid marketing. A good review will attempt to remain unbiased, and those are still easy to find if you look.
If the trade is that you get an early sample but you cannot do an honest review, what you have sold is your reputation.
More specifically, the Epyc 4004 series as far as the Zen 4 models are concerned. What a strange coincidence to borrow the numerical designation of another famous microprocessor. Somebody at AMD must be havig too much fun.
Yeah, but I think the problem is that the reviewers generally don't see too much of a difference most of the time so they are pretty happy to just let it go and call it a manufacturing variance. If they really put work into buying multiple retail examples (RAM, CPU and Mobo setups) and put time into testing... the risk of expense is very high, the risk of loss of relationship with the manufacturer is very high (and risk to future exclusives or well timed pre-release demos, etc.), and the potential benefit is fairly low. Until someone finds real indications of very obvious discrepancies between preview hardware for press reviews and retail products, it remains a theoretical problem and when even the low hanging fruit is probably a few percent difference in performance, they're just not going to take all those risks and incur costs just to come up with, "Eh, it's a bit different but not much. Could be just a glitch."
The times when there are real changes in preview or initial release product vs. retail or v1.1 hardware, the manufacturers tend to get caught... eventually. Look at SSDs that switch from TLC to QLC or omit the DRAM or change controllers without an obvious hardware version bump or whatever. There are people watching and they catch on after a bit.
I think the situation can be covered mostly with the idea that CPUs are more complex than ever, RAM is more complex than ever, motherboards are more complex than ever and the motherboard makers are especially incentivized to push the limits of performance with 'out of the box' settings that are not particularly sane/safe for all hardware combinations. Honest marketing should cover that to some degree, but until they are legally required to do so, they will let it slide and the press people will probably not dig into it unless someone gets vocal about it. Preferably someone with very little to lose.
You know 3 or more different reviewers have done this? Gone out a bought a dozen to 20 duplicates of multiple different CPUs and tested them?
Hardware Unboxed
Gamers Nexus (fairly recent)
View: https://www.youtube.com/watch?v=PUeZQ3pky-w
LTT
The typically variation was a couple of % performance wise and a bit more on power-use (required to reach that performance).
Heck Der8auer even commented on one of those discussions about the amount of variation he'd seen in ~4,000 CPU's of the same class for overclocking.
There doesn't seem to be massive variation between what is reviewed (ignoring extreme overclocking) and actually acquired if you go out and purchase the same combo of parts that were reviewed. Buildzoid has had a couple of digs at motherboard manufacturers though and if I remember correctly seems to be of the view that they'll certify a motherboard as rated for 8000MT if they can find a single CPU out of a large pile that allows them to hit that.
The claim that CPU review samples are exceptionally good silicon is false and has been debunked multiple times. I think IceStorm's earlier comment of "golden samples" was aimed at highly binned overclocker RAM kits that are often handed out along with the CPU review samples.
I agree that using such RAM is showcasing the CPUs in a non-representative scenario. But I wouldn't call that RAM a golden sample, because it can be bought off the shelf, at a hefty premium. And it comes with all the consequences of overclocking, both good and bad.
BTW, from a 10000ft view it is always statistically unlikely that early review samples are golden samples. The product is new and fabrication has seen just enough tuning to meet a target spec. The overall number of units produced is still fairly low. Both of these constraints mean that there simply don't exist many specimens yet, that perform drastically above the mean.
As the production run continues on, design and fabrication get more finetuning. And the total number of units keeps growing. This gradually increases the number and quality of samples that perform above the spec. Eventually, once almost every single sample is a golden one, the spec can be updated to shift to the new mean. So we get a product refresh.
I agree that using such RAM is showcasing the CPUs in a non-representative scenario. But I wouldn't call that RAM a golden sample, because it can be bought off the shelf, at a hefty premium. And it comes with all the consequences of overclocking, both good and bad.
BTW, from a 10000ft view it is always statistically unlikely that early review samples are golden samples. The product is new and fabrication has seen just enough tuning to meet a target spec. The overall number of units produced is still fairly low. Both of these constraints mean that there simply don't exist many specimens yet, that perform drastically above the mean.
As the production run continues on, design and fabrication get more finetuning. And the total number of units keeps growing. This gradually increases the number and quality of samples that perform above the spec. Eventually, once almost every single sample is a golden one, the spec can be updated to shift to the new mean. So we get a product refresh.