Review AMD Ryzen 9 9900X3D Review: 3D V-Cache's Middle Child

[teeejay94]

Questionable pricing is people who are selling the 5800X3D and 7800X3D for 600$ still , while the brand new 9800X3D is 640 or so dollars Canadian on NewEgg. If anything I think the second hand market has lost their minds, how is a 7800X3D, which doesn't perform as well as its younger sibling, how is that worth the same amount of money as a 9800 x3d, the answer is it isn't you can't have two CPUs of different generations be the same price it just doesn't make sense and we're not going to start following that pricing model just because a bunch of people in the second hand want to get their money back

 

[abufrejoval]

Sometimes the belly button perspective from gamer publications has me boil over: I guess today is one of those days.

I know headlines need to sell your content, but when you're smart enough to note "AMD has been curiously reluctant for the press to test the Ryzen 9 9900X3D", then you should be smart enough to take the hint: this product isn't targeting your audience.

And that's why AMD isn't giving you those sample, so you don't go about and spread opinions which in their wisdom they deemed ignorant and misleading.

And I am sorry to say: they were right not to give you those samples because you're wrong judging the price performance value of those chips, for those who buy the majority of them.

If they really were bad value, the price would change. AMD really hates putting chips into landfills (Llano!)

You and I and most readers of TH wouldn't and perhaps shouldn't buy these chips, because those €100 savings for that gaming rig we're putting together may come with compromises we don't really want to face down the road: overcapacity seems more attractive than max economy.

But for bulk buyers e.g. of game developing workstations, those €100 saved may add up to significant savings for the planned duration of a mass deployment, even a bonus, while those CPUs will do an excellent job at their intended target, much better than what they replace.

And the main advantage of those bulk buyers is that they can predict much better the workloads they'll be putting on them, e.g. because they design the product and don't see as much benefit in overprovisioning as you and I might see, with workloads much less known yet.

Note that while V-cache does indeed seem to help with game performance, chances of six-to-eight parallel threads consistently benefiting from V-cache in a game-experience-changing manner actually still is rather slim: such a game wouldn't sell, unless 8-core V-cache CPUs were they only thing around. And even then, it would be hard to do.

So having "only" six V-cache cores isn't as much of a practical impediment as you might want to make it, as long as the critical threads stay with those caches and there is enough head-room for all the other threads not to preempt the critical ones from the cache cores.

Which brings the focus to the benefit of those extra lesser-cache cores: and there all that time I've spent waiting on shader re-compilations even on my 16-core machines at least provides a hint: more cores mean less wait which means productivity.

It's not quite significant enough for me to go EPYC and when there are hundreds or thousands of developers working on huge games, the price per seat becomes a rather big lump sum worth a matching product, EPYCs for a few, matching spreads for others and then perhaps a crowd significant enough to match those 9900X3D chips, pepole who will run the game without hitting a performance wall with "only 6" V-cache cores, yet also design and compile assets in a way where those extra cores save enough time to make them worth that price point.

Not that I believe these 6-core CCDs are ever really a design target: they remain chips that failed the top bins, but they pile up to numbers where turning them into a SKU makes sense for a niche, even with V-cache and on EPYC die carriers: another reason not to push them as consumer mainstream products...

If everybody believed six core Ryzens are all you need for top gaming, that would destroy AMDs bottom line based on an 8-core design!

Now, that actual price point may never actually be that close to MSRP, but your headline suggestion that the price needs changing, lacks substance to the point where I'd say it's more wrong than the price.

And in fact quite a few people buying only CPUs with fully enabled CCDs may be overspending in terms of actual need... mostly feel better that they'd done the best they could in terms of performance ...sacrificing a bit of economy.

 

[JarredWaltonGPU]

Sometimes the belly button perspective from gamer publications has me boil over: I guess today is one of those days.

I know headlines need to sell your content, but when you're smart enough to note "AMD has been curiously reluctant for the press to test the Ryzen 9 9900X3D", then you should be smart enough to take the hint: this product isn't targeting your audience.

And that's why AMD isn't giving you those sample, so you don't go about and spread opinions which in their wisdom they deemed ignorant and misleading.

And I am sorry to say: they were right not to give you those samples because you're wrong judging the price performance value of those chips, for those who buy the majority of them.

If they really were bad value, the price would change. AMD really hates putting chips into landfills (Llano!)

You and I and most readers of TH wouldn't and perhaps shouldn't buy these chips, because those €100 savings for that gaming rig we're putting together may come with compromises we don't really want to face down the road: overcapacity seems more attractive than max economy.

But for bulk buyers e.g. of game developing workstations, those €100 saved may add up to significant savings for the planned duration of a mass deployment, even a bonus, while those CPUs will do an excellent job at their intended target, much better than what they replace.

And the main advantage of those bulk buyers is that they can predict much better the workloads they'll be putting on them, e.g. because they design the product and don't see as much benefit in overprovisioning as you and I might see, with workloads much less known yet.

Note that while V-cache does indeed seem to help with game performance, chances of six-to-eight parallel threads consistently benefiting from V-cache in a game-experience-changing manner actually still is rather slim: such a game wouldn't sell, unless 8-core V-cache CPUs were they only thing around. And even then, it would be hard to do.

So having "only" six V-cache cores isn't as much of a practical impediment as you might want to make it, as long as the critical threads stay with those caches and there is enough head-room for all the other threads not to preempt the critical ones from the cache cores.

Which brings the focus to the benefit of those extra lesser-cache cores: and there all that time I've spent waiting on shader re-compilations even on my 16-core machines at least provides a hint: more cores mean less wait which means productivity.

It's not quite significant enough for me to go EPYC and when there are hundreds or thousands of developers working on huge games, the price per seat becomes a rather big lump sum worth a matching product, EPYCs for a few, matching spreads for others and then perhaps a crowd significant enough to match those 9900X3D chips, pepole who will run the game without hitting a performance wall with "only 6" V-cache cores, yet also design and compile assets in a way where those extra cores save enough time to make them worth that price point.

Not that I believe these 6-core CCDs are ever really a design target: they remain chips that failed the top bins, but they pile up to numbers where turning them into a SKU makes sense for a niche, even with V-cache and on EPYC die carriers: another reason not to push them as consumer mainstream products...

If everybody believed six core Ryzens are all you need for top gaming, that would destroy AMDs bottom line based on an 8-core design!

Now, that actual price point may never actually be that close to MSRP, but your headline suggestion that the price needs changing, lacks substance to the point where I'd say it's more wrong than the price.

And in fact quite a few people buying only CPUs with fully enabled CCDs may be overspending in terms of actual need... mostly feel better that they'd done the best they could in terms of performance ...sacrificing a bit of economy.

Talk about having a belly button perspective! The whole point is that the 12-core isn’t a great deal relative to the 16-core or the 8-core X3D chips. Ever. Period. There’s nothing you can do to actually make it the better choice.

Do you want to do game development? Do you need a lot of CPU cores? Then you should absolutely spend the extra $100 to get the 9950X3D. Because you’re not just spending $100 more on a $600 CPU. What you’re really doing as a game developer is buying a complete system. $100 extra on a system that might already cost $2500? It’s a drop in the bucket for the extra 25% performance boost in heavily threaded workloads.

If you’re not doing heavily threaded workloads, conversely, the 9800X3D is still the better choice. Because it’s just as fast for gaming purposes as a 16-core model but costs less. (Both run like an 8-core chip in games and cache-friendly workloads.)

There’s literally no situation where the 12-core X3D chip is the best choice. And AMD knows this, and that’s why it didn’t sample the chips to ANY reviewers. We don’t just test games, we test all sorts of applications. And that’s the conclusion no matter what workload you run.

If you don’t need the X3D, you can make a valid argument for the vanilla 9900X. But even then, the 9950X upgrade would be better. It’s why this is the “middle child” — the older and younger siblings get all the attention, for good reasons.

Why does the 9900X3D exist? Because AMD does have 6-core CCDs, and it doesn’t want to cannibalize 9800X3D sales with a 9600X3D. Some people will think eight cores aren’t enough, but they’ll balk at the price jump to 16-cores. The 9900X3D on paper looks like a reasonable middle ground, but in reality it only caters to the uninformed.

 

[abufrejoval]

Talk about having a belly button perspective! The whole point is that the 12-core isn’t a great deal relative to the 16-core or the 8-core X3D chips. Ever. Period. There’s nothing you can do to actually make it the better choice.

Not even lowering the price? Wasn't that your point?

Do you want to do game development? Do you need a lot of CPU cores? Then you should absolutely spend the extra $100 to get the 9950X3D. Because you’re not just spending $100 more on a $600 CPU. What you’re really doing as a game developer is buying a complete system. $100 extra on a system that might already cost $2500? It’s a drop in the bucket for the extra 25% performance boost in heavily threaded workloads.

Again, from a gamer perspective, I agree, I wouldn't buy a 6/12 core part either, not even if it was €100/300 part. But that's just my navel and yours.

If you’re not doing heavily threaded workloads, conversely, the 9800X3D is still the better choice. Because it’s just as fast for gaming purposes as a 16-core model but costs less. (Both run like an 8-core chip in games and cache-friendly workloads.)

There’s literally no situation where the 12-core X3D chip is the best choice. And AMD knows this, and that’s why it didn’t sample the chips to ANY reviewers. We don’t just test games, we test all sorts of applications. And that’s the conclusion no matter what workload you run.

Again if that were universally true, there is more wrong with it than the price, contradicting your headline.

And that means you haven't really done your job investigating.

If you don’t need the X3D, you can make a valid argument for the vanilla 9900X. But even then, the 9950X upgrade would be better. It’s why this is the “middle child” — the older and younger siblings get all the attention, for good reasons.

Why does the 9900X3D exist? Because AMD does have 6-core CCDs, and it doesn’t want to cannibalize 9800X3D sales with a 9600X3D. Some people will think eight cores aren’t enough, but they’ll balk at the price jump to 16-cores. The 9900X3D on paper looks like a reasonable middle ground, but in reality it only caters to the uninformed.

I guess before we can make educated guesses as to why these chips exist (and at these prices), we need a little more info.

E.g. I'd be pretty sure that AMD bins CCDs before they decide which ones will get V-cache.

I'm also pretty sure they prefer to sell all CCDs as EPYCs, because that's greater revenue, at least initially.

The next question the is how late in the game they need to decide on packaging as desktop/HX-laptop, TR or EPYC.

There is in all likelyhood also a niche market for lower-core V-cache EPYCs, because in HPC, EDA etc. there is niche for nearly everything.

So do they put V-cache on piles of "bad-binned" 6 core CCDs to sell them at a better price? Quite possible if we knew their manufacturing overhead for adding V-cache, and when they need to decide that we'd have a better guess.

I am mostly telling you that I am more inclined to believe that you didn't do your research properly than believing that AMD simply doesn't know how to price.

Especially if now you're saying that wouldn't help anyway...

 

[thestryker]

The x900 series since Zen 3 has been the one that gets discounts faster and a lot of the time bigger and they're always in stock (the 7900X3D was actually cheaper than the 7800X3D at one point). This is something that has gotten worse over time. To me that would indicate AMD doesn't make them for the retail market despite still selling them as boxed processors because it seems foolish to sell something the market isn't particularly interested in.

While I wouldn't buy one of the x900 parts period I can understand why someone might buy the non-X3D ones. If you don't need maximum single CCD performance and do a lot of MT it makes sense to go from an 8-core to 12-core part assuming you cannot afford/justify going to 16-core. With the X3D parts one is presumably buying it for the extra cache which only comes in a single CCD. That means now you're getting a compromised experience when compared to the 8-core part. At that point the only benefit is the extra 4 cores for multithreading since the clock speed scaling is no longer dramatically different with Zen 5.

The only reason I can think of that AMD even sells X3D x900 parts is that they make more money even with the extra CCD than selling an X3D x600 part even if they sell poorly.

 

[JarredWaltonGPU]

Not even lowering the price? Wasn't that your point?

Again, from a gamer perspective, I agree, I wouldn't buy a 6/12 core part either, not even if it was €100/300 part. But that's just my navel and yours.

Again if that were universally true, there is more wrong with it than the price, contradicting your headline.

And that means you haven't really done your job investigating.

I guess before we can make educated guesses as to why these chips exist (and at these prices), we need a little more info.

E.g. I'd be pretty sure that AMD bins CCDs before they decide which ones will get V-cache.

I'm also pretty sure they prefer to sell all CCDs as EPYCs, because that's greater revenue, at least initially.

The next question the is how late in the game they need to decide on packaging as desktop/HX-laptop, TR of EPYC.

There is in all likelyhood also a niche market for lower-core V-cache EPYCs, because in HPC, EDA etc. there is niche for nearly everything.

So do they put V-cache on piles of "bad-binned" 6 core CCDs to sell them at a better price? Quite possible if we knew their manufacturing overhead for adding V-cache, and when they need to decide that.

I am mostly telling you that I am more inclined to believe that you didn't do your research properly than believing that AMD simply doesn't know how to price.

Especially if now you're saying that wouldn't help anyway...

[Rolling my eyes...]

Of COURSE changing the price changes the potential attractiveness of a part. That's a given. Please don't treat us like we're idiots. At the current prices, 9900X3D (like 7900X3D before it) makes almost zero sense.

And from AMD's side of things, I'm not sure it makes sense either. Because let's say it gets two CCDs that can't function as 8-core chips. Well, on the one hand it can do a 7900X3D for $600, with one of the chips getting (relatively expensive) X3D cache. Or it could just sell two 9600X chips and be done, increase market share, and not sell an underwhelming chip to an uninformed person.

Based on Ryzen 7 9700X vs Ryzen 7 9800X3D prices, it seems AMD feels a $150 difference for the stacked cache is justifiable. ($200 to go from 9900X to 9900X3D, or from 9950X to 9950X3D.) And yes, some of that is profit. The point being, two partially disabled chips at $279 (yes, they're selling at $229 now at some places), with no stacked cache, versus one $599 chip with cache. That's only a $40 total difference, for something AMD otherwise wants to charge $150~$200 to deliver.

The real problem is that people want higher tier parts to be universally better. Core Ultra 9 285K is faster in every way compared to Core Ultra 7 265K, and that's in turn faster in every way than Core Ultra 5 245K. But because of stacked cache, while Ryzen 9 9950X3D is generally faster or equal to both 9950X and 9800X3D, the 9900X3D is slower than both the 9950X and the 9800X3D in certain tests. It's a compromise, and still an expensive one at that.

Cut the price to below $550 so that it's less expensive than the 9950X — which would still be faster in almost all heavily threaded tasks — and now you're only $80 more than the 9800X3D. And the 9800X3D is still faster in games.

TLDR:
1) If you care about games and don't care about multi-threaded so much? 9800X3D for $480.
2) If you care about games and maximum multi-threaded performance? 9950X3D for $700.
3) If you don't really care about games but you care about multi-threaded performance? 9950X for $550.

The 9900X3D lacks a clearly defined market niche. If you only kind of care about games, enough to want more than a 9900X but not enough to want a 9800X3D or 9950X3D, but you also kind of care about multi-threaded, but not enough to go to the 9950X or 9950X3D? Great. 9900X3D might be for you. Welcome to the "I don't really know what I want or why" category of CPUs! How much should that cost? Probably $550 tops, and that lead back to the idea that AMD would be better off not making that part and only offering 9700X, 9800X3D, 9900X, 9950X, and 9950X3D.

 

[abufrejoval]

That still sounds like consumer navel gazing and saying AMD is stupid, to me.

As we've seen with Intel, tunnel vision can occasionally go on far too long and that's been true for AMD as well at various points in their journey.

But in this case I'm still thinking that understanding both the production logistics better and where and how they sell the chips would make for a better explanation than AMD just being stupid.

If I'm wrong, we'll just see these CPUs drop to really low prices, if there are in fact enough of them to even cause a tiny glut somewhere. Perhaps they are simply putting theoretical chips out there and don't ever actually manufacture them, if they don't get ordered? Unless you know if they make prices for chips they have left over or chips to match a full range of price points, how can you judge their "wisdom"?

And btw. I've seen quotes for the original Zen 3 V-cache of just $20 total extra for the entire V-cache in manufacturing, which is clearly not how they are selling them.

Of course, even if that was that cheap in manufacturing originally, that still could have been very difficult to scale back then. Today the scaling logistics should be much better managed while cost of the operational manufacturing overhead should be down as well. The fabbing cost of the SRAM shims? Who knows, but not completely decoupled from the base CCDs, either.

If I can't prod you to dig deeper, fine. You're not Anandtech, I understand that. But there is far too many people out there, including some big egos on Youtube, who criticize before they analyze and understand the bigger picture.

 

[JarredWaltonGPU]

You're not Anandtech, I understand that.

LOL. Eleven year AnandTech alumni, thanks. There are a lot of reasons that bastion of tech journalism is defunct.

Nowhere do we say "AMD is stupid." Those are your strawman words. We say, "This is a product that's going to be difficult to sell, given the competing products and prices." Every product exists for a reason. Mostly, that reason is "to make money." That doesn't mean every product, at the proffered price, is worth buying.

IF there was a great reason for this product to exist, or even a good reason, with its given specs and pricing, AMD would absolutely be sampling it. AMD itself knows that this is a product caught between the proverbial rock and hard place, and helping reviewers tell that story doesn't benefit AMD. Ergo, it doesn't sample the part.

This is the same reason why mediocre laptops, SSDs, etc. (and cars and all sort of non-tech products) don't generally get widespread review coverage.

Anyway, you win. Congrats. You're right and it's an awesome product and you should buy it.

 

[abufrejoval]

The x900 series since Zen 3 has been the one that gets discounts faster and a lot of the time bigger and they're always in stock (the 7900X3D was actually cheaper than the 7800X3D at one point). This is something that has gotten worse over time. To me that would indicate AMD doesn't make them for the retail market despite still selling them as boxed processors because it seems foolish to sell something the market isn't particularly interested in.

While I wouldn't buy one of the x900 parts period I can understand why someone might buy the non-X3D ones. If you don't need maximum single CCD performance and do a lot of MT it makes sense to go from an 8-core to 12-core part assuming you cannot afford/justify going to 16-core. With the X3D parts one is presumably buying it for the extra cache which only comes in a single CCD. That means now you're getting a compromised experience when compared to the 8-core part. At that point the only benefit is the extra 4 cores for multithreading since the clock speed scaling is no longer dramatically different with Zen 5.

The only reason I can think of that AMD even sells X3D x900 parts is that they make more money even with the extra CCD than selling an X3D x600 part even if they sell poorly.

That's the point I'm trying to raise: unless you understand the logistics of how these chips are going through the various stages of production and assembly and how flexible AMD can balance markets and inventory (which has a cost, too), it's tough to understand the distribution and pricing strategy.

Remember those GPU-less APUs that came from the top rim of the wafer? Might have been two per wafer, but eventually there were enough of them to turn then into a product. E.g. CCDs with only two working cores may truly wind up in the trash pile, but quad cores? With a full cache complement? Those might actually just fit some EDA or genomics use case...

There have been so many examples of how AMD has been trying to turn even the last bits of usable silicon into some niche chip that they still could sell somewhere somehow, that I believe it's a major science there.

Whether AMD has gone ahead and produced massive amounts of 9900X3D or is simply testing the water is hard to tell without that knowledge and judging without knowledge should just be marked as opinion or speculation.

 

[abufrejoval]

LOL. Eleven year AnandTech alumni, thanks. There are a lot of reasons that bastion of tech journalism is defunct.

Anyway, you win. Congrats. You're right and it's an awesome product and you should buy it.

No, I lost because I couldn't get you to find out: you're just pissed off I called you on not going deeper.
I know that is expensive, but unless you go just a little beyond opinion, TH might not have a USP and be next.

No I'm not trying to win an argument, I'm trying to get you fill some of the big gap your sister publication left open when it closed. There is far too many opinions out there, including my own. What I want is information.

And I might have mentioned, that I wouldn't ever buy anything but a top binned chip for myself, because I'm not buying bulk.

And I *did* agree that for most of the TH readership it's the same. I only disagreed that AMD is just stupid to offer the 9900X3D or at that price and that you should either moderate your judgment or (ideally) dig deeper.

 

[helper800]

No, I lost because I couldn't get you to find out: you're just pissed off I called you on not going deeper.
I know that is expensive, but unless you go just a little beyond opinion, TH might not have a USP and be next.

No I'm not trying to win an argument, I'm trying to get you fill some of the big gap your sister publication left open when it closed. There is far too many opinions out there, including my own. What I want is information.

And I might have mentioned, that I wouldn't ever buy anything but a top binned chip for myself, because I'm not buying bulk.

And I *did* agree that for most of the TH readership it's the same. I only disagreed that AMD is just stupid to offer the 9900X3D or at that price and that you should either moderate your judgment or (ideally) dig deeper.

You should become a freelance reviewer mining for that tantalizing data you want, if there is such a market for such information.

 

[abufrejoval]

You should become a freelance reviewer mining for that tantalizing data you want, if there is such a market for such information.

I don't monetize every facet of my life (it sure helps, that I don't have to...) and in the quest of knowledge, wisdom is its own reward (yeah, I just liked the sound of that).

To be honest, I am quite shocked that nobody else even seems to be interested to find out, if AMD needs to "bind early" or "late" for producing sellable SKUs. I remember that this facet alone in object oriented programming was as big as the shism between Catholics and Protestants.