Discussion AMD Threadripper 3000 series announcement and reviews

[Kocicak]

24 cores for 1400 and 32 cores for 2000 USD?

That is expensive. It does not put much pressure on Intel. It also does not motivate people too much to change platforms from AM4 to sTRX4. I must say I am surprised.

I bet that some people at Intel are pretty relieved now.

 

[CHADBOGA]

More rumors... and conspiracy.. Jeesh, nobody likes facts ????

What are the facts?

Humans have the ability to use deduction, so why not use it when and where appropriate?

 

[jamescox]

I think AMD is causing a lot of the demand due to how well they have been knocking it out of the park with Zen 2.

Zen 2 die sizes are a lot bigger than mobile dies.

That isn’t true. Zen 2 dies are smaller than a lot of mobile die since it is just the cpu core clusters and an infinity fabric interface. Most mobile die have cpu, gpu, and a bunch of other hardware. A zen 2 cpu chiplet is only about 75 square mm. An Apple A13 is closer to 100 square mm. The IO die is huge, but that is on 14 nm.

As far as the prices are concerned, this not unexpected to me and doesn’t seem out of line. Comparisons with intel 18 core are completely bogus. The price performance curve goes up quickly at the high end. The $1000 intel 18 core part is more comparable to the 16 core 3950x at $750. The intel part is a much lower end part compared to the 24 and 32 core AMD parts. Intel doesn’t have anything comparable.

The 3950x is two high performance cpu chiplets and a very small, cheap IO die. The 3960x and 3970x use 4 high performance cpu chiplets and a much larger and more expensive IO die. If they just doubled the 3950x price (2x the chiplets and probably more than 2x the IO die size), that would be $1500. It doesn’t make any sense for them to be much cheaper than the advertised prices. While they have the same core counts to the 2970WX and 2990WX, they will perform much, much better. They could come out with a cheaper 16 core part later; I don’t know how much of a market there would be for that though. It would probably be quite expensive and wouldn’t perform much better in many cases than the 3950x. There isn’t any NUMA issues as there was with the previous WX parts.

 

[DrMrLordX]

@jamescox

What makes you think the I/O die in TR3 is going to be that much more expensive than any of the other I/O dice AMD had sold to date? GF 14nm is cheap, GF 12nm is cheap. Take your pick.

Otherwise, it is as @Markfw indicated: AMD is likely bumping up the price points for 3960x and 3970x to take advantage of market position. AMD needs to be very careful so as not to build the same animosity that people have had towards Intel for years wrt pricing/segmentation. At least for now, most features (SIMD instructions, overclocking, etc) are common to AMD's entire product lineup, so they haven't gone quite as batty as Intel did in previous years.

 

[Kenmitch]

What are the facts?

3950x requires AMD ComboPI 1.0.0.4 which is currently making it's rounds. I'd lean towards it's taking longer to whip the Agesa into shape than anticipated.

 

[coercitiv]

disregard

 

[soresu]

You pay (will pay) 750 for 16 cores at AM4, one could expect to pay max. 1500 (+ 50 or 100 extra for the different platform, pricier package etc.) for 32 cores.

Nope - even on the AM4 platform, the 3950X costs 50% more than 3900X for only 1/3 moar cores, it's not unreasonable to expect AMD to get more than a little pricier on their premium HEDT platform, especially when they have the current advantage in process tech (and therefore core count and power) behind them.

Having said that, I would be interested to see what TDP they could get by using a more reasonable clock frequency like 3 Ghz for a 3970 non-X/E model.

Papa needs a new efficiency king, and the 3970X is the same TDP as the 2990X - even with it's much increased performance for certain workloads, I'm less than impressed when I know they can field a much more efficient option for a few hundred mhz less.

 

[Hitman928]

https://www.reddit.com/r/Amd/comments/dsy4kw

The socket change also sets us up nicely for future development and scalability of the Threadripper platform, both on a near- and long-term basis.

 

[Kocicak]

The 3rd Gen Threadripper will have 88 total PCIe Gen 4 lanes with 72 usable (CPU+motherboard).

What are the "unusable" 16 lines used for???

 

[dnavas]

What are the "unusable" 16 lines used for???

So, here are two pics worth scanning:

https://images.anandtech.com/doci/15062/CPU1.jpg

https://images.anandtech.com/doci/15062/Chipset.jpg

From the CPU I count a total of 64 lanes -- 48 general, 8 downlink to the chipset, 8 more that are likely to go to NVME or SATA.
From the chipset you've got 8x uplink, 8x general, another 8x that could be more SATA or more NVME (I also assume half of those "1x4 PCIe 4.0" should have been a "4x1 PCIe 4.0" :sigh: )

So, by my count, that's 80 total, assuming you don't double-count the up/down-link. Of course, 8x of them are in use by the chipset, and (apparently?) 8x on both sides are reserved. Or the reserved are what's being shown, in which case if you double-count them there are 88. I dunno -- it isn't clear to me whether the "reserved" are or are not distinct from the 8x that are in-use. Either way, though, you've got 56 free on the CPU side and 16 free on the chipset side for 72 total -- the rest are for the chipset, one way or another.

 

[Kocicak]

OK, so we have 56 lines coming from the CPU and 16 coming from the chipset - 72 usable lanes total.

Then we have 8 lanes used for both way communication between CPU and chipset.

And then we have 8 more lanes reserved for one way communication - sending data from the CPU to the chipset, which are presently not in use. What would be the usage scenario for those?

 

[Topweasel]

Math isn't that hard. https://www.anandtech.com/show/15062/amds-2019-fall-update/4
48x PCIe 4.0 for PCIe Slots.
2x NVME PCIe 4.0 x4 from CPU = 56
8x PCIe 4.0 to Chipset = 64 Lanes.

From the chipset you have 8x PCIe for PCIe slot.
Plus the choice of 2 configurations that have 4x PCIe 4.0 NVMe choices.
So 16 usable PCIe 4.0 lanes on the chipset.

56+16= 72 PCIe Lanes.
+ 8 for the Connection between CPU and Chipset = 80 active PCIe 4.0 lanes. With 8 of them not addressable. So 72 addressable PCIe 4.0 lanes.

Edit: Didn't realize the 8x PCIe uplink coming from the chip-set to the CPU. So 88 PCIe Lanes. 16 of them Reserved for CPU to Chipset communication. That's 4x the connection of 570x (2x Guess I had been missing that the the x570 did the same thing). That explains the added lanes in the Chipset. It's going to be hard to bottleneck that connection.

 

[UsandThem]

https://www.tomshardware.com/reviews/amd-ryzen-9-3900-review-eco-mode

Ryzen 3900 (non X) CPU review.

It's too bad they are only available to OEMs.

 

[Mopetar]

https://www.tomshardware.com/reviews/amd-ryzen-9-3900-review-eco-mode

Ryzen 3900 (non X) CPU review.

It's too bad they are only available to OEMs.

Does it really matter? AMD announced their new Eco Mode a few days ago that let's you set your 3900X to run at 65W instead of 105W so you can probably expect the same clock and boost settings.

I don't know if everything being sold as a 3900 to the OEMs is capable of acting like a 3900X, but it's presumed that if that were the case, AMD wouldn't sell a 3900 for substantially less than the 3900X so it isn't as though you're missing out on some kind of massive cost savings with this chip. So the 3900 may be some lower binning or just 3900X given a different configuration that's more suited to what OEMs want to offer.

Either way you'll be able to get the same from a 3900X with Eco Mode, so you're not really missing anything.

 

[UsandThem]

Does it really matter? AMD announced their new Eco Mode a few days ago that let's you set your 3900X to run at 65W instead of 105W so you can probably expect the same clock and boost settings.

Yes, that info is in the Tom's article.

The thing that users miss out on is saving some money. Just like the 3700X vs 3800X and the 3600 vs 3600X, users have the option to save some money on CPUs that are close in performance yet use less power.

People can always buy a high end CPU (like the 9900k or 3900X) and reduce it's power consumption and performance with a few UEFI clicks. However, it still doesn't save money by doing that.

 

[DrMrLordX]

Eco Mode is weird. It doesn't do exactly what you'd expect. Also, a 3900 isn't a Threadripper.

 

[Topweasel]

Does it really matter? AMD announced their new Eco Mode a few days ago that let's you set your 3900X to run at 65W instead of 105W so you can probably expect the same clock and boost settings.

I don't know if everything being sold as a 3900 to the OEMs is capable of acting like a 3900X, but it's presumed that if that were the case, AMD wouldn't sell a 3900 for substantially less than the 3900X so it isn't as though you're missing out on some kind of massive cost savings with this chip. So the 3900 may be some lower binning or just 3900X given a different configuration that's more suited to what OEMs want to offer.

Either way you'll be able to get the same from a 3900X with Eco Mode, so you're not really missing anything.

Yeah this is every X vs. Non-X part. If AMD was selling a 3900 retail first they would need slotting. It would be more than a 3800x, so more then $400. Really it would have to be $450 or $475. So not the worlds greatest savings. The typical savings for a non-X part is $30. So I am going to go with $475. At that price difference it just becomes one of three debates, do you need something to run a lower TDP, if so get the 3900. Do you want the fastest while leaving it mostly at default? Get the 3900x. Are you willing to overclock and are willing to take the chance that you might not hit the same limit if you had spent $25? Get the 3900.

I said it before but AMD is unlikely to offer the 3900 in retail because it doesn't have a whole lot of space there and if their isn't a distinct difference between the dies used in either it doesn't make sense for AMD to offer the 3900 as long as the 3900x is basically always sold out. If the 3900x sales starts to wain a bit then maybe. But honestly the launch of the Eco-mode basically tells me that AMD doesn't plan on offering the 3900 or a 65w 3950x alternative in retail. If you have a TDP limitation you want to work with, get the best and click the single switch in the Bios. At these prices it seems nitpicky to go after such a small difference (relative) in price. Specially when the guys that will do so are the ones that are going to be trying to overclock it to the higher level or beyond.

 

[UsandThem]

Eco Mode is weird. It doesn't do exactly what you'd expect. Also, a 3900 isn't a Threadripper.

Oops.

I posted it while on mobile, so I only saw "3000 series". I thought I was posting it in the Ryzen 3000 review thread.

 

[Mopetar]

Yes, that info is in the Tom's article.

The thing that users miss out on is saving some money. Just like the 3700X vs 3800X and the 3600 vs 3600X, users have the option to save some money on CPUs that are close in performance yet use less power.

People can always buy a high end CPU (like the 9900k or 3900X) and reduce it's power consumption and performance with a few UEFI clicks. However, it still doesn't save money by doing that.

That assumes that they’re actually different parts. In the Tom’s article the test the 3900 with an overclock:

Synthetic stress tests aside, after uncorking the power limits via overclocking, the 3900's increased power consumption during the application benchmarks largely equates to the consumption of the overclocked 3900X (except for the multi-threaded AVX y-cruncher workload). As expected, that results in the tuned 3900 delivering roughly the same performance as the 3900X in all of the tests. Oddly, it even offers the same performance in y-cruncher as the 3900X, but consumes 12W less. This could simply boil down to silicon quality.

Maybe it really is two different bins but they’re so close that in reality it doesn’t matter. If that were the case then I wouldn’t expect a substantial price difference. Maybe $50 at most.

Without more data it’s plausible that these are the same chips only with a separate name for the OEM configuration with the lower TDP that makes it easier to build acceptable cooling.

Even if they were different parts, AMD might never sell a 3900 to the consumer market if only because it means that they can’t cover OEM orders as easily. However that requires knowing a lot about their production capacity and business agreements to say with any certainty so it’s conjecture on my part.

What I’m essentially saying is that I’m not sure there’s a world where you get to buy a 3900 and get the cost savings you imagine. Hell, just dropping to 65W on the 3900X might be enough to realize those supposed savings depending on usage, power costs, and how much added costs there are to cool the room the CPU is in.

 

[MrTeal]

Exactly. But their pricing is not out of line with previous TR4 chips. I paid $1800 for my 2990wx, and the new 3970x will blow its doors off for $200 more. I paid $1200 for my 2970wx. I got luck with my 2nd 2990wx at only $1100.

Sorry, probably poorly worded on my part. That's what I was getting at, the previous generation already had a premium for the more capable platform and price didn't scale linearly with performance. There's no reason any should have expected to pick up a 24 core TR for double the price of a 3900X.

 

[dnavas]

the previous generation already had a premium for the more capable platform and price didn't scale linearly with performance.

Correct. The 2700x shipped at $330 and the 2950x at $900, a premium to double of $240 (plus value of two stock coolers, which for the sake of brevity, I'm going to ignore). This is less than the current premium between the 3900x and the 3960x.

There's no reason any should have expected to pick up a 24 core TR for double the price of a 3900X.

I do wish that when people mean "I don't understand why anyone would have..." they wouldn't say "I deny you the right to believe...." Maybe ask? Of course I have a reason for believing that the 24 core TR would be closer to double 3900x than the 2950x was to the 2700x. There is an engineering similarity between the 3900x and the 3960x and the 2950x that isn't shared with the 2700x. Which is to say that the "premium" between the 2950x and the 2700x was not completely an artifact of market segmentation and other such nonsense. Now, is the "glue" between dies worth $240? Well, no, probably not. The extra packaging is probably not worth even a fifth of that. But, nevertheless, there was "value" there. That physical value doesn't exist in this generation. And yet, the premium this generation is higher.

It gets worse. The 2990WX sold for EXACTLY twice the 2950X, producing a very engineering-centric core-scaling between the top and the bottom of the TR line. The 2970WX went for $50 LESS than pure core-scaling, so, get this, last year they rounded down for you! (It's de-minimus, though, let's be honest.)

This year, even if you accept this new higher AM4 vs TRX40 premium, the pricing department isn't done with you. Pull out a calculator and do $1300/24*32 and ask yourself who used to be in charge of pricing decisions, who you think is now, and whether you want those people to remain in charge. Or you could say "oh well, they have no competition, woe is us, we have no pricing power as mere consumers" -- I didn't upgrade my CPU for seven years while Intel played those games. I have no interest in playing that game with AMD either. Your mileage may vary. And that's okay -- seriously. At least this generation the fp throughput makes a compelling argument for upgrade. I'm not sure it's going to excuse the motherboard pricing, but that's a whole nother ball of wax....

 

[Topweasel]

The general rule of thumb is that AMD was always the value option. That value option that what ever compute power you got from AMD you got cheaper then the comparable Intel solution. As far as I can tell this still exists. But what we are seeing is AMD stretching its legs to allow for higher Margin. AMD has earned the right to be able to sell their CPU's what they are worth, realistically I don't see what the issue is with a 24c monster, super high performance CPU that will put others to shame, at 1400 is. It's inline with with the compute performance per dollar of the 10980xe after it's 50% price cut. I get it's no longer and only because of those Intel price cuts the Sleeper CPU you got at steal where you can giggle inside because the Intel guys paid so much to get so much less. It's only real damning statistic is a mediocre increase in price to follow up what is going to be a major increase in performance. Considering the demand they are seeing on the consumer and server side (both having priority) and the material costs (all that 7nm and a huge 14nm IO die) to go with the increase in performance, it really isn't a major shift for AMD other then they aren't purposely driving down prices to drive up value against Intel, which again without the unprecedented price cut from Intel, these chips would still come off as a major bargain.

 

[dnavas]

I agree with nearly everything you say there. I diverge here:

... you got at steal where you can giggle inside because the Intel guys paid so much to get so much less.

To be clear, I was never giggling at the Intel guys (aka: people who bought Intel CPUs), I was swearing at Intel. I wasn't giddy at the production of good hardware at value pricing, I was just suddenly not ridiculously ticked off at a company that was busy guzzling greenbacks (well, I had another choice, anyway). That difference is probably illustrative of my basic thinking :> . The details, though, wind through market theory and consumer/producer balance of power, which is well off-topic.

I look forward to seeing what AMD does next round.

 

[jamescox]

@jamescox

What makes you think the I/O die in TR3 is going to be that much more expensive than any of the other I/O dice AMD had sold to date? GF 14nm is cheap, GF 12nm is cheap. Take your pick.

Otherwise, it is as @Markfw indicated: AMD is likely bumping up the price points for 3960x and 3970x to take advantage of market position. AMD needs to be very careful so as not to build the same animosity that people have had towards Intel for years wrt pricing/segmentation. At least for now, most features (SIMD instructions, overclocking, etc) are common to AMD's entire product lineup, so they haven't gone quite as batty as Intel did in previous years.

I don’t know why you think it would be that cheap. It is cheaper than 7 nm but there is still demand for 14 nm. A lot of parts that are not bleeding edge will be built on 14/12 nm for years to come. It is a huge amount of silicon with the die size. It is close in size to high end 14/12 nm gpus, which people are still paying a lot of money for. If they use the Epyc IO die in ThreadRipper, then that is ~4x the silicon of the AM4 IO die. You would definitely be paying a premium for that level of performance, but I am fine with AMD making some better margins. This is pretty far from Intel charging huge prices for 4 core chips of the ridiculous prices they charge for HEDT parts. The intel cut their HEDT prices by ~50% and they still aren’t going to be competitive with AMD parts. It is nowhere near as bad as it used to be. When the i7 processors first came out, it went from around $500 to $1000 just for a tiny increase in clock speed. If intel actually had competing parts with AMD 24 and 32 core ThreadRipper, then we would probably be praising AMD for how low their prices are compared to intel. How much is Intel’s 28 core that would be outperformed by the new ThreadRipper processors?

 

[DrMrLordX]

It is cheaper than 7 nm but there is still demand for 14 nm.

Really? AMD has essentially abandoned all their 14nm/12nm GPU production. I think there's a 12nm run of RX 580s out there, but that's about it. About the only thing they're seriously ordering from GF in quantity now are I/O dice and some chipset dice. Picasso may still be in production, but keep in mind that AMD has got to start clearing the channel of those to make room for Renoir. It would not surprise me if the spigot on Picasso shuts off soon if it hasn't happened already. Does GF have some other major clients out there gobbling up wafers? I'm not personally aware of any.

It is a huge amount of silicon with the die size. It is close in size to high end 14/12 nm gpus, which people are still paying a lot of money for.

Who is still paying a lot of money for Vega64 dice? Besides, according to at least one (now dated) estimate:

14nm wafer cost

14nm wafer cost

webcache.googleusercontent.com

(sorry, it's a cache; the primary site is not responding right now)

they estimate the price of a 14nm GF wafer at $4000. I'm thinking that price is too high given the drop in demand, but whatever. We'll go with $4000. Anyway, with a die size of ~416mm squared, I'm thinking you can get . . . what, ~129 dice from a 300mm diameter wafer? That's about $31 per die.

That's expensive? And that's Rome. TR3 may not use the same I/O die.

 

[lobz]

Really?

I think he meant general demand for 14nm wafers in the foundry industry. The majority of TSMC's 2019 revenue comes from >7nm wafers.