AMD has had a great 2019, but the year isn’t over yet. Team Red will deliver a pair of upgrades for the desktop CPU market in a few weeks with the launch of its 16-core Ryzen 9 3950X (first teased last summer) and the upcoming 32-core Threadripper 3970X and 24-core Threadripper 3960X. There’s also a brand-new platform for the new Threadripper CPUs, and a low-cost Athlon part for $50.
Let’s round it all up.
Ryzen Hits 16 Cores
First up, the Ryzen 9 3950X, arriving for review on November 14 with retail availability on November 25. This CPU was advertised in June for fall availability, only to be delayed as AMD struggled to keep Ryzen on store shelves. The new chip will price at $750, a significant increase over AMD’s 12-core family. It’s clear that 12-core is meant to be the sweet spot at $500 and below. The jump from 12 to 16 cores is a 1.5x price increase for a 1.33x core count improvement. This isn’t bad by the standards of top-end parts, but it’s definitely a halo part. AMD may honestly have positioned the chip like this to keep demand tamped down. Both the Ryzen 9 3900X and the Ryzen 9 3950X use two 7nm die, but the 3950X requires that both dies be fully functional. With the 3900X, AMD can use die recovery to fuse off bad CPU cores and still sell the processor.
AMD’s proposed price stack and product mix for the future. We should note that this chart is only valid until Cascade Lake hits the market, which is when Intel’s Core i9-10980XE will drop to a $999 price point and offer better price/performance ratios against the Ryzen 9 3950X than the current Core i9-9920X. AMD is claiming some decisive wins against the competition.
The Ryzen 9 3950X may only have a 105W TDP, but AMD is recommending heavy-hitting cooler. The CPU will not ship with one, but customers are recommended to use a 280mm AIO liquid cooler, if not something even larger. The company is also introducing a new “Eco Mode,” which will allow CPUs to operate as if they were one TDP bracket lower than otherwise.
Athlon 3000G: AMD’s New Low-Cost APU
AMD is also announcing a new APU today, at the low cost of $49. The new 2C/4T CPU is still based on 12nm Zen+ silicon (no luck for those hoping this was a stealth Zen 2 chip). The new CPU picks up +300MHz compared to the Athlon 2000GE with the GPU grabbing another 100MHz. Both gains should be significant — a 2C/4T core can use the boost, and the modest Vega 3 GPU is still going to scale from clock speed.
AMD claims the 3000GE is a very solid comparison against the Pentium G5400 at $73, with substantial performance improvements in multiple applications. As a low-end APU it should be a solid part for customers with very light computing needs. The chip is also overclocked for anyone who wants to try and pick up a bit more performance on the cheap.
Ryzen Threadripper 3970X: Killer Performance, New Platform
Finally, there’s Threadripper. The Threadripper 3960X is a 24C/48T part, with a 3.8GHz base clock, 4.5GHz boost clock, 140MB of onboard cache, and a $1400 price tag. The 3970X is a 32C/64T part with a 3.7GHz base clock, 4.5GHz boost, 144MB of cache, and a $2000 price tag. The claims AMD is making for performance are substantial, to say the least.
Last year, AMD launched the 32-core Threadripper 2990WX, but the chip’s Windows 10 performance was throttled by deficiencies baked into the Windows 10 scheduler and how it assigned workloads to the core. The architectural and structural changes made in the shift from TR2 to TR3 should obviate these problems and allow the CPU to hit its full potential, performance-wise. AMD’s $1400 price point on the 24-core chip is a 1.5x increase in cores for a 2x increase in price, while the 32-core CPU is a 2x increase in cores for a 2.67x increase in price. Intel’s Cascade Lake being priced at $999 makes it a well-positioned upgrade relative to the 9900KS (roughly a 2x increase in cores for a 2x increase in price), but this is highly unusual for Intel. The Core i9-9980XE, for example, is still a $1900 – $2000 CPU — 2.25x more cores than the 9900K, for 4x more money.
The 7nm chips inside Threadripper use symmetrical configurations without dummy die in either a 6+6+6+6 or 8+8+8+8 configuration. The unified I/O die and memory controller design avoid the NUMA issues that plagued the 2990WX and should neutralize the core contention problems that limited performance on that CPU under Windows 10 (the 2990WX was a much better chip under Linux than in Windows, even with Core Prio).
AMD is moving to position Threadripper in a leadership position to maximize its CPU revenue and drive greater adoption of its chips in high-end and boutique workstations that use these sorts of chips. At the same time, it’s brought a 16-core to desktop to extend the core count of its mainstream CPUs. If you bought a first-generation Ryzen in 2017 on an X370 motherboard, you likely now have the option to step up to as many as 16 CPU cores and a significant uplift in CPU IPC after less than three years.
All of that performance requires a new platform, and AMD’s TRX40 is here to foot the bill. TRX40 supports up to 256GB of RAM using 32GB DIMMs and it significantly expands overall platform bandwidth. There are 64 lanes of PCIe 4.0 bandwidth in total, divided up as follows: 48 lanes for general chipset use (GPUs, accelerators), 8 lanes for chipset downlink (fixed), and then a pair of options for the eight lanes remaining.
Think of those last eight ports as being divided into two quads of four ports each. Motherboard vendors can wire up a PCIe 4.0 x4 slot, an NVMe PCIe 4.0 x4 slot (for SSDs), or attach four SATA ports. This “Pick One” option is available for both quads, meaning that a board vendor could choose to have four additional SATA ports attached to one set of four PCIe 4.0 lanes while using the other quad for an NVMe x4 slot. In this case, those extra SATA ports would be hanging off the CPU directly rather than through the chipset.
One of the reasons why AMD has a new socket on TRX40 is because it wanted to increase platform chipset bandwidth. Up until now, the northbridge and southbridge have been linked by a PCIe x4 3.0 link. With TRX40, AMD is adopting a PCIe 4.0 x8 link, effectively quadrupling bandwidth. This is the equivalent of a full-sized x16 GPU slot being used for communication between the CPU and its chipset.
TRX40 is not forwards or backwards compatible with AMD’s previous TR4 socket or motherboards. While the CPUs look physically identical and the sockets are the same, a 1000-series or 2000-series Threadripper will not accept a TRX40 CPU or vice-versa. AMD’s TRX40 is designed to be flexible, with a great deal of potential customizability. There was no mention of a “TRX80” as has been rumored.
Conclusion
Overall, this is a momentous set of launches for AMD. The lack of upgrade compatibility for first and second-generation Threadripper owners is an unfortunate loss, but the 32-core Ryzen 9 3970X looks as though it could rewrite the rules of high-end desktop performance. The 16-core Ryzen 9 3950X at $750 is priced at a premium above AMD’s other cores, but AMD was always likely to establish a high-end market for its chips. Intel has done no different for decades, and AMD has played a critical role in making desktop computing cheaper over the past 2.5 years. At the same time, the company wants to establish itself in the kind of markets where OEMs don’t even talk to you unless you’ve got a multi-thousand dollar CPU to field. That means establishing Threadripper as a player in higher-end markets.
There were consistent rumors of a 64-core Threadripper emerging this generation, but no such chip has tipped up and AMD hasn’t said anything about confirming one. It’s not surprising that AMD would want to keep its halo CPU core counts for the server market. Intel has higher core count Xeons coming next year, with 38-core Ice Lake server parts and 48-core Cooper Lake socketed chips coming as well (built on 10nm and 14nm, respectively). With socketed 48-core chips coming, AMD may be waiting to see how Intel introduces these chips and where it positions them before making any moves of its own. Alternately, it may simply plan to keep Threadripper as a 32-core platform for now, or to launch a 64-core chip at a later date.
AMD's 16-core Ryzen 9 3950X, 32-core Threadripper 3970X Available November 25 - ExtremeTech
Wanted to add this thought too, now that reviews are starting to be released. Many of them say words to the effect that the lines between consumer and HEDT are being blurred, but that's not really the case. Yes the core counts may be getting up there, but Threadripper, as compared to Ryzen, has ECC memory capability (unofficially so does Ryzen, but it is motherboard dependent), quad channel memory controllers, and -many- more PCIe lanes and interconnect bandwidth.
And to throw this in, while the 3950X has a listed peak boost speed single core of 4.7ghz, AnandTech found that the 3950X peaks at 4450mhz single core turbo over a period of time, whereas TomsHardware DID show peaks of 4700mhz but only for exceedingly short periods of time even with a custom open loop liquid cooler with dual 360mm radiators keeping the temperatures well below the maximum.
Wanted to just tack on one more thing from the Techspot review. If you get a 3950X, don't overclock it, and think twice before enabling PBO, The gains are minimal, but power consumption goes through the roof, talking about a 40% increase in power for a 7% increase in Blender performance on a manual overclock.
So, this is what I posted over at OCN (https://www.overclock.net/forum/10-...ipper-big-fail-missed-opportunity-amd-17.html
This isn't quite true. AMD originally valued the PCIe and quad-channel memory at a $150 price premium. See 1900X at $650 and 1800X at $500.
Now, AMD, in part, believed the marketing department that a mainstream CPU with the same core count would destroy sales of the HEDT part with the same core count. But this is a mistake.
When first released, you had the 1700, 1700X, 1800X, and 1900X all with 8-cores. That is pretty crowded and this is when their die binning was not great, so there was no great distinction on frequency between the 1700 and the 1900X. The 1900X had a huge flaw in its design, namely that the memory channels were on a second die and that could NOT compete against the performance of the mainstream chips, especially with the 8-core 1700 being overclockable to about the same performance of the 1700X, but priced at $270, give or take, compared to $330 for the 1700X and $500 for the flagship 1800X on AM4.
Intel saw the same thing. They tried to put a quad core on the X299 platform (Boy did that go wrong). Intel's strategy was to step gamers and consumers onto the more profitable HEDT platform through a slow transition. That was the reason for the quad core Kaby. But AMD's release crushed those plans, along with MB MFRS saying enough with this lane segmentation and having to engineer around BS gimmicks.
Fast forward to today. How is this situation different from the two stated above. First, there will only be 1 16-core on the AM4 platform, not 3 chips. To build on that, the cheapest of those chips was less than half the cost of the HEDT chip, and the second cheapest was about half the price, both of which performed better for gaming regarding memory. Second, everyone knows 16-cores is for MORE than gaming. It is at pro-sumer levels of performance and core count. As such, it isn't going to be the go to for gaming machines, rather for cheap render stations, youtubers, etc., that are not where stepping into the HEDT market would make a huge difference, necessarily. Meanwhile, having an HEDT offering fills a gap left in their product lineup (explained in a moment). Those that need the lanes or memory bandwidth (which should always be examined on a per core or per thread basis along with looking at whether your software will bottleneck below a per core or thread memory bandwidth) would get the HEDT, whereas those that do not would get the mainstream offering.
With that spelled out, let's then examine the benchmarks. The data for the 3950X puts it at or slightly ahead of the 9980XE. That is fine. But if you are choked on memory or need the PCIe lanes, the AM4 platform is a non-starter. Period. What that really is an advertisement for, due to pricing, is that if 1950X or 2950X owners, whose motherboards were roughly copy-pasted from X299 designs and did not fully take advantage of leveraging the PCIe lanes of the TR products in some instances, need an upgrade and expected to upgrade this year want something that fits that budget point, the 10980XE will now be their upgrade path (since drop in is no longer on the table). That is, of course, if you are able to overlook the security issues with the platform. The 18-core, due to speed and IPC, would be roughly in line with a 16-core threadripper on HEDT. A supported decent board for the 18-core for around $300 puts that platform at around $1300.
Now, let's look at the alternatives and price points. A 12-core on AM4 is MSRP of $500 and the 3950X 16-core is $750. The X570 MB is going to be $250-400 for a decent one. So, for switching to the AM4 platform, you have around a $1000 pricing for the 3950X that could go higher depending on features needed (up to $600, so $1350 unless REALLY burning money for the Aqua, etc.). The Intel highest end main HEDT lineup, excluding the overclockable Xeon, runs $1000 and can be had with motherboard around $1300-1450. And, if you need those lanes and memory bandwidth per core/thread for your specific programs, that is now the recommendation at that price point, especially if your software does not support nor scale past 22 threads (software coding in some instances is just not necessarily supporting the crazy (in a good way) increase in cores and threads). With the 24-core, you start at $1400, about $100 more than the ORIGINAL pricing of the 24-core 2970WX, whereas the 2990WX increased to $2000 from either $1700 or $1800 launch price. I don't have a problem with that as Epyc benchmarks showed the centralized I/O die corrected the NUMA issues in M$ software/OSes (which were due, in part, to legacy support of Intel's glued processors from long before and M$ not properly having a good scheduler to deal with these chips, stale data issues related to multiple hops to get to memory, etc.). But, the change in forcing a new socket, something not needed for the extra lanes on Epyc platforms and the memory argument for AM4 which had to be routed in the PCB, along with them blaming the PCIe lanes to the chipset when it was already shown AMD considered doing 8 lanes on AM4 but would not have needed a new socket and the reasons given for not doing so were heat and power draw, not a new socket, suggests the arguments on the necessity of the change are lies. Everyone can figure out the VRM issues are central, as power delivery for 24+ core chips required active cooling on the ZE, etc. But I digress.
In any case, that means for an upgrade, you will have to buy a new platform, and with the cheapest Gigabyte board being in the $400 range, while the ones with the features people really may want or need are in the $600 range, you have a buy-in of $1800-$2000, rather than a drop in of $1400. Moreover, AMD pushing buying the 2970WX or 2990WX for existing Zen/+ HEDT owners is almost an insult at this point, as it has been shown, depending on your workloads, the 18-core Intel chips may outperform those handily.
So, in mainstream, AMD wins the recommendation except for OCing or gaming, where if you are going max FPS and it is solely a gaming machine, Intel is still the top. Still, many will be consuming AMD chips for the $200 price point to $400 price point. Then comes the AM4 productivity. If you do NOT need the memory per core/thread nor the PCIe lanes, then the $750 ($1000-1150 platform cost with CPU) 3950X is a good chip. But if you are like me, you need around 44-PCIe lanes, give or take. That would bring you into Intel's offerings (which I recommend AGAINST buying any Intel 10-core or 12-core HEDT chips, they are not worth it, grab the 9900K if you don't need the PCIe or memory bandwidth per core/thread as it will be a better performer, generally, or get the AMD 3900X or 3950X), which is the 14-core for $780 and the 18-core for $1000 (or around $1100-1400). This is ignoring power budget, for the moment. Intel had the special auctioning of high binned 14-core chips awhile back. Those were shown to even beat the 18-core at specific tasks and workloads. As such, they should not be ignored. Comparing the 3950X to the 7960X, though, the 7960X got spanked handily in workloads. As such, it comes down to platform features. So, if you need the HEDT platform offerings, although fewer lanes than AMD's, Intel's platform fills the gap in-between the AMD AM4 offerings and AMD's Threadripper 3000 offerings, while having IPS (instructions per second, calculated as IPC*Frequency) roughly on par with the Zen 2 chips, even if slightly behind, but still being a large enough uplift from 1950X and 2950X chips that it should be considered, even if older in design.
Finally, anything above that level brings you to AMD's newest 3000 series TR lines. The 32-core will neuter Intel's OCable Xeon with 28-cores coming in at $3000 and with a MB costing $1000-2000, thereby meaning a platform cost of $4000+. AMD effectively killed this chip, and that is without discussing the 48-core and 64-core variants, who even though lacking AVX-512 have been shown on Epyc that the high core count and AVX-256 support overcomes to give equivalent performance with Intel's chips with roughly half the core count using AVX-512.
This is the pricing lineup and how my recommendations have shaped to date.
Now to address your statement Intel cannot compete with AMD TR. That is false. By cutting the prices in half, I have shown Intel now has a command of the HEDT space below the platform cost of $1400. AMD has NO COMPETING PRODUCTS in this segment. As mentioned on software utilization, some programs cannot fully use all those threads (not a bad problem, but can be a factor on if the higher core count chips are needed in that specific deployment). You also have the PCIe lanes or memory bandwidth per core/thread being a potential issue on how some software is coded, which would rule out AM4 from consideration. That means Intel has taken the budget friendly HEDT space WITHOUT any competition from AMD whatever. AMD has taken the crown atop the best performing HEDT products, toppling Intel in fantastic fashion. But what built Zen and Zen+ adoption WAS THE VALUE PROPOSITION. You could grab a 16-core AMD product on HEDT with more lanes over grabbing a 10-core Intel product for the same price. Guess what made sense for a lot of users? Now, Intel has the value proposition. It doesn't matter that it is on an older process, they were far enough ahead, and benchmarks show, that AMD's 16-core HEDT product would have only been a bit better than the 18-core offering while both the 1950X and 2950X products are much lower on performance than the 18-core. That means Intel is now not only the value proposition, but the logical upgrade path for 1950X and 2950X owners who cannot wait for the 2021 platform.
I think many people are forgetting to look at the actual product segmentation and how consumers will evaluate the product options. Because of that, they are only looking at "what is the biggest and best in that category." Just like the 2080 Ti is not the top selling regarding volume of cards, so to will the AMD top dog products not be moving the most volume in the segment. With TR and TR2, AMD flipped the script among DIYers so that they were outselling Intel 2:1 based on the value proposition. They just lost the value proposition. Do you think they are going to maintain their volume in the HEDT segment then? Probably not. Even Intel's purchasers of the $2000 18-core chips could not keep Intel in the majority on a volume basis for the segment. So thinking AMD can do it is a bit absurd.
And, just to stress the point, had AMD had a single 16-core HEDT chip, priced around $900, with the $150 price premium for quad channel and PCIe lanes, even with potentially higher cost MBs than Intel would require, they would get rid of any consideration of the 14-core or 18-core Intel CPUs, thereby having squeezed Intel from top to bottom out on value. That single lack of a product, even with changing the socket, literally is what will likely be considered AMD's folly and a big missed opportunity.
Finally, having more than two years support on the TRX40 would be a mistake. The current TR4 socket is a variant of the SP3 socket for Epyc. The leaked roadmap from an AMD presentation showed the need for SP5 in 2021 for Epyc CPUs, which will be for DDR5 and PCIe 5.0. At that point, changing the socket would make sense as that is when large platform changes will be made. This change in socket this generation seems more about insufficient power delivery systems on X399 boards and not wanting consumers blowing up their hardware (see Nanny-state). But, considering the changes needed on the server platform to handle the new changes in 2021, why should we assume the TRX40 is sufficient?
Instead, if AMD leveled with consumers and said that the HEDT platform will be adopting a similar cadence to Intel in supporting only two generations per socket, I think consumers would be happier as they can plan for lifespan of their platform. Simply put, tell consumers that the HEDT platform is their new playground for cutting edge features, so the limited length is for specifically adding value to the product stack while not being held back by socket compatibility. Customers will accept that rational. It also is no different than Intel's normal cadence, so the market is primed without expectations, which is why this launch will likely not be as good of one for AMD on TR. Personally, I would not recommend buying AMD TR in 2021 with Zen 4 unless a socket change accompanies the chips, at least from my current perspective.
The real reason some consumers are mad about the non-compatibility with TR4 X399 boards was they were given little warning. AMD leaked out non-compatibility starting around August, maybe a bit earlier. This was meant to prime the market for needing a new board. That was not enough time and was speculation, while having other rumors cut against it. As such, some consumers were blindsided by the decision. That creates negative goodwill moving forward. In other words, consumers are not so mad about price shock as they are about consumer expectations not being met. With how AMD used Ryzen support until 2020 in February of 2017, but used AM4 until 2020 starting by at least July of 2017, consumer expectations could have been formed that were not managed by the change of phrasing by the time they began purchasing the X399 platform.
But putting that aside, the problem, nonetheless is consumer expectations and the improper management of those expectations by AMD. Instead, this close to Christmas shopping, they spring an extra minimum cost of $400 (averages are closer to $600 for a new motherboard) when customers already had their budgets set for the new model of chips. This, I feel, is why some have complained heavily about the change that has come. That extra $400 at launch could equate to numerous other gifts.
I hope this analysis finds readers well regarding this topic.Here is answering other comments that arose after my posting in the other forum.
I am sorry but 20 PCIe4 lines = 40 PCIe3 lines. Intel drops 2066 to compete with AM4 for this reason too. So oldie 2066 is not a HEDT anymore like former HEDTs 2011 and 1366.
Again Intel has nothing against the new TRs excluding W-3xxx for the same or higher prices
So what? The problem is I'd have to buy daughter boards with mux chips to split out the PCIe4 lanes to PCIe 3 lanes to actually use my add-in cards. THAT IS UNACCEPTABLE and raises the cost to equal or above levels of just buying an Intel HEDT platform.
You CANNOT count lane bandwidth like that. You MUST look at how it can be used. And since most cards are not yet optimized for PCIe 4.0, you do not get nearly the benefit that is possible there.
That is the problem with that argument. If you CANNOT use it with your cards, or have to use convoluted processes to make it usable, why bother when you have a product in that price range which can satisfy your need WITHOUT the contortions to use it?
So you are wrong and that argument is absurd ab initio. It is a distraction to make those that haven't optimized workstation HEDT platforms or those not fully utilizing their lane count to think maybe the mainstream platform is okay. But it isn't. You get bottlenecked with the chipset lanes, granted those being equivalent to a PCIe 3.0x8 lanes, while each card in a slot takes up physical space while reducing how to get what you need on the lanes you want. So you can split x16 to x8/x8, but without a daughter board with mux chip, you are wasting half the bandwidth on each slot. Quickly, when you start planning it out, the AM4 platform IS insufficient. PERIOD.
Yes I canI can use 3 video cards, 3 nvme, bunches of ssd and 128 memory on AM4. Top 570x mobos gave me that. It’s not a HEDT but much better than Intel offers on mainstream.
And for the cost of the 3950X, you can just buy the 18-core and not have to fuss.
Let's break down how that would work, 3 graphics cards and NVMe. Those three cards are PCIe 3.0 unless being 5700/XT. That means that you would be stuck with an 8/4/4 situation. But, those are physical lanes. Because you have PCIe 3.0 products filling them, they are running at PCIe 3.0 speeds as 8/4/4, meaning there goes half the bandwidth for those lanes wasted AND if the GPUs are high-end enough, two of the cards are being choked by the x4 lanes, meaning you already are suffering on performance.
Then, with the NVMe, you have 1 with direct connection to the CPU. That is fine, but might be wasting bandwidth if PCIe 3.0 in the slot. The other two run through the chipset. That is equivalent to PCIe 3.0 x8, which is fine, but also is competing with ALL OTHER PRODUCTS hanging off the chipset, which may include your SSDs, anything plugged into sata potentially, possibly networking, etc. I'd have to see the MB schematic to give what it is sharing with.
Do you not see the problem?
