Monday, June 10th 2019
AMD Readies Ryzen 9 3950X 16-core Processor to Awestrike Crowds at E3
When AMD launched its Ryzen 9 3900X 12-core/24-thread processor at its Computex 2019 keynote, our readers commented on the notable absence of a 16-core SKU, given that a "Matisse" multi-chip module with two 8-core "Zen 2" chiplets adds up to that core-count. Some readers noted this could be a case of AMD holding back its top performing part in the absence of competition in the segment from Intel. It turns out, the company was saving this part up for an E3 2019 unveiling.
The Ryzen 9 3950X maxes out "Matisse" MCM with 16 cores, 32 threads via SMT, a staggering 64 MB of L3 cache (72 MB including the 8 MB of total L2 cache), and a stunning 105-Watt TDP figure that's unchanged from the company's TDP for the 3900X. The Ryzen 9 3950X is clocked at 3.50 GHz, with a maximum boost frequency of 4.70 GHz. The company is yet to reveal its price, but given that the $499 price-tag has already been taken by the 3900X, one could expect an even higher price. It remains to be seen if the 3950X will launch alongside the rest of the series on 7/7.
Source:
VideoCardz
The Ryzen 9 3950X maxes out "Matisse" MCM with 16 cores, 32 threads via SMT, a staggering 64 MB of L3 cache (72 MB including the 8 MB of total L2 cache), and a stunning 105-Watt TDP figure that's unchanged from the company's TDP for the 3900X. The Ryzen 9 3950X is clocked at 3.50 GHz, with a maximum boost frequency of 4.70 GHz. The company is yet to reveal its price, but given that the $499 price-tag has already been taken by the 3900X, one could expect an even higher price. It remains to be seen if the 3950X will launch alongside the rest of the series on 7/7.
90 Comments on AMD Readies Ryzen 9 3950X 16-core Processor to Awestrike Crowds at E3
Turn the argument around: I do prefer my CPUs to be air cooled and still run close to max capacity ;) It wouldn't be good if you had to add half the price of the CPU in cooling to make it work proper, would it?
Thing is, AMD under Su has been really big on small dies to maximize profits. Even Radeon VII has a small die. Small dies are rarely the optimal path for enthusiast performance. Of course, having a large die doesn't guarantee outstanding performance either (as we saw with Fiji). If those dies are full of irrelevant quasi-relevant stuff like Tensor, then maybe what you're getting is help with hot spots more than the highest performance. (Or, if the architecture has a bottleneck...) It does seem to make some sense, though, to go to chiplets to spread heat out, since we're at 7nm now — despite the latency. However, if we're talking about best performance we would be far nearer the reticle limit. Enthusiasts should be wary about getting excited about expensive products like Radeon VII that ship with small dies and high clocks. That's not only probably not optimal for performance-per-watt, it's not optimal for pure performance. What it is optimal for is margin for the company peddling it. Small dies with high stock clocks are great for planned obsolescence.
Hyperbole isn't very useful in this subject area, either. 105 watts is very little. Even if we were to see a return to the 250 watt 9590 (it used more than the 220 it was rated for), that would hardly be all that outlandish when compared to what high-performance GPUs use. This is an industry where people used to run triple SLI. Anandtech and others included triple GTX 480 and 580 systems in their benchmarks. I am reminded of the sudden obsession with power consumption that happened with televisions once marketeers starting pushing LED backlit TVs. Suddenly, it was utterly shameful to not have the most power-sipping LED technology. Anyone buying a CCFL model was clearly wanting to drag knuckles and doom the planet. It's fascinating to see how marketing pressure can change the narrative so rapidly and effectively.
If people are really so interested in performance-per-watt, then why enable companies to slap 3 fans on a 1660 and a high voltage — in search of a clock that's well outside of the efficient limit of the chip? People buy cards like this. In fact, I don't think any 3rd-party GPUs with decent cooling don't have some amount of overclock on them out of the box. GPU manufacturers know that enthusiasts care more about performance than they care about performance-per-watt efficiency. If Fiji, Vega, and Polaris had been able to lead in raw performance then sales would have been considerably higher, despite the worse performance-per-watt (outside the rare games that are more efficient on recent GCN). I have no doubt that if someone were to put out a 500 watt GPU that leads in performance over Nvidia's top card by 25% it would sell very well, or which equalized performance at half the price.
Server CPUs typically have lots of cores with high performance-per-watt and a low TDP. Enthusiasts were sold server-centric designs by AMD before. Remember Bulldozer/Piledriver? 8 cores with weak FPU performance back in 2011/2012 was definitely not optimal for gamers. It was, though, more useful for servers. What many don't realize is that AMD sold a good number of Piledriver Opterons for supercomputers. Those computers mainly used GPUs for the heavy lifting, making the cheapness of Piledriver attractive. Yes, Intel's Xeon had better performance-per-watt than Bulldozer/Piledriver so AMD didn't do so well in the server market. (That doesn't mean the design, though, wasn't more server-oriented than it could have been.) The supercomputer market cared less about power consumption, although that's becoming more of a marketing focus lately. AMD pioneered the many small cores design philosophy with Bulldozer and what we're seeing with Ryzen is an extension of that philosophy, only without CMT.
Something like a 180 watt TDP would be perfectly reasonable for a 16 core 32 thread desktop CPU. 105 seems too low to maximize performance. It also seems like a way to enable mediocre board VRMs to get less heat from customers. This way, AMD props up the practice of selling "8 phase" boards that are really 4, and so on.
However, all this said, I like performance-per-watt a lot because it makes cooling without a high level of noise easier. Regardless, though, quiet cooling and high performance are at odds. The enthusiast point of view typically favors the latter, unless that person has deep enough pockets to offset the higher heat with better dissipation hardware.
If you think it's weird to advocate for higher performance (higher TDP) over something very conservative like 105 watts, consider the explosion of AIOs in popularity (not their penchant for literally exploding, as happened with my H50). Dual fan AIOs are pretty standard these days for enthusiasts and their CPUs. Why shouldn't they be running cheaper air coolers instead? They're running dual and triple fan AIOs to gain performance.How successful that is depends on design factors like power delivery robustness. This can be seen, for instance, with GPUs that can't get further because of weak VRM and/or not enough/large enough connectors. How easy it is also depends on various decisions. I think most anyone would rather have a CPU perform higher at stock so their overclocking effort is less work. There are also people who don't want to or who can't overclock (work computers).
If you release low TDP parts across the board then board makers can sell their fake/weak VRMs and enthusiasts will be stuck paying the extra extra premium for the few top-end boards that actually have good performance. If your CPUs, though, are more demanding out of the box, there is less room for fakery.
There is also the potential issue of the number of layers. I have read that 6 layers are really needed for PCI-e 4 and yet Gigabyte is apparently going to be selling 4 layer 570 boards. Perhaps one reason for the 105 watt TDP is to make it possible to cut down on the number of layers needed. I don't know if there is a connection.
Can we expect 7GHz in single core ?
All I'm saying is I don't see a point in either Intel or AMD releasing some crazy 300w CPU that requires water cooling out of the box. If you are an enthusiast, you buy the K series CPU (or HEDT model) and you do that yourself.