AMD Launches Threadripper 3970X, 3960X and Smokes Intel's New 18-Core CPU (hothardware.com) 44
MojoKid writes: Intel and AMD have been duking it out in the high-end desktop processor space lately. AMD's return to competitive footing versus Intel has propelled the company forward and the brand has a loyal, passionate following due the competitive performance-per-dollar its 3rd Generation Ryzen processors bring versus Intel offerings. Today, both companies have launched new flagship many-core CPUs, the Intel Core i9-10980XE, which is an 18-core chip, and the AMD 3rd Gen Threadripper 3970X and 3960X, which are 32-core and 24-core chips, respectively. Intel's Core i9-10980XE brings a lower price of $999 and competes more favorably versus AMD's lower-end 16-core Ryzen 9 3950X that's priced at just $750. Meanwhile, the new AMD Threadripper 3960X at $1399 and Threadripper 3970X at $1999 leave Intel's fastest desktop chip in the dust in multi-threaded workloads, sometimes by a wide margin. In addition, while Threadripper 3960X and 3970X pull only about 26 to 36 Watts of additional power versus Intel's new Core i9-10980XE, they do it with 33-77% more core resources. Regardless, it's impressive how the tables have turned, as AMD is now firmly entrenched with some better value propositions in high-end desktop processors, and better performance in many cases as well.
chip to chip (Score:4, Interesting)
finally AMD have figured out how to do on chip interconnects and are actually selling them at a decent price
I hope AMD are wildly profitable
I hope intel bundles FPGA on all their server chips soon
Re:chip to chip (Score:5, Informative)
I hope AMD are wildly profitable
From their latest quarterly earnings:
Our first full quarter of 7nm Ryzen, Radeon and EPYC processor sales drove our highest quarterly revenue since 2005, our highest quarterly gross margin since 2012 and a significant increase in net income year-over-year, (...) For the fourth quarter of 2019, AMD expects revenue to be approximately $2.1 billion, plus or minus $50 million, an increase of approximately 48 percent year-over-year and approximately 17 percent sequentially
AMD is in full comeback mode, not only because Zen is actually good but because Intel is still fumbling with their 10nm process. And with AMD in both the PS5 and Xbox Scarlet they'll probably see some nice semi-custom boosts in the near future too. I think they got very lucky with both the crypto mining craze and Intel running out of steam but now their future looks pretty bright, though they really could use a new "Zen for GPUs" architecture too. Despite that they mention Radeon I think Navi is clearly the underdog in that market. But hey, as long as the CPU side starts making money again there's room for investments. Also they need it to compete better with Intel in the laptop market anyway, they perform better than Intel but also at higher power.
Re: (Score:1)
Navi (RDNA) is looking pretty swell too, I presume you know about W5700. It's Navi. Can't speak for you, but iWant.
Re: (Score:1)
Navi is clearly the underdog in that market
By which silly metric?
RDNA based Navi beats competition two+ size the die size and is competitively priced.
Re: (Score:2)
I imagine the only reason AMD didn't do this with Zen 1 was lack of money to develop the separate IO die.
Re: (Score:2)
I don't know about that. The I/O die does add latency, but uniformly across every core. I think the main reason is the windows scheduler sucked balls when it came to dealing with non-uniform access and core layout. But on the other hand the I/O die should save money as you can to it with an older process and lets you use the limited fab capacity to make as many chip-lets as you possibly can. It may also make it easier to integrate the GPU on package for their APU offering.
Re: chip to chip (Score:2)
On old chipsets from around P2 P3 etc... memory was connected to the chipset even for SMP machines and this is much the same.
Re: (Score:2)
If you're talking P2 ere terminology, the I/O die is the north-bridge. And you're defiantly adding latency when you have to go over DMI (P2) or the infinity fabric, vs doing it with a ring bus and controller on die. Granted doing it on package isn't much latency, but it is some amount, and shows up on certain benchmarks (mainly those that hit data sets much larger than cache or that require a lot of round trips with a co-processor. And the Thread-ripper 1000/2000 series when you were accessing memory local
Design and skill (Score:1)
Re: (Score:2)
Re: (Score:1)
A 3rd and any other brand is also free to enter the market.
See how their experts do...with a new CPU design.
If Intel drops the price, makes a faster CPU at a lower price again... the market might move back.
Re: (Score:2)
A 3rd and any other brand is also free to enter the market.
Not really. Intel and AMD share a patent cross-license agreement that makes them an effective duopoly for X86 CPUs. (And other CPU architectures simply aren't viable in the mainstream desktop market.)
Re: (Score:2)
Someone has to make more cores faster again.
Design a new CPU that is next gen fast. Not just add more average cores again and again.
Something to "CPU" way beyond 8K ready GPU designs.
Re: (Score:3)
If I were you, I'd get on the task of smashing those physics barriers to higher core speeds right away. Your Nobel Prize is just waiting to be collected.
Re: (Score:1)
If more adding more cores cant keep up with the needs of displays and games?
Stop larger display design?
Stop game art complexity?
The next gen CPU has to keep up.
Until then its more cores
"Intel designers aging..."? (Score:2)
A previous comment:
Intel has insufficient management, in my opinion. [slashdot.org]
Comment removed (Score:5, Interesting)
Re: (Score:2)
But what Intel did NOT see coming was finding out their entire design at a fundamental level would be a security nightmare
I was with you till you got to this point. Speculative execution has had zero to do with Intel's failings and AMD's success and hasn't factored into the purchasing decision of anyone other than a handful of Slashdot posters and cloud computing providers (and even then I'm not sure that it's pushed them). Also you got the timeline wrong. Zen came out before the security issues did. Hell when Meltdown and Spectre hit AMD was already announcing the imminent and highly anticipated release of Zen+ on the back of
Re: (Score:2)
Re: (Score:2)
Ice lake looks to be a widening of the same sort of core. 5 Decoders rather than 4, 2 more execution pipelines, and a 50% increase of reservation stations. But this only helps if you can find the ILP, and if your OoO is secure.
We don't have a lot of details on Zen3, but the rumor is that it's a new u-arch that will likely focus in improvements the the cache hierachy, power, and on the I/O die.
No rumors on how exactly they'll change the pipeline if at all (maybe AVX 512 and a few more reservation stations.
I doubt it (Score:1)
Threadripper 3990X 64 Core / 128 Thread confirmed! (Score:5, Informative)
AMD is having a 1-2-3 knock out punch!
* They just confirmed a 64-core / 128-thread Threadripper 3990X [twitter.com] is coming in 2020! This will be REAL interesting to see what price point it comes with -- especially compared to EPYC.
* The TR 3970X is kicking ass and taking names! However there IS one place where Intel the holds the lead: Computational Fluid Dynamics -- due to it having high memory usage. STARS Euler3d [hothardware.com] shows the i9-10980XE and i9-9980XE (slightly) beating the Threadrippers 3960X and 3970X.
However, I think this bears some more investigation as TechReport reported back in their TR 1920X review [techreport.com] that:
Is the Intel Fortran compiler using AVX-512 [wikipedia.org] ?
Also does anyone have Stockfish benchmarks, and how compilation scales with multi-core for the TR 3960X and TR 3970X?
Re: (Score:2)
Similar tests performed 8 years ago on an Opteron system also showed a bias towards Intel. The article explains it as being the result of Intel having a lower latency memory connection. There were also other, more significant, reasons for the 2011 test results but they are probably not applicable to the 2019 hardware. It is posted here [anandtech.com].
Fluid dynamics involves numerous interdependent tasks. It does a great job of stressing the memory system. CPU caches do not help as much as in other applications. H
Re:Threadripper 3990X 64 Core / 128 Thread confirm (Score:4, Interesting)
Found a review [anandtech.com] that details memory latency for an EPYC 7742 and Xeon 8280. The cache latency is much lower for the EPYC. However, memory latency is lower for Intel (~115ns vs ~89ns). So the difference between the Intel and AMD CPUs in the Euler3D benchmark can probably be attributed to a difference in memory latency.
Re: (Score:2)
The notion that a foot is a light-nanosecond has led me to some interesting visualizations of how high speed circuits operate. Instead of thinking of it as 89 ns or 115 ns away, imagine that it's 89 feet or 115 feet away, and you have to suck the data down a tube that long. I guess this comes from the "little guy in your head" trope, like the scene in "Everything You Always Wanted to Know About Sex". I'm taking the perspective of the "little guy in the CPU", and in that view, it's easier to imagine latency
Re: (Score:2)
I've got my doubts. Raytracing also strongly stresses the memory system. It's about as random access of a problem as you can muster (OpenRL (RIP) not withstanding).
Re: (Score:2)
Raytracing also strongly stresses the memory system.
Yep, memory bandwidth.
It's about as random access of a problem as you can muster
Nope.
Bandwidth-bound, not latency-bound, or else GPU's still wouldnt be doing it.
Re: (Score:2)
Thank-you for digging that up! That definitely sheds some more light on the situation.
More benchmarks on memory latency is definitely needed!
Re: (Score:2)
AMD is having a 1-2-3 knock out punch!
Actually the one place where Intel holds a wide lead is in market share. That's the real target. The hope is that AMD's technical benchmark wins translate into more market share. If that doesn't happen or slows down, then that's really bad news for AMD. Depending on the specific market share report [tomshardware.com], AMD has around 18% desktop share, 4% server share, and 15% mobile share. More importantly, the annual increase in market share is 3-5%, depending on the segment. In any case, it's clear that it's going to
Re: (Score:1)
Re:Threadripper 3990X 64 Core / 128 Thread confirm (Score:4, Informative)
Beyond pure benchmarks you get more PCIe lanes with Threadripper, and they are PCIe 4 where as Intel only offers V3. That means each lane has 2x the bandwidth.
On Ryzen it makes a huge difference with the CPU-Chipset bridge having about 4x the bandwidth that Intel's parts do. Quad NVMe RAID? 10g ethernet? Several USB 3.2 ports? Even the chipset PCIe slots have enough bandwidth for them now, not just the ones linked directly to the CPU.
Also with the Threadripper benchmarks most of them seem to have been done on air cooling. Modern CPUs dynamically adjust frequency to maintain preset power and thermal limits, so if you switch to water cooling you can get higher boost clocks for longer.
Re: (Score:2)
The chiplet design has an inherently higher memory latency, which is why on some workloads it struggles to keep up. But it seems most use cases the increased cache can hide or eliminate most of the disadvantage.
Competition -- finally! (Score:3)
Re:Competition -- finally! (Score:4, Interesting)
Yep, the last time AMD was the clear winner was when they brought out the Athlon, through about the Athlon XP. They had the best performance then, now they have the best performance again. I hope they kick Intel right in the dick for years to come. I'm looking forward to buying some of these umpty-hojillion core processors in used servers a few years from now.
Meanwhile, my FX-8350 is still doing what I need it to do, even though a budget AMD CPU is something like twice as fast now. We truly live in the future.
Re: (Score:3)
It saved their company last time, and its the only way to save their company this time also.
Re: (Score:2)
[--] and these CPU's are bloody expensive for relatively smaller returns on performance. If anything they are both neck and neck at the moment[--]
If you refer to the new Threadripper CPUs, they are far ahead from being neck-and-neck. I've seen reviews like "Total Intel HEDT annihilation". Even the non-HEDT Ryzen 3950X can outperform the 10980XE in many benchmarks.
If you refer to the desktop parts, then sure, in some gaming benchmarks the Ryzen parts can be some 5-10% behind the fastest Intel's parts. Basically Intel 9900K(S) and 9700K are the only Intel CPUs that make any sense at the moment, and even those if you're exclusively gaming, with a 2080 T
Re: (Score:1)
I hope they kick Intel right in the dick for years to come.
We don't want them to kick Intel too hard. If they're too successful they will become the next Intel. We need them to tick tock back and forth. That's ideal for consumers.
Meanwhile, my FX-8350 is still doing what I need it to do, even though a budget AMD CPU is something like twice as fast now. We truly live in the future.
A Ryzen 3900x is about twice as fast single-threaded and 4x multi-threaded.
Re: (Score:2)
I'm still coasting on a Phenom II, I never saw a point in moving to the FX generation. But you know that meme where the guy turns around to look at a hottie, while standing next to his girlfriend? Yeah, that's me right about now.
Re: (Score:2)
Re: (Score:1)
Have they fixed the RDRAND bug yet? (Score:1)
Re: (Score:2)
UEFI updates using AGESA 1.0.0.4 and higher fix the problem. The problem is that not all mobo makers have been updating to the latest AGESA versions(1.0.0.7) when they're released. To give another examples AGESA 1.0.0.3 which is still being shipped on budget and high-end gaming motherboards caused Red Dead Redemption 2 not to launch either, but reverting to 1.0.0.2 or updating to 1.0.0.4 it worked just fine.