12 ms·
AMD Ryzen 3rd Gen 'Matisse' Coming Mid 2019: Eight Core Zen 2 with PCIe 4.0
- reacharavindh 8y agoQuestion from a curious sysadmin: How are these AMD GPUs for machine learning and math related stuff? I have only heard of Nvidia CUDA in this context. Does anybody do ML work on AMD GPUs? Does the popular libraries like PyTorch, TensorFlow etc support both CUDA and AMD's equivalent?
- microcolonel 8y ago> Question from a curious sysadmin: How are these AMD GPUs for machine learning and math related stuff? They're just fine, though I think a tuned implementation on either tends to show NVIDIA ahead. > I have only heard of Nvidia CUDA in this context. Does anybody do ML work on AMD GPUs? NVIDIA is still king here, but if AMD continues to develop their Radeon Instinct stuff, it could be competitive for some applications; it then becomes a question of whether or not there will be enough people who know something other than CUDA (which is not currently supported on AMD GPUs, though work is under way to make a compatible runtime).
- celrod 8y agoHIP (part of the ROCm stack) is Fairlane similar to CUDA. They also have goodies like tensorflow-rocm, available through pip or a docker container. That stack is also open source. And writing in it is supposed to keep you compatible with NVidea. Wanting to support them -- and hoping they become more competitive in ML -- are why I bought Vega GPUs. Although, I haven't found much time to actually try and get any of my software running on a GPU, let alone optimized. As a huge fan of avx512, I would like to try -- maybe I can get even better performance on a graphics card. But with AMD, the lack of support for the uninitiated is apparent. Not much in the way of existing software and resources like online tutorials. I'd like guides on how to optimize kernels, organizing wavefronts and discussing memory movement. Maybe I haven't looked hard enough. Especially painful in Julia, where almost half a year after 1.0, the only supported way to use GPUs on any library is with CUDANative. "Cross platform" GPUArrays.jl's open cl backend still hasn't been updated. Means all my coding will be done in HIP. Which is fine.
- sremani 8y agonote: Not DL practitioner. As of today, if installing pytorch w/ GPU option the assumption is you have a nvidia gpu. Same thing seems to be the case for Tensorflow (did not install tensorflow lately) The first mover effect of Nvidia/CUDA should not be under-estimated.
- jjoonathan 8y agoThere's a lot of CUDA-specifc code out there. Even if you think you know exactly what you need, even if it claims to be compatible with AMD cards, you should seriously consider the possibility that you are wrong on one of those counts. I was, twice. Twice I had to sell my AMD cards, buy NVIDIA cards, and eat the spread. By comparison, just paying the "green tax" up front would have been cheap. I want AMD to win, I really do, but I can't personally justify taking chances on them anymore.
- jandrese 8y agonVidia provides way more support for the ML community than AMD, and the ML community reciprocates by buying nVidia hardware. You can use AMD stuff, but it means being on the bleeding edge and doing the debugging yourself instead of being able to ask the community for help (via Google searches). So yes you can use AMD hardware for it, but you end up making a lot more work for yourself to do so, and for little to no benefit to yourself.
- jamesblonde 8y agoROCm is AMD's "CUDA". TensorFlow is now up-to-date - but you have to work with AMD's fork of tensorflow, they haven't upstreamed their changes yet. The new GPU of interest for DeepLearning is the MI60. It has 32GB of RAM, 1TB/s memory bandwidth. Performance on fp32 looks slightly lower than the V100 (resnet50 on tensorflow, see below). But there are no tensorcores, so resnet50 would go 3x faster using the TensorCores. Personally, i think TensorCores are not that useful in the general case, as not enough programmers that i have come across are writing their code to use them. They are good, but nichey. The important thing we don't know is the price. Expect it to be released by the end of Q1 or early Q2. Reference: https://wccftech.com/amd-radeon-mi60-resnet-benchmarks-v100-tensor-not-used/ https://wccftech.com/amd-radeon-mi60-resnet-benchmarks-v100-...
- TwoNineA 8y agoInteresting, according to Anandtech, there might be space for something else on the package: https://images.anandtech.com/doci/13829/cpu44.jpg https://images.anandtech.com/doci/13829/cpu44.jpg Maybe a second Zen2 die?
- Symmetry 8y agoOr integrated graphics.
- jandrese 8y agoProbably for an iGPU for the budget model all in one chips.
- DCKing 8y agoThere were some rumors before this announcement that AMD would announce SKUs with 1 CPU die (6-8 cores activated) + 1 GPU die or with 2 CPU dies (8-16 cores activated). But AMD didn't announce any SKUs - seems they're a bit behind on schedule.
- hajile 8y agoThey are launching processors in the middle of the year. If they reveal too much 4-6 months out, Intel has way too much time to adjust their prices, change their marketing, put pressure on suppliers, etc. This leaves a lot of guesswork. Intel has to be prepared for multi-chip, chip + CPU, faster chips, slower chips (maybe this was a golden sample), etc which makes it much harder to adjust in preparation.
- okl 8y agoAlso, if they reveal too much prospective buyers might be incentivized to delay their purchase. As stupid as that might sound, it is a thing.
- mbroncano 8y agoIt even has a name: Osborne effect https://en.m.wikipedia.org/wiki/Osborne_effect https://en.m.wikipedia.org/wiki/Osborne_effect
- oliveshell 8y agoWow, performance competitive with i9-9900K at nearly half the power draw would be huge if it turns out to be true. Can’t wait for benchmarks. What an architecture Zen is.
- agumonkey 8y agoI wonder how AMD is planning long term. Zen did put them back on track, but I hope it's not a one shot era and that they will find ways to get strong for a decade or so.
- sitkack 8y agoChiplets and power are the next dimension of competition. Yield, chiplets vastly improve yield. The limit as chiplet size drops, yield basically goes to 1. Erasure Coding for Silicon. And then power, part of that is process and part is design. I am not skilled enough to call that, but as leakage current drops, then does power draw, static power draw, allowing for frequencies to scale over larger ranges. Hell, maybe went with some novel clockless design. Either way, this is momentous.
- kissiel 8y ago
- bratao 8y ago"Identical Performance to the Core i9-9900K, At Just Over Half The Power" This looks like trouble for Intel!
- jandrese 8y agoYeah, but that's comparing a mid 2019 chip to a chip you can buy today. Intel will probably make some announcement about their chip that's faster than this one right about the time this hits the shelves. It's the neverending rat race for chipmakers.
- zanny 8y agoUp to this point Intel has been competing with AMD by dramatically cutting their previously massive margins between what was possible on their tech and what they were milking consumers for with their enthusiast platforms. Intel won't be able to compete with this with another one of their Arizona fabbed 14nm chips. The techs 8 years old now, but they still haven't demonstrated a readiness in anything newer yet. This might be the pants down moment Intel can't bounce back from a few months later.
- fleetfox 8y agoIt looks like Intel 10nm desktop CPUs are going to be available only in 2020.
- wmf 8y agoIt's possible that Intel will announce a 10C Comet Lake but then AMD can quickly counter with a 16C Ryzen. Then Intel has no response unless they can cram an 18C -X die into an 115x socket.
- qball 8y agoI'm not so sure about that. The Intel playbook has, for the last 8 years, been either "same architecture, same clock speed, less power draw" (remember, Sandy Bridge [2011] and Broadwell [2014] perform within 10% of each other given the same clock speed- a consequence of chasing thin-and-light laptops and tablets) or "same architecture, higher clock speed, more power draw" after Zen hit store shelves. Instructions per clock haven't changed significantly in many years; Intel has just slowly gotten better at hitting higher clocks because smaller transistors pull less power. Overclockers have known for years that 5GHz is about the maximum that Intel's architectures can do under standard conditions (i.e. not liquid nitrogen), and the bottleneck there has primarily been due to heat generation and not processor instability. It hasn't escaped Intel that these kinds of clocks were easily possible- they had special CPUs sold at those maximum speeds for high-frequency trading applications (not sold to the general public), and that fact was being kept in reserve as a strategic advantage against potential competition. But that advantage has already been used up (not that AMD has one at the moment either; Zen and Zen+ have practically zero overclocking headroom on their fastest examples). Intel will need a new architecture to compete, and that's still likely at least a year down the road- and unlike the last time this happened (the Pentium 4 was supposed to reach 5GHz, but capped out at 3.8 for heat and power draw reasons) they don't have another architecture waiting in the wings to save them.
- bluedino 8y ago>> Cinebench is an idealized situation for AMD I wouldn't say it's especially suited for it, but AMD generally does well because of the large amount of cores/threads. On the other hand, something like 7Zip, AMD has had a huge advantage compared to Intel, even going back to the Bulldozer days. What is it about that code that runs so well on AMD?
- entropicdrifter 8y agoCompute (rather than I/O) heavy and highly multithreaded stuff skews towards AMD. Most applications benefit more from Intel's caching and single-core turbo boost features than they do from AMD's more discrete style of multithreading and weaker single-core mode tech.
- zozbot123 8y ago> Compute (rather than I/O) heavy and highly multithreaded stuff skews towards AMD. This. It's worth noting that newer, legacy-free programming languages make it a lot more feasible to parallelize compute-heavy parts of the code, and to deal with the "Memory wall" (i.e. the rise of memory-bandwidth as a bottleneck especially in high-CPU-frequency, high-core-count systems) by working with efficient, low-level representations of data. This will help not just AMD, but also newer entrants in this market segment such as ARM vendors.
- wbl 8y agoDo the rooflines of these chips show that pattern? Intel has a big vector unit with some seriously useful instructions in it, pushing the arithmetic speed up.
- mikenew 8y ago> AMD generally does well because of the large amount of cores/threads True, but in this case they're comparing an 8C/16T chip against the also 8C/16T i9-9900K, so this is a pretty apples-to-apples comparison.
- ineedasername 8y agoAMD seems to be making a push for the server market at the moment too. It looks like a good time for it as Intel struggles to get production on 10nm while AMD is already at 7nm for some of their range. But Intel has an incumbency advantage. I'm no expert on the server market though. Could someone with a bit more knowledge chime in with a how big a threat this really is?
- dunpeal 8y agoI don't see how this "incumbency advantage" would help Intel, exactly. The server market is highly competitive. If AMD puts out superior product at an attractive price, nobody is going to say "but we've used Intel for a decade, how can we suddenly switch to AMD?!".
- the8472 8y agoThings aren't always that simple. For example AMD may be cheaper per core, but if you're running AVX heavy code then intel has a throughput advantage. At least it did with Zen, Zen2 has some improvements in that area, so we'll have to measure again. Moving to the next intel generation is usually simple, incremental change. Moving between vendors means a lot of re-evaluating if you're crunching numbers. If your workload needs a lot of PCIe lanes for IO or you need cheap systems with ECC ram then the story changes quite a bit.
- d3v 8y agoThats exactly what they would say. Reliability track record, existing software/project restraints and enterprise relationships matter more than the price tag. That's why Intel can still charge rediculous amounts for enterprise chips for some time.
- hajile 8y agoI remember a time when Opteron had a very large part of the server market. It's about lack of competition, not about how businesses love Intel charging a fortune for their product. Who's making chips to compete with Intel? AMD, IBM, and nobody else (ARM isn't high-performance yet). IBM's lock-in with POWER is scary to businesses. AMD wasn't competitive for most server needs. That left Intel to charge whatever they wanted (and they did). With EPYC, there was finally a competitive offering at a great price (not to mention a better track record with meltdown and the constant stream of spectre variants). All the same software runs on the AMD chips too.
- gigatexal 8y agoI’m hoping to get a steak on the 2xxx lines. These 3xxx chips if they bench well should be awesome. Can’t wait to see reviews.
- erdewit 8y agoDoes anyone know why the IO-die is so large? Does it contain L3 cache or something?
- Tuldok 8y agoIt's 14nm. Also, it was speculated to have L4 cache.
- wmf 8y agoThe Zeppelin (Ryzen 1xxx) die is 212 mm2 and its CCXs are 88 mm2 leaving 124 mm2 for the uncore. Now we see that the Matisse (Ryzen 3xxx) IO die is 122 mm2 — virtually the same size on the same process for the same functionality. (I don't see how L4 cache would fit in there BTW.) As to why it's so large, I guess connecting cores, memory, and PCIe at extremely high speed just requires a lot of transistors. Intel's uncore seems to be far smaller; I'm not sure how or why.
- pmarcelll 8y agoThe Ryzen chips offer more PCIe lanes and ECC memory support. These chips also use the Infinity Fabric high-speed interconnect, which is not needed by Intel's single-chip designs. The two CCX units in Zeppelin also need to talk to each other, so probably all this complexity just adds up.
- hitpointdrew 8y agoThe IO specific transistors don't scale down well, so they left it at 14nm.
- pmarcelll 8y agoI thought the main reason is that AMD still needs to use GF's manufacturing capacity (according to their Wafer Supply Agreement) and it also makes the chiplet design cheaper (the high-performance node is only used where it really matters).
- 8y ago
- vbezhenar 8y agoNot released AMD CPU performs a bit worse than already released Intel CPU. They are still behind. But it's very intriguing. If they managed to push 5+ GHz, it would be awesome CPU.
- rukittenme 8y agoYou mean slightly better. The AMD CPU outperformed the 9900k. Wit the same core count. Which means AMD was likely turboing to 5GHz during the demo.
- sp332 8y agoThe article body shows the AMD chip slightly behind the i9. The chart just after it shows two different figures - "pre-brief" it's behind, "on stage" it's ahead. Not a huge margin in any case.
- dralley 8y agoPower consumption was significantly better. I'd say they're still ahead.
- vardump 8y agoWell, if they can turbo at 5 GHz while consuming nearly half as much power as the equivalent Intel part... it sure does sound like something that I might buy.
- mikenew 8y agoIt has almost identical performance at a little over half the power consumption. They're ahead.
- dijit 8y agoWhat are you comparing to what? Price/Perf goes to AMD every time, and increasingly often; some parallel workloads can simply perform better on AMD CPUs. I work on The Division 2 (My managing director was on stage during the keynote) and we get better performance in quite a lot of area's using AMD CPUs. Interesting fact: AMD EPYCs in our DC run with less power consumption per compute than our Xeon Skylake servers. What I later learned was that Intel takes the average power consumption and lists it, where AMD lists the max. Not sure why that is.
- m0zg 8y agoAlso, natively wider SIMD (256 vs 128 bit), though still no AVX512 IIRC. I'm waiting on third gen Threadrippers in particular. Now that will be one heck of a chip for a quad GPU deep learning workstation.
- hajile 8y agoIntel does avx512 at 30+% slower clockspeeds. AMD should be able to do an avx512 at full clockspeeds in double the cycles. For equivalent architectures, you'd expect Intel to be faster at workloads that are only avx512. They'd be slower at mixed workloads (the avx unit slows down, but so do all the other ALUs executing in parallel along with the SMT thread). Most importantly, 512 is 256 more sets of hardware which is a big addition to core size and power usage for what is a very fringe workload. Saving that while getting better performance in common workloads seems like a great tradeoff.
- m0zg 8y agoI'm aware. But it's not just about clock speed per se. AVX512 has some _very_ specialized instructions that are specifically designed to speed up matrix-matrix and matrix-vector multiply. Even at a significantly slower clock those instructions improve things by a lot if you do linear algebra. Most of linalg is done on GPUs nowadays, but not all of it.
- bitL 8y agoAs a consequence, TR/Ryzen is very slow in "classical" ML in Python (I can tell), because MKL/AVX2 seems to be way slower. This could be fixed with Zen 2; AVX512 then will be likely at 2/3 of Intel's performance without slowing down the rest of the system during computations.
- m0zg 8y agoI agree with this assessment. Frankly, GPUs are so much faster for dense linalg than even the highest end Intel (or AMD) chips that anything that can be done on GPUs should be done there. But that requires one to know how to work with them, if the primitives don't already exist, and most people are oblivious to the finer points.
- yazr 8y agoWhats the cache-coherency throughput of AMD vs Intel? Isn't this a major, major issue with >=8 cores ?! EDIT: yes, of course, lots of server code has very low memory contention.
- berbec 8y agoWhat I'm salivating for is an AMD 6c/12t mobile chip. With this power draw on desktop, chop the thermals and boost clock a bit and you have something truelly frightening on the battery life & burstable performance scale.
- ric129 8y agoBy "mobile" I'm guessing you mean laptops, and not phones. Energy consumption of AMD's x86-64 CPUs are only impressive when comparing with Intel, they're not competitive with current ARM CPUs.
- ben-schaaf 8y agoConsidering last generation's APUs used the same "chiplets" as the desktop processors, I'd wager the next generation of APUs will have a single 8c16t zen2 die.
- berbec 8y agoOoh.. A 3500G, 6c with Vega 15 (or whatever) would be nice!
- nicoburns 8y ago> This suggests that AMD’s new processors with the same amount of cores are offering performance parity in select benchmarks to Intel’s highest performing mainstream processor, while consuming a lot less power. Almost half as much power. > >How has AMD done this? IPC or Frequency? Am I missing something? Wouldn't this be expected given that this new AMD processor is running on 7nm while Intel hasn't really launched in 10nm processors yet. AMD is a process generation ahead...
- stoobs 8y agoThe processes are measured differently, so 7nm isn't physically 3nm smaller than 10nm in this case. I believe they said in one of the articles that more details on how they achieved the power saving will be forthcoming nearer the launch date.
- throwaway2021 8y agoI'll never buy AMD again, I had nothing but problems with Raven Ridge on Linux: https://bugzilla.kernel.org/show_bug.cgi?id=196683 https://bugzilla.kernel.org/show_bug.cgi?id=196683
- jtl999 8y agoYeah. I'm concerned about similar issues too.
- throwaway2021 8y agoNo need to be concerned, it's working fine with Arch Linux (Linux 4.20, mesa 18.3). It's still a problem with Ubuntu 18.04.1 (kernel 4.15), but Ubuntu 18.04.2 will hopefully be more stable with kernel 4.18.
- grecy 8y agoI assume PCIe 4.0 will support the next-generation of Thunderbolt?
- Dylan16807 8y agoI'm not sure what you mean. It should be straightforward to connect a thunderbolt bridge to any version of PCIe. If you specifically mean "without going over 4 lanes", then that's probably true.
- grecy 8y agoCurrently devices have Thunderbolt 3 (up to 40 Gbps), I'm wondering if PCIe 4.0 will increase that. I assume it would be called Thunderbolt 4, and I assume it will be faster.
- wmf 8y agoPersonally I would prefer Thunderbolt 3 to get cheaper instead of a new generation that stays expensive. Who needs 80 Gbps Thunderbolt 4 now and why?
- mmrezaie 8y agoExternal GPUs and Asics. Also, this is a connection that can be shared and the more is probably the better since we all probably have notebooks and they all gonna not have most ports we need.
- Dylan16807 8y agoIncrease it how? That's like asking if faster ethernet is going to increase thunderbolt speed. No, it's a totally different way of sending data over wires. A theoretical Thunderbolt 4 that's twice as fast could easily be fed with a PCIe 3 connection. It could even be fed with a PCIe 2 connection!
- grecy 8y ago
- sitkack 8y agoThis is going to own cloud. Costs are linear with power.
- stoobs 8y agoEven the previous generations of Epyc are surprisingly cheap compared to Intel - Azure has a L8s_v2 8 core with 64 GB RAM and decent storage for under £100/month, which is less than half the price of anything else. Obviously, core speed may not suit every application, but it's pretty eye-opening!
- sigi45 8y agoI was looking to buy a 9700k and wasn't considering amd for probably 8 years. Even with Nvidia and Amd, amd came out and than I bought a 1060. I might just wait this time for amd.
- reiichiroh 8y agoA Geforce GTX 1060 GPU?
- vkaku 8y agoMeanwhile ... their Radeon VII is still on PCIe 3 I'd rather they release a better chipset with at least 2 PCIe 3s than come up with these fast I/O paths not even commercially usable today. Such agony.
- bryanlarsen 8y agoReally? The mi60 uses the exact same chip as the Radeon VII, and it supports pcie4. They will gimp the Radeon VII in some places so they don't completely kill mi60 sales, so dropping pcie4 makes sense in isolation but is completely whack in the bigger picture. Doesn't the right hand talk to the left hand at AMD? Hopefully there's time to reverse this crazy decision.