17 ms·
Details Emerge on China's 64-Core ARM Chip
- faragon 10y agoThat's direct competition for Cavium (american multicore MIPS -max 48 cores/chip on Octeon III SoCs- and ARM -max 48 cores/chip on ThunderX SoCs- chip maker). If they manage to sell 64-core ARM chips for 50$, even using just 28nm (maximum available on mainland China foundries -AFAIK-), it could hurt Cavium badly and start worrying Intel.
- dogma1138 10y agoThey aren't going to sell these chips anywhere close to 50$, high end mobile phone SoC's cost 60-100$ these days, even the Chinese SoC's are not that cheap MediaTek's Helio X25/X20 are "cheap" because they aren't Snapdragon 823 Exynos 8893 100$ price parts, but they are around 30-50$ also (and could be even more, since a bargain bin MT6582 MediaTek ARMv7 SoC costs 8-10$). This isn't going to be fabbed in China, almost none of the chinese SoC's are fabbed locally the foundries are not built for that the process is too old and overall they are not intended to produce such complex silicon, there is much more to a process than just the lithography and even that is not straightforward because you actually need to do tricks to get to sub 28nm today as you often end up using higher wavelength light since extreme UV and other similar technologies aren't there yet. The 64-core SoC's are going to be multi-100$ parts, probably closer to the 1000$ mark regardless if they are chinese or not, Cavium SoC's are about as expensive or even more expensive than some of the Intel parts they are competing against. If China introduces a 64 core ARMv8 SoC at 50$ Intel won't be worried, it would panic to the point of them using the backdoor everyone on HN think they have via AMT rigging the election so Trump can win and then convincing him to nuke China might actually happen. That said a 50$ 64 core ARMv8 SoC is pretty much impossible to make even in China, by China, for China.
- faragon 10y ago"That said a 50$ 64 core ARMv8 SoC is pretty much impossible to make even in China, by China, for China." Allwinner A20 (4 core ARMv8 Cortex A53, 512KB L2 cache, 128KB L1 cache, 2 core Mali 400, video interface, etc.) costs 6$ in volume. So I guess just in plain die area, the 64 core Mars ARMv8 (32MB L2 cache) could take just 8 times the area of Allwinner A20, without including L3 cache. That would mean 48$ just in silicon, with naive extrapolation. Also, there are two versions of the chip: the high-end ("Mars") and a low-end ("Earth"), may be without L3 cache ([1]). With progressive die reduction, it could be even cheaper, if demand reaches enough critical mass (e.g. billions of chips for routing devices, servers, etc.). In addition to Cavium, there is another potential competitor: Mediatek (after Atheros aquisition), selling crazy cheap SoCs for SOHO routers. What if they start in the high-end and server market? (they would need just IP available from ARM for the inter-core memory routing, for making cheap SoCs). [1] https://tweakimg.net/files/upload/HC27.24.320-Mars-64core-Gao-Phytium-v1.0.pdf https://tweakimg.net/files/upload/HC27.24.320-Mars-64core-Ga...
- dogma1138 10y agoSoC production isn't simply a question of die size it's a question of yield. Larger dies would have smaller yields simply because say you have 25% defected dies with an 4 core SOC, when you scale it up to something 8 times it's size your defects are going to have a very severe effect on your yields to the point where you can't actually produce SoCs above a certain die size simply because it's more likely than not that all or almost all of them will have defects. Also the older A53 copy paste SoCs are cheap, the licensing is cheap, and little to no R&D went into them, Allwinner, MediaTek etc. all have sub 10$ bargain bin SOCs (The A20 is after all a 3-4 year old 55nm SOC that wasn't cutting edge when it came out ;)) but their high end SOCs aren't dirt cheap they are just sufficiently below the likes of Qualcomm and Samsung to be competitive. Don't get me wrong eventually can get there, but the question is do you even need to get there, by the time you can get a Mars like SoC down to a 50$ price point would that SOC be even relevant? Technology constantly evolves x86 and even ARM CPUs are not cheaper, they even increase in costs in as the likes Samsung and Qualcomm charge more year after year and even MediaTek is now stepping into the higher costs higher margins game also.
- xamlhacker 10y agoAtheros was acquired by Qualcomm, not Mediatek.
- faragon 10y agoYes, thank you. Mediatek acquired Ralink.
- faragon 10y agoI meant Allwinner A64 (4 core ARMv8 Cortex A53, 512KB L2 cache, 128KB L1 cache, 2 core Mali 400, video interface, etc.), and not A20 (4 core Cortex A7).
- faragon 10y agoI meant Allwinner A64 (4 core ARMv8 Cortex A53, 512KB L2 cache, 128KB L1 cache, 2 core Mali 400, video interface, etc.), and not A20 (4 core Cortex A7).
- 10y ago
- PieterH 10y agoIt's more a question of when than whether, isn't it... $1000 for a CPU today means $50 in six years, if there is real competition, which there is here.
- faragon 10y agoWell, I just would like a 64 ARMv8 core cheap board for less than 100$, in few years :-D
- dogma1138 10y agoAlso 6 years old, people aren't building servers today with Xeon E7530's.
- GrumpyYoungMan 10y ago>It's more a question of when than whether, isn't it. No. That kind of cost reduction came mostly from process shrinks and there have been plenty of articles on HN explaining why that's coming to an end in the very near future. Even Intel, the process leader, had to end its "tick-tock" cycle with its latest set of chips.
- astrodust 10y agoIt does seem like Intel is going to see their market share in the server space erode not from traditional competitors like AMD, but by upstarts using ARM. Intel's Xeon Phi experiment hasn't produced anything cost-effective and workable yet, but if they can get that to run at low power they might have an answer to the swarms of ARM systems on the horizon.
- dogma1138 10y agoXeon Phi isn't an answer to ARM, Intel has the Xeon SOC which is directly competitive with ARM and it's pretty damn good. http://www.intel.co.uk/content/www/uk/en/processors/xeon/xeon-processor-d-family.html http://www.intel.co.uk/content/www/uk/en/processors/xeon/xeo... No 64 cores, but 16 X86 Xeon Cores @ 1.3ghz with HyperThreading will give pretty much any high end ARM SOC a run for their money, a TDP of 45W also doesn't hurt ;) Intel has made a bet that by the time ARM would get their performance to x86 levels Intel would get their power consumption to ARM SOC levels and Intel seem to have been correct on that bet. The Xeon Phi is also pretty interesting Knight's Landing and onwards provides pretty good performance for deep-learning and other applications that companies like NVIDIA were swinging at, it's still not on the same performance level as NVIDIA's top tier of compute products but it's close and the x86 compatibility is likely to have a sufficient advantage in certain scenarios to make the Xeon Phi really competitive. NVIDIA was planning on having x86 compatibility for it's HPC parts however both AMD and Intel killed that dream in court.
- CravAnon 10y agoI suspect with the embargo on Intel, China is highly motivated to push ARM ahead of Intel quickly. If you think Phytium and the Chinese government aren't tied at the hip you're kidding yourselves. Sure, they can't fab at 14nm but they likely will soon.
- dogma1138 10y agoThey can fab on 14/10nm just like all the other Chinese soc companies do, MediaTek is doing 20, 16 and 10nm SoC's if they couldn't they wouldn't be competitive regardless of their price.
- biot 10y agoIs Phytium owned by the Chinese government? Otherwise, why call it China's chip? It's like saying America's Xeon.
- rch 10y agoI've been thinking about this lately: what's the difference between how China supports companies, and the U.S. giving Intel a big DOE contract?
- nickpsecurity 10y agoChinese companies often copy and improve on each others' work. U.S. companies often patent and/or copyright the work then sue many with profitable remixes. The U.S. inventions can be like throwing money into a big hole if wrong company ends up with the patent or copyright. I noticed this distinction studying Shenzhen vs US tech sector.
- twblalock 10y agoHonestly not a lot. People who complain about the Chinese government helping Chinese companies need to recognize that the US has been doing similar things for decades. China does seem to give more support to its companies than the US does, but that's a difference of degree, not of kind. In fact, the development of Silicon Valley as a tech hub in the first place was heavily dependent on the military-industrial complex -- Fairchild had considerable military contracts, for example, and the companies that followed have had such contracts as well. This is a situation where free trade and national security can be in conflict -- it would simply be irresponsible for the US or China to trust non-domestic companies with military projects, satellites, etc. But most of the companies that work on those projects also compete in the consumer market, where government subsidies are considered trade barriers.
- jomamaxx 10y ago"Honestly not a lot." and "China does seem to give more support to its companies than the US does, but that's a difference of degree, not of kind." I'm sorry but this is not true. Chinese companies are often a direct part of national strategy and are created by, led by, supported by and defended by the state. The collusion between many Chinese companies often goes far, far beyond 'government contracts'. It's rational for any government to prefer local companies for strategic purchases, particularly military etc.. Nobody is going to bat an eye if a state related carrier choses to buy planes from their national aerospace entity. But China will force companies from other countries to give up their IP - then hand it over to state-owned or related companies, then 'hold up' the foreign company in approval processes and other forms of tariff barriers - while the 'state owned' company gets a head start in the local market. And obviously having direct government ownership in many of these companies means a material intervention far beyond what the US does. The US does not own any piece of Apple or Intel, and makes investments in entities like GM only very reluctantly under negative conditions to keep them alive - not to promote a national strategy. Finally - because of the heavy state owned banking system and the total political control over monetary policy - issues like lending and financing are also heavily politicized, which is another form of direct state intervention. The means by which the US or UK intervene are on average completely different from how it's often done in China.
- wmf 10y agoWhat the article doesn't mention is that no independent benchmarks will be released, no English documentation exists, and the chip won't be available outside China.
- isjustajokebro 10y agoDo Intel and Amd release Chinese documentation?
- stass 10y agoThis is not a fair comparison. English is an international language of science and technology unlike Chinese and other languages.
- chrischen 10y agoThat doesn't preclude Chinese from being the same, or becoming the same.
- duaneb 10y agoI agree wholeheartedly. I don't necessarily understand alienating billions of potential customers on either side, though. Intel and AMD should release mandarin documentation.
- jomamaxx 10y ago"That doesn't preclude Chinese from being the same, or becoming the same." + It won't. The international language already English and that won't change (India, North America, Europe - most of Africa, the Middle East are already English as international lang. even China esp. Hong Kong, Singapore) - but Chinese may become a 'sphere of influence' language in surrounding countries - much like Spanish or French. Of course it kind of that already. + The issue of not releasing documentation in English is odd, given that English docs would enable them to access maybe a 5x greater market. And for the cost of what? Clearly, it's a strategic decision on their part - there's some kind of impetus behind it. Possibly they don't think it's ready. Who knows? But it's not likely an oversight. It was a decision.
- niftich 10y ago> 128 MB of L3 Wow! I suppose when you have 64 cores to feed... But what's the play here besides promoting indigenous development? Alone, these cores are not remarkable, and getting 64 of them fed constantly is only useful for long-running HPC workloads. To me the most interesting thing here is the 'Hawk' interconnect that allows some of these dies to be replaced with custom processing units; hardware encode/decode blocks sn a common option in consumer devices today, but any custom block can be made for specialized workloads.
- trsohmers 10y agoThe 128MB L3 is actually off chip... it is made up of 8 "CMC" (Cache & Memory Chips) that combine DDR3 and 16MB of L3. Each "panel" of 8 cores is connected to a single CMC. All I know is that will be extremely expensive and low yield, and as far as I know, they haven't shown any of these "CMC" chips. 120W for the base chip, and even more power for L3 and DRAM (plus them being off die) will make it even higher latency. This is pretty DOA in terms of being a real competitor against Intel.
- TazeTSchnitzel 10y agoDon't some Intel Xeons have 128MiB of L4? But L3 is another story, for sure.
- Coding_Cat 10y agoI think the iris graphics ones can use the vram as shared L4, and that's up to 256 MiB IIRC.
- Animats 10y ago2,892 pins, 16 memory controllers. Hm. How much is shared? Is this a cluster of 4-CPU shared memory multiprocessors on one chip, with inter-system channels? Or are all the caches linked, so that it looks like one giant multiprocessor? This setup has lots of memory boards in sockets, some distance from the CPU. If you're not sharing memory, it might be faster to have fewer CPUs per chip and put the CPUs and the memory on the same board.