22 ms·
Now that AMD has closed the gap, and passed Intel’s per core performance in the 16-core CPU market, it has a platform with more RAM capacity and more PCIe lanes
by zaroth 8y ago
Now that AMD has closed the gap, and passed Intel’s per core performance in the 16-core CPU market, it has a platform with more RAM capacity and more PCIe lanes along with more performance than the Intel Xeon Gold 6142M, at around a quarter of the price.
What does Intel have so far up its sleeve that AMD has to virtually give away its chips like this?
- gameswithgo 8y agoAX2 and AVX512 But probably more of a factor is that people are used to them and trust them.
- pythonaut_16 8y agoIsn't Zen 2 adding support for AVX512?
- bitL 8y agoYes, but a slow one. They just bump AVX units from 128-bit to 256-bit, making it Haswell-style. Skylake-X has a full 512-bit wide AVX unit.
- bpye 8y agoDoesn't Skylake-X have the whole downclocking issue when running AVX-512 workloads though? Unless you can use almost exclusively AVX-512 then it doesn't make sense. https://blog.cloudflare.com/on-the-dangers-of-intels-frequency-scaling/ https://blog.cloudflare.com/on-the-dangers-of-intels-frequen...
- gameswithgo 8y agoAVX and AVX512 both do downclocking. Likely AMD will have to downclock too, pretty fundamental limits at work there.
- bitL 8y agoWith 7nm they might not need to downclock that much.
- hajile 8y agoAMD verified they would be running AVX2 at full clocks. Doing AVX-512 would happen at half-rate, but the processor would continue at full speed. In contrast, skylake downclocks base clocks by 30+%. Not only does the AVX calculation downclock, but all the other calculations in other ALUs/FPUs also downclock. For pure AVX-512 workloads, Intel would be a little faster, but for mixed loads, AMD should be much faster overall. AMD says they don't downclock AVX2 (about 2/5 down the page) https://www.anandtech.com/show/13578/naples-rome-milan-zen-4-an-interview-with-amd-cto-mark-papermaster https://www.anandtech.com/show/13578/naples-rome-milan-zen-4... See page 13 for AVX clockspeeds https://www.intel.com/content/dam/www/public/us/en/documents/specification-updates/xeon-scalable-spec-update.pdf https://www.intel.com/content/dam/www/public/us/en/documents...
- paulmd 8y agoAVX-512 isn't just a wider vector, it also fills in some holes in the AVX instruction sets. It's still useful to support it, even if they are only doing it at AVX2 throughput levels.
- dragontamer 8y agoCitation please. I never heard of Zen2 supporting AVX512.
- shaklee3 8y agoAvx512 is very new, especially to the server market. There likely aren't many people absolutely relying on it yet.
- coredog64 8y agoISTR seeing some benchmarks where using AVX512 really hurts your throughput if you're running multithreaded code. Again, vague memory, but the reason was that it winds up heating up the core that's running the AVX512 code and causing all the other cores to thermally throttle.
- wstuartcl 8y agoYes, exercising AVX512 on intel (very much at all) thermal throttles in most workloads.
- drewg123 8y agoThe mindshare of mid- to upper-level execs within its customer's organizations. EG, in today's "Nobody ever got fired for buying IBM", "IBM" is Intel.
- Traster 8y agoMy background is in the FPGA industry which gives an absolutely fantastic example of this exact dynamic. FPGA companies produce a new generation of chips every 3-5 years or so. There are only two companies, with market share split 60:40. In the last 2 decades we've seen one produce a fantastic product and the other screw up, and then it flips, and then it flips back. Over that time the market share of these companies went 60:40, 62:38, 58:42, 60:40. The actual number of customers who switch to the better product is tiny. Why? * All the existing knowledge in the company is about one platform, it's incredibly expensive to develop the skills on the other toolchain. Port all your software etc. * The existing products are all from one vendor so you save loads of effort if all the products are basically similar. * There are existing relationships with the company you're with. * You know that if you do switch, all that cost of switching may result in only a few years of using the better product. * The IP you're buying works better with the vendor you're currently on. * You don't know what the real world performance will actually be for your application. So yeah, you could move from Intel to AMD, but the chip is only a tiny part of that cost.
- Fnoord 8y agoGiven the future doesn't look very bright for Intel, given ARM at one hand and AMD at the other (with their 10 nm and 7 nm having multiple setbacks) I wouldn't count on Intel doing well in the future. It might end up like Nokia or Microsoft. I wouldn't invest in Intel...
- vrazj 8y agoBrand value, I guess? All AMD hardware (CPUs and GPUs) I've bought in the past was crap, I'm never buying anything from them again, no matter what they do. It's always the same: on paper they are better and cheaper, but once I have them at home they are worse and, in case of the GPUs, they have pathetic drivers. I've always felt ripped off after buying AMD, it's never worth it just to save some pennies. This has been my experience, I wonder if others have felt the same.
- dkersten 8y agoI've had both Intel and AMD cpu's and both AMD and Nvidia GPU's (I guess you could add Intel too for their integrated gpu's) and my AMD experience has been generally quite pleasant and problem-free.
- kllrnohj 8y agoWhenever AMD's CPUs have been price/performance competitive they've always been a fantastic choice. Especially now, when with Ryzen/Threadripper/Epyc you not only got more cores for the less money, you also got better security. AMD's GPU drivers have been perfectly fine since the 9700 Pro. That's a 16 year old myth at this point, let it go.
- throwaway2048 8y agoIts really not a myth, AMD drivers on windows have completely terrible openGL performance for instance.
- kllrnohj 8y agoDo you have any benchmark comparisons to support this claim? All I can find in searching are random threads of people saying this, but nobody providing any actual evidence or comparisons. The few games using OpenGL on Windows I can find show AMD's performance being perfectly competent: https://www.pcgamer.com/doom-benchmarks-return-vulkan-vs-opengl/2/ https://www.pcgamer.com/doom-benchmarks-return-vulkan-vs-ope... https://www.hardocp.com/article/2014/05/21/wolfenstein_new_order_performance_review/4 https://www.hardocp.com/article/2014/05/21/wolfenstein_new_o... The nvidia cards were on average a bit faster, but that was true in DX as well. Meaning it wasn't just "lol amd opengl drivers" And of course Linux testing, where OpenGL is far more common, shows no major disparity, either, which you already know hence the qualifier of "on windows" but that qualifier makes no sense. It's going to be the bulk of the same code between windows & linux for the driver, as has also been rather well tested & verified. Which gets back to "it's a goddamn 16 year old myth" that literally spawned out of Nvidia's hyper aggressive opengl optimization of Quake 3.
- mirekrusin 8y agoMarket share?
- dman 8y agoA near monopoly
- StillBored 8y agoWhy do you think $1500 for a 16 core part is giving it away? There was a time not long ago where $1500 for an intel xeon was a solidly high-end part. This part is really nice, but it looks pretty low end when compared with dual socket machines with 128 cores. Its likely AMD is making a healthy profit on it and has an estimate about how much of the market they are going to gain by significantly undercutting intel vs the profits per part. AKA if knocking 10-20% off their proffit increases their market by 30-40% its a no-brainer.
- paulmd 8y agoIt's probably more that they're clearing out the old chips before launching their 7nm-based Zen2 Epyc chips next quarter. They've already sampled out to hyperscalers. But, generally, Epyc does have some deficiencies. It's essentially NUMA-on-a-package, and each NUMA node is itself essentially a pair of 4-core processors jammed together (each with its own cache, with all the usual problems that brings). That doesn't work for everything... for example GPGPU compute doesn't really like to be split across NUMA. Inter-core and memory latency is also much higher than on Intel platforms. Intel is playing with much larger building blocks, their die is 28C and they can scale up to 8 sockets, while AMD has 32C per package they can only scale up to 2 sockets because they are actually four dies inside already (both systems scale to 8 dies). A lot of stuff is fine with those tradeoffs, particularly the stuff you use server processors for. And it's pretty cute if you want a lot of storage or lanes. And it's pretty cheap to manufacture due to their smaller dies (although of course TCO is much larger than just the cost of the CPU). But it's not for everyone. Next quarter they're moving to 8 dies in a package, on 7nm, with updated AVX2 and probably support for the AVX-512 instruction set (at half throughput).
- tracker1 8y agoThe Zen 2 line is definitely looking interesting to say the least. I think where the Ryzen is going could be interesting in the server space as well. Ryzen 3600G rumored for example having GPU and CPU cores through the common IO interface. It wouldn't surprise me to see similar approaches in the workstation/server space. Wonder if there's been consideration for ARM compute units even. It's definitely an interesting approach.