Azul launches Vega2 Java machines768 cores and 768GBBy Charlie Demerjian: Monday 04 December 2006, 16:41AZUL UNWRAPPED its newest chip the 48-core Vega2 today. It is the successor to last year's Vega1. This Java processor triples performance from the year-old part and has an attractive feature set.Let's start out with the systems. Azul does not sell chips, only complete machines. There is a two-chip, 96 core, 48GB model 3210 and a four-chip, 96-192GB model 3220. Both are 5U machines, with the 3210 pulling about 580W max and the 3220 pulling about 1000W. There will be eight and 16 socket machines coming in 2007.The machines are all SSI or Single System Image, all cores can see all the memory. This may not sound like much of a feat, but remember that each chip is 48 cores. This means the biggest box has 768 cores and up to 768GB of memory, every last bit accessible to each CPU.The way it works is that each core is a simple CPU that runs one thread. The cores have 16K L1 cache, and 8 of these are connected to a single L2 cache. Those L2 caches are connected to an internal mesh, and a Vega2 is 6 of these clusters. There are 4 on chip memory controllers, supporting 12 DIMMs per CPU. The mesh connects them all in a non-blocking way.The CPUs externally are connected via another mesh, 16 sockets means 15 links per CPU. The link is a SERDES link similar to PCIe2 in that it has an embedded clock. If you are not getting the picture, everything is connected to everything else as directly as possible.What you end up with is a monster compute engine that plugs into a host PC. The cores themselves are tuned to run Java and other languages that live in a VM. The ISA for the Vega2 cores does not directly map to Java, but they map pretty well to the concepts in VM based languages.One interesting bit is that the Vega2 ISA is different from the Vega1 ISA. Between the generations, a lot was learned about what works, what does not work, and how things can be sped up. Azul is claiming that Vega2 is 3x faster than Vega1, and if 2x of that came from the core doubling, the other 1x probably came from the new ISA. Since the user never sees, touches or cares about the bare metal, Azul can tweak away with impunity, and they do.In the end, you get a box. Said box plugs into your rack and runs Java really stinking fast. The numbers they showed were SPECjbb2005, and they scored a little over 850K, beating out a Fujitsu 128 chip SPARC64 that scored over 800K. The Azul box starts at $49,995 for a 2-way and probably around $500K for the 16-way, so it is notably cheaper than the Fujitsu machine.For perspective, the 16 chip 384-core Azul 3840 (Vega1) scored about 325K, a 2-way Core2 at 3.0GHz hits about 100K, and an Opteron takes up the rear at closer to 50K. Azul holds the absolute record and probably the perf/watt record too.One very nice partnership is BEA, they have ported the entire portfolio of Java software to the Vega based machines, so you have a decent set of apps right off the bat. Since Java is portable (no snickering people!), theoretically you can put an Azul box on almost any machine and offload the Java processing to it.This brings up the most important question. We asked a fiend with a long history of Java coding experience. He has written good portions of several engines that run on the Azul boxes, and also has a Vega1 box at his beck and call. He gives it a thumbs up with a few caveats.The main one is your code must be able to utilise multiple CPUs. If each Vega2 core is equivalent to a 100MHz Intel core, and your software can only use one thread, you have a very slow box with 767 idle cores. If your code can use multiple cores, you have a machine that can truly fly through code.Since most server code is very amenable to threading, he thinks this is not a problem for the majority of applications out there. They will just run fast on the big box humming away in the corner. One nice feature of this type of setup is consistency. Since you have so many cores, chances are they are not all working at the same time, a percentage of them will be idle at any given time.This means that any task you fire off will be immediately serviced leading to a fairly predictable execution time. If you only have four vastly faster cores, each one will be heavily loaded, so a task might finish quickly or it might take a bit. Each Vega2 core may be slower, but it is more 'reliable' in a timing sense.The down side to this type of box is there is no local storage, so anything persistent needs to be offloaded back to the host and saved. For transactional systems, this type of logging is critical and imposes a huge overhead on Azul boxes. They still are faster than most equivalent machines, but with the overhead end up with less of a crushing victory.A CPU launch would not be a launch without the requisite list of partners declaring their unrequited love for the machine. Azul lined up some big ones, with BEA an obvious believer for porting the code, and BT another big one. BT has ported their Openreach B2B gateway to Azul, and since it is the world's largest gateway, yours won't be nearly as big or demanding.In the end, the user sees a box that runs code faster, and it runs the software they want or it does not. If it does, it can be a very clear winner for a number of scenarios but not all. The geeks see an interesting architecture that is actually commercially viable instead of simply a curiosity. Whatever you see it as, you can buy one now from Azul. µ
Meet Larrabee, Intel's answer to a GPUHello CGPUs, goodbye NvidiaBy Charlie Demerjian: Friday 09 February 2007, 14:36WE FIRST TOLD you about Intel's GPU plans last spring, and the name, Larrabee last summer. That brings up the question of just what the heck it is, other than the utter death of Nvidia.Intel decided to talk about Larrabee last week to VR-Zone (nice catch guys), so I guess that makes it open season on info. VRZ got it almost dead on, the target is 16 cores in the early 2009 time frame, but that is not a fixed number. Due to the architecture, that can go down in an ATI x900/x600/x300 fashion, maybe 16/8/4 cores respectively, but technically speaking it can also go up by quite a bit.What are those cores? They are not GPUs, they are x86 'mini-cores', basically small dumb in order cores with a staggeringly short pipeline. They also have four threads per core, so a total of 64 threads per "CGPU". To make this work as a GPU, you need instructions, vector instructions, so there is a hugely wide vector unit strapped on to it. The instruction set, an x86 extension for those paying attention, will have a lot of the functionality of a GPU.What you end up with is a ton of threads running a super-wide vector unit with the controls in x86. You use the same tools to program the GPU as you do the CPU, using the same mnemonics, and the same everything. It also makes things a snap to use the GPU as an extension to the main CPU.Rather than making the traditional 3D pipeline of putting points in space, connecting them, painting the resultant triangles, and then twiddling them simply faster, Intel is throwing that out the window. Instead you get the tools to do things any way you want, if you can build a better mousetrap, you are more than welcome to do so. Intel will support you there.Those are the cores, but how are they connected? That one is easy, a hugely wide bi-directional ring bus. Think four not three digits of bit width and Tbps not Gbps of bandwidth. It should be 'enough' for the average user, if you need more, well now is the time to contact your friendly Intel exec and ask.As you can see, the architecture is stupidly scalable, if you want more CPUs, just plop them on. If you want less, delete nodes, not a big deal. That is why we said 16 but it could change on more or less on a whim. The biggest problem is bandwidth usage as a limiter to scalability. 20 and 24 core variants seem quite doable.The current chip is 65nm and was set for first silicon in late 07 last we heard, but this was undoubtedly delayed when the project was moved from late 08 to 09. This info is for a test chip, if you see a production part, it will almost assuredly be on 45 nanometres. The one that is being worked on now is a test chip, but if it works out spectacularly, it could be made into a production piece. What would have been a hot and slow single threaded CPU is an average GPU nowadays.Why bring up CPUs? When we first heard about Larrabee, it was undecided where the thing would slot in, CPU or GPU. It could have gone the way of Keifer/Kevet, or been promoted to full CPU status. There was a lot of risk in putting out an insanely fast CPU that can't do a single thread at speed to save its life.The solution would be to plop a Merom or two in the middle, but seeing as the chip was already too hot and big, that isn't going to happen, so instead a GPU was born. I would think that the whole GPU notion is going away soon as the whole concept gets pulled on die, or more likely adapted as tiles on a "Fusion like" marchitecture.In any case, the whole idea of a GPU as a separate chip is a thing of the past. The first step is a GPU on a CPU like AMD's Fusion, but this is transitional. Both sides will pull the functionality into the core itself, and GPUs will cease to be. Now do you see why Nvidia is dead?So, in two years, the first steps to GPUs going away will hit the market. From there, it is a matter of shrinking and adding features, but there is no turning back. Welcome the CGPU. Now do you understand why AMD had to buy ATI to survive? µ
No sir, graphics is something we change with regularity.
