LucidLogix is got a lot of attention with their vendor-independent approach to scaling graphics cards. The promise of putting both ATI and nVidia graphics cards [and Intel's Larrabee in the future] in a single system is somewhat of a dream. Thus, we decided to sit down and talk with Offir Remez, co-founder, President and VP of Business Development at LucidLogix Inc. In a brief interview, we disclose how LucidLogix created the new I/O chip that can handle
Who is LucidLogix?
BSN*: Hi Offir, thanks for taking the time to talk with us. Can you tell us a little bit more about LucidLogix?
Offir: I am co-founder, president and VP of Business Development over at LucidLogix. We are a four year old company and we worked hard on this technology. As you know, any decent engineering project takes 3-4 years of development in order to come to market with a mature product. That is the hard way but we decided to pursue it and now the effort is being rewarded.
The company has 60 employees, most of them working in engineering. The work of the team was awarded with more than 50 patents in relation to our technology, and we are now ramping up our business in Taiwan and in the USA [LucidLogix is an Israeli company, Ed.].
BSN*: If we understood correctly, the company is fiscally stable [Venture Capital-backed] with one of investments being made by Intel Capital. Can you elaborate a little bit more on the relationship?
Offir: There are no secrets that we have a relationship with Intel. Their investment arm analyzed the company, found logic in our business model and decided to invest in us. Usually, this kind of investment is quiet but our technology is such that it is not regarded as a secret. Our technology approaches [graphics] scalability from another perspective and we consider that we’re more efficient in it. That is the reason why Intel Capital and other investors invested in us and have high expectations. But not as high as we have.
How Hydra technology works?
BSN*: How did you got to want to develop a vendor-agnostic approach to multi-GPU market?
Offir: First of all, I am a father and both my son and I are gamers. With the graphics cycle of new graphics coming every four months and the balance of power continuously shifting between ATI and nVidia, it was a frustrating experience to have to change the system every couple of months. After a while, I started to analyze what was going on and saw that they [GPU vendors, Ed.] are approaching the graphics from the wrong perspective. Graphics technology is now computational, not fixed-pipeline. There is no need to render a whole frame – that approach can only go so far.
BSN*: Can you tell us how your scaling technology actually works? When we discuss about multi-GPU scalability, ATI and nVidia have each GPU rendering frames, such as AFR [Alternate Frame Rendering], SFR [Split Frame Rendering] and several other implementations.
Offir: What limits your scalability are frame dependencies, due to use of primitive AFR modes. Today’s games have such dependencies so big that in many cases, you even have to go five frames back to take the information from it. Even with triple buffering, five frames are too big of a drawback for the GPU because the chips cannot go back that much and that results in increased latency – they simply don’t have enough cache to do that. Thus, scaling performance is nowhere near where it is supposed to be.
BSN*: What is needed from game developers in order for Hydra to work?
Offir: Nothing. We have a real-time task distribution engine. The frame buffer stream is being analyzed and checked on-the-fly. By approaching graphics on computational level we resolved frame dependency. We tested our new Lucid 200 chip on 100 games and it worked out of the box. We did have some artifacts, broken pixels and so on but that was solved with our own driver.
BSN*: How did you resolve the issue, have you talked with developers?
Offir: Like I said already, we didn’t talk to game developers about our problems. Everything was solved by frame buffer stream analysis. As we improve our detection algorithms the code will be even more intelligent.
BSN*: First demonstrations of Lucid Hydra technology were in HPC [High Performance Computing], not in graphics space. Did you do that because of simplicity working with Tesla cards or??
Offir: Our chip has many cores – we have a DNA processor capable of handling very high traffic between the CPU and the GPU or GPGPU. We don’t use over-the-top connector. We designed the Hydra chip to control all the traffic and analyze the data stream.
Internally, our chip works at 200 Gbps [25 GB/s] or five PCIe Gen2 x16 lanes. This resulted in somewhat unexpected interest from the HPC market, especially Japan.
Our Asian partner [ELSA Japan] came to us and asked us to use derivative of the chip for their extender GPGPU box. The chip was used not for scaling performance, but for scaling connectivity. At first, we didn’t thought of it at all, but engineers at ELSA were thrilled with the achieved performance. In the past, such boxes had to use controllers from the companies such as PLX and performance often got lost in the process.
They tested Tesla GPGPUs, GeForces, Quadros and even connected storage controllers together with the GPU, removing one large bottleneck.
BSN*: In the end, they potentially gave you a whole new business line.
Offir: ELSA took our evaluation board [now a retail product ? Adventure 2000], did all the proper certifications and took it to market. Their part comes with four PCIe Gen2 slots and Elsa ships the part with two slots being used for GPU Computing [Tesla C1060] and two slots for 1TB SSDs [OCZ Z-Drive]. Fifth PCIe Gen2 lane is being used for external communication.
All the communication is handled by Lucid LT24102 chip and the boards; there are no connections back to the CPU! This is what made Elsa’s engineers so enthusiastic. With this move, we can see that PCI Express is stepping onto Infiniband territory. However, after all being said and done this is not our focus market. Our focus market is consumer PC and notebook gaming segment.
Lucid’s Hydra 200 chip – the beginning of something new
BSN*: On IDF, you officially introduced Hydra 200 chip with MSI being the first vendor out with their Big Bang motherboard. In order to gain market traction, naturally you need to have more than one vendor pushing your product. Can you tell us with how many top tier motherboard makers are you working with?
Offir: We work with majority of Top 5 motherboard makers. I cannot disclose the number right now, but we have multiple chips on offer and you should see them starting to come to market during the final quarter of this year [starting on October 1st, ending on December 31st, Ed.].
We are targeting upper mainstream and high-end market segments. Just like in any [smart] business, you start at the top and then bring the product into the mainstream.
BSN*: Initially, you’re targeting Core i5 and Core i7 systems, correct.
Offir: We hope to go into the high-end market and give best performance possible. Our main goal is to solve the issue of new GPUs arriving every quarter or so, and for that a debut in high-end market and going upper-mainstream is the best way. Our partners are targeting the Core platforms with best possible GPU scaling.
BSN*: Can you disclose us what manufacturing process are you using for Hydra-capable chips and what are the thermals achieved by your parts?
Offir: Certainly. So, Lucid 100 or the first generation was built using 130nm process and with Lucid 200 series, we moved onto 65nm process. This helped our power consumption, which we consider to be really good. Lucid LT24102 chip eats up only 6W of power and practically does not require a heatsink at all, driving the cost down. As I said earlier, it supports five PCI Express Gen2 X16 lanes.
We also have two lowed end chips, out of which one is which is packed in only 18x18mm [whole Lucid 200 package is 23x23mm] and that chip consumes only four watts. Capabilities-wise, both the small footprint LT22114 chip and regular sized, cost-effective LT22102 chips come with 3-Way PCIe Gen2 x8 instead of five.
BSN*: Who would you consider to be biggest competitors for Lucid 200 chips?
Offir: In any case, our competitors are native solutions such as ATI CrossFire and nVidia SLI, followed by bridge products such as nForce 100/200 bridge chips. But for being vendor-agnostic, Lucid 200 is the only solution on the market.
BSN*: Thank you for this conversation.
Long story short – Hydra is [almost] here. After four years of development, LucidLogix is shipping products to leading motherboard vendors and we already have a schedule for reviewing Hydra parts coming from the several top tier vendors.
We cannot give out a final word in this article without mentioning the reason why Intel Capital invested in LucidLogix [and the one that neither LucidLogix nor Intel will ever confirm]; Larrabee’s original roadmap. The plan looked like this: Larrabee comes out in June 2009, LucidLogix delivers Big Bang-like motherboards in Q4’09, and in November 2009, Intel introduces GPU scaling which is completely autonomous from Intel’s engineers. That was the plan. However, LucidLogix delivered a cross-vendor multi-GPU product between 15-24 months ahead of Larrabee and now nVidia and AMD have a good new player to think of [and time to maybe re-think the AFR/SFR strategy].
MSI Big Bang motherboard is launching on October 29, 2009 and a new era will begin in the multi-GPU field. There is no doubt in our minds that the shift in graphics happened already. Furthermore, it’s quite odd to see that both AMD and nVidia kept pushing AFR modes instead of creating a computational scaling path. If Lucid Hydra technology proves to be efficient, both companies will have no choice but to abandon their own "primitive" ways and move forward with the technology.
Our take on LucidLogix is positive. This is the first vendor to approach GPU scaling on a whole new level, not thinking about frames, but rather streams of data. We have to evaluate the product in practice, but from a business and technology standpoint – these guys and girls are here to stay.