WEBVTT 00:00:00.400 --> 00:00:09.840 To understand why a bizarre device like this one here exists, we need to travel 00:00:05.680 --> 00:00:12.559 back to the time of its release. 00:00:09.840 --> 00:00:17.520 The year was 2006. High performance consumer SSDs were still 2 years away, 00:00:15.360 --> 00:00:20.960 an eternity in computer hardware terms. And even the fastest consumer hard drive 00:00:19.520 --> 00:00:25.519 on the market, the Western Digital Raptor X, barely stood out from the 00:00:23.039 --> 00:00:31.840 pack. So, there was a very real appetite for faster boot devices. Meanwhile, 00:00:28.720 --> 00:00:33.600 Intel's Core2 Duo Conro platform was one 00:00:31.840 --> 00:00:40.960 of the most compelling upgrades for gamers in years, but it required an 00:00:37.760 --> 00:00:43.440 upgrade to shiny new DDR2 memory, 00:00:40.960 --> 00:00:49.760 resulting in an abundance of Castaway DDR1 memory sticks. So, there to 00:00:46.559 --> 00:00:52.559 capitalize on these two conditions was 00:00:49.760 --> 00:00:59.120 Gigabyte. The idea here was really simple. Here we go. Take some of the 00:00:56.239 --> 00:01:04.720 leftover DDR1 memory that you already have lying around and instead of 00:01:01.440 --> 00:01:07.360 spending thousands of dollars on an SSD, 00:01:04.720 --> 00:01:18.880 take it, chuck it on this thing, and use that as your system boot drive. 00:01:18.880 --> 00:01:24.880 I never had one of these when they were new. Back then, I was busy spending all 00:01:22.400 --> 00:01:29.600 my money on textbooks I'd never read for a degree that I'd never finish. That 00:01:27.439 --> 00:01:34.400 worked out great. So, when I saw this on eBay for like 30 bucks, I figured, what 00:01:32.159 --> 00:01:38.079 the hey, let's take a look at it. Let's start with the PCI connector. Now, on 00:01:36.560 --> 00:01:44.960 the surface, you might think that this looks very similar to a modern PCI 00:01:40.720 --> 00:01:46.880 Express SSD, but you'd be wrong for a 00:01:44.960 --> 00:01:52.079 number of reasons. Starting with the connector at the bottom. This is a PCI 00:01:50.079 --> 00:01:58.960 connector and this interface capped out at just 133 megabytes per second. That 00:01:55.840 --> 00:02:01.840 is even less than SATA 1 at 150 00:01:58.960 --> 00:02:07.200 megabytes a second. But that wasn't even its main problem. PCI is what's called a 00:02:04.880 --> 00:02:14.400 shared bus, which means that every device on it has to fight over that 133 00:02:11.200 --> 00:02:16.879 megabytes a second. And thing is, many 00:02:14.400 --> 00:02:21.680 older motherboard designs not only had their PCI slots hooked up to it, they 00:02:19.280 --> 00:02:26.800 even had onboard devices like their built-in sound card and network 00:02:23.440 --> 00:02:30.160 interface running off of it already. So, 00:02:26.800 --> 00:02:33.280 this slot was only suitable to use for 00:02:30.160 --> 00:02:35.280 power. It's also got a SATA port, so 00:02:33.280 --> 00:02:41.040 that's how it actually communicates with the rest of your system. Unlike PCI, 00:02:38.239 --> 00:02:46.640 SATA ports are not forced to share the same bus. So that means full performance 00:02:43.280 --> 00:02:48.239 to this port all the time. Well, at 00:02:46.640 --> 00:02:51.599 least as long as you don't run into a bottleneck elsewhere in your system, 00:02:50.080 --> 00:02:55.760 like on the link between the south bridge and the north bridge. I remember 00:02:53.760 --> 00:03:00.480 people used to hit those all the time when they were running multiple drives 00:02:57.440 --> 00:03:03.760 in RAID zero. Now, let's talk about this 00:03:00.480 --> 00:03:06.319 chip right here. This XYLink's FPGA 00:03:03.760 --> 00:03:12.080 probably represents a big chunk of the cost of the IRAM. FPGAAS offer 00:03:09.440 --> 00:03:18.640 incredible flexibility, allowing one to basically manufacture a processor in 00:03:15.200 --> 00:03:22.000 software, then program that processor 00:03:18.640 --> 00:03:24.640 into a blank one. Now, the cost is that 00:03:22.000 --> 00:03:30.159 you lose some efficiency and your per unit costs go way up. If this was a high 00:03:28.159 --> 00:03:34.400 volume product, they'd create custom silicon for it. But Gigabyte, even at 00:03:32.879 --> 00:03:40.640 the time, admitted that their first production run was a measly 1,000 units 00:03:37.360 --> 00:03:42.720 for the entire world. So clearly custom 00:03:40.640 --> 00:03:47.200 silicon wasn't going to be a problem. Let's talk about why they needed this 00:03:44.799 --> 00:03:52.000 processor, though. Fundamentally, the principle of storing data is the same 00:03:49.360 --> 00:03:57.760 between RAM and hard drives. You got your ones and you got your zeros. But 00:03:54.720 --> 00:04:00.959 the way the binary data is organized on 00:03:57.760 --> 00:04:03.439 the device is completely different. A 00:04:00.959 --> 00:04:08.080 hard drive uses a physical spinning platter covered in circular tracks that 00:04:06.000 --> 00:04:12.400 are divided up into sectors. So when your operating system asks for a 00:04:10.000 --> 00:04:17.040 particular bit, it tells your hard drive where to move the read head to grab the 00:04:14.560 --> 00:04:23.759 information it needs as it goes whipping by at 7200 revs per minute. By contrast, 00:04:20.720 --> 00:04:27.040 RAM stores bits, so zeros and ones, in 00:04:23.759 --> 00:04:29.759 cells that are laid out in a giant grid 00:04:27.040 --> 00:04:34.000 and that can be instantly accessed as long as you know the intersecting row 00:04:32.000 --> 00:04:38.240 and column, which is known as an address. Well, that's exactly the 00:04:36.560 --> 00:04:43.600 performance advantage that we're after here. But it also means that you can't 00:04:40.880 --> 00:04:50.000 just treat this thing exactly the way that you would a hard drive. So, our 00:04:46.960 --> 00:04:52.000 FPGA is programmed to act as a 00:04:50.000 --> 00:04:57.199 translation layer between our SATA interface here and the memory controller 00:04:54.639 --> 00:05:04.000 that's built into it. Now, let's look at the memory slots. These are operating at 00:04:59.919 --> 00:05:05.840 DDR1200 MHz. And while modern memory 00:05:04.000 --> 00:05:10.800 modules and their controllers would operate at much higher frequencies, even 00:05:08.400 --> 00:05:14.960 with the inefficiency of using SATA instead of NVMe like we would for any 00:05:12.560 --> 00:05:22.560 solid state drive today, that is a theoretical limit of 1.6 GB per second 00:05:19.680 --> 00:05:29.840 between our FPGA and our RAM here. That's as fast as a decent modern SSD 00:05:25.280 --> 00:05:32.240 and way faster than a hard drive. Well, 00:05:29.840 --> 00:05:35.280 SATA bottleneck notwithstanding. So, let's get the rest of these RAM sticks 00:05:33.680 --> 00:05:39.600 on here. Let's fire this thing up. All right. So, as you guys probably saw, I 00:05:37.520 --> 00:05:45.199 already had Windows fired up on this thing before we started. So, okay, just 00:05:43.120 --> 00:05:50.560 going to go ahead put all my RAM back into place. Let's throw this on the one 00:05:48.560 --> 00:05:55.520 accessible PCI slot we have on this board 00:05:52.880 --> 00:06:01.039 and fire it up. Well, that's curious. It's empty. No 00:05:58.560 --> 00:06:06.400 bootable device. So, our Gigabyte IRAM is showing up as 00:06:03.280 --> 00:06:08.880 our first boot device, 00:06:06.400 --> 00:06:15.759 but we failed to boot. My device that was full not half an hour 00:06:11.919 --> 00:06:20.000 ago is now empty. And there in lies the 00:06:15.759 --> 00:06:22.960 Achilles heel of the Gigabyte RAM. DRAM 00:06:20.000 --> 00:06:28.080 is what's known as volatile memory. So each cell on a memory chip, by the way, 00:06:25.520 --> 00:06:33.360 if you like this shirt, ltstore.com, is made up of a transistor and a capacitor. 00:06:30.960 --> 00:06:38.720 The transistor acts as a switch, and the capacitor either gets filled up with 00:06:35.520 --> 00:06:42.160 electrons, that's a one, or emptied of 00:06:38.720 --> 00:06:45.759 electrons, that's a zero. The problem is 00:06:42.160 --> 00:06:48.720 that if left alone with no power, all of 00:06:45.759 --> 00:06:56.319 the cells gradually lose their charge and zero out. No data left means no 00:06:53.360 --> 00:07:01.360 Windows boot. So, we should probably get Windows reinstalled. 00:06:58.319 --> 00:07:04.639 Going to use my super definitely genuine 00:07:01.360 --> 00:07:05.919 Windows XP CD here. Actually, that's not 00:07:04.639 --> 00:07:11.319 probably anymore. I have checked this. It does in fact have Windows XP on it. 00:07:07.840 --> 00:07:11.319 Can fix that. 00:07:20.319 --> 00:07:29.199 All right. So, here it is. Our whopping 4 GB of unpartitioned space. Now, you 00:07:26.319 --> 00:07:33.840 might wonder, why are you using Windows XP, Linus? And the reason is that it's 00:07:32.160 --> 00:07:39.199 actually the most recent version of Windows that will run on a 4 gig drive. 00:07:37.440 --> 00:07:44.160 I think I should even have about a gig left over when we're done here. All 00:07:41.520 --> 00:07:49.280 right. So, while that fires up, let's talk about what happened there. 00:07:47.440 --> 00:07:54.400 This was the first thing Brandon asked me about. This weird battery case 00:07:52.319 --> 00:08:01.199 looking thing over on the far right of the card. That is exactly what it is. 00:07:58.240 --> 00:08:06.800 And what it does is it takes the power that's being given to the card through 00:08:03.039 --> 00:08:10.000 the PCI slot and stores a little bit of 00:08:06.800 --> 00:08:12.560 it so that in the event that I turn off 00:08:10.000 --> 00:08:16.879 my computer, my operating system doesn't immediately 00:08:14.240 --> 00:08:21.919 disappear and I need to completely reinstall it and all my boot drive 00:08:19.440 --> 00:08:25.680 applications all over again. Now, it wasn't a perfect solution, and in fact, 00:08:23.680 --> 00:08:31.440 Gigabyte only rated it for, I think it was either 16 or 18 hours of power loss, 00:08:28.800 --> 00:08:36.479 but it did mean that in the event of an accidental power supply unplug or a 00:08:34.000 --> 00:08:41.599 short power outage, you weren't stuck starting fresh in Windows XP, you know, 00:08:39.200 --> 00:08:45.760 changing your uh I was always partial to the the chess pieces picture, you know, 00:08:43.760 --> 00:08:51.120 and all that good stuff. So, it's funny. I had actually intended to do some, you 00:08:49.120 --> 00:08:56.880 know, comparative benchmark testing between our IRAM and this old hard drive 00:08:54.240 --> 00:09:00.240 here. But what I realized is that once you've got the operating system 00:08:58.160 --> 00:09:07.360 installed on the IRAM, you've only got about I don't know 800 mgab of space 00:09:04.959 --> 00:09:12.880 left over. So there it's not really a whole lot to work with. Oh wow. And I 00:09:09.760 --> 00:09:15.200 just installed one game and it hit 365. 00:09:12.880 --> 00:09:19.600 Problem number two is even if we did want to do game loading times, for 00:09:16.800 --> 00:09:22.640 example, there's no Windows XP driver for the Titan XP graphics card. I have 00:09:21.360 --> 00:09:25.760 to go get another graphics card. You know what? They benchmarked it to death 00:09:24.399 --> 00:09:30.480 back when it was released. It's a little bit faster in terms of real world 00:09:27.920 --> 00:09:35.440 performance and a lot faster in terms of synthetic performance, particularly when 00:09:32.160 --> 00:09:38.959 it comes to random reads and writes. So 00:09:35.440 --> 00:09:42.399 you can see here we're at a right steady 00:09:38.959 --> 00:09:44.000 125 or so megabytes per second. Unlike a 00:09:42.399 --> 00:09:48.399 hard drive which would start at the outer edge of the platter like 00:09:45.920 --> 00:09:52.000 physically and then taper off in terms of performance as it makes its way to 00:09:50.240 --> 00:09:57.440 the data that's stored closer to the center. 00:09:54.160 --> 00:10:01.760 Not so here. And this is where things 00:09:57.440 --> 00:10:04.320 get really crazy. That access time 00:10:01.760 --> 00:10:09.440 literally 0.0c 0 milliseconds. 00:10:06.560 --> 00:10:13.760 Hard drives would be all over the place in here. 00:10:12.320 --> 00:10:19.680 Now, I'm just going to go ahead and restart. Oh, no. I turned it off. 00:10:17.200 --> 00:10:23.760 Actually, that's okay because as long as the power supply still has power, you 00:10:22.240 --> 00:10:28.640 can actually see there's an indicator LED here on the card. It's getting 00:10:26.399 --> 00:10:35.040 standby power off the slot and that's keeping the memory cells constantly 00:10:31.360 --> 00:10:37.760 refreshed so that they don't drain. But 00:10:35.040 --> 00:10:43.040 if I were to do this, it's gone. So, cool things about the 00:10:41.519 --> 00:10:47.760 IRAM. Really, really fast in spite of its 00:10:46.000 --> 00:10:52.640 interface bottleneck, especially compared to hardware that existed at the 00:10:50.000 --> 00:10:58.800 time. But its downfalls were of course the aforementioned interface bottleneck, 00:10:55.040 --> 00:11:01.360 the cludgy power loss resiliency, and 00:10:58.800 --> 00:11:05.360 the poor capacity. Because, you know, for sky's is the limit at any cost 00:11:03.680 --> 00:11:10.000 enthusiasts, something like this honestly doesn't 00:11:07.040 --> 00:11:16.560 even seem that crazy except that it was rated for 8 GB of maximum capacity. And 00:11:14.399 --> 00:11:22.000 practically speaking, there was actually no way to achieve that. Unbuffered DDR1 00:11:19.440 --> 00:11:28.880 DIMs only went up to 1 gig per stick. And while ECC server DDR1 memory modules 00:11:25.839 --> 00:11:31.600 were made in 2 gig capacities, this card 00:11:28.880 --> 00:11:35.839 has no support for ECC memory. And then the final nail in the coffin was, of 00:11:33.200 --> 00:11:40.640 course, that it was really expensive. If you had the memory lying around, and if 00:11:38.399 --> 00:11:46.399 you were way too lazy to just flip it on Craigslist, you could use that memory. 00:11:43.040 --> 00:11:50.320 But otherwise, at around $90 per stick 00:11:46.399 --> 00:11:52.880 at the time, you were looking at $500 to 00:11:50.320 --> 00:11:58.480 fully load this thing up. That is pretty steep for a five or at best 10second 00:11:56.240 --> 00:12:02.640 boot time or game loading time advantage. So, with the high cost and 00:12:00.560 --> 00:12:06.639 inconvenience as major factors, Gigabyte probably didn't make many more than the 00:12:04.240 --> 00:12:10.720 first thousand of these. But the IRAM was definitely a sign of the direction 00:12:09.040 --> 00:12:16.320 that computer storage was heading. In fact, Intel's Optane technology is a low 00:12:13.600 --> 00:12:20.399 latency solidstate storage tech that can now be found on memory modules just 00:12:19.040 --> 00:12:26.000 like, you know, they look just like these ones and used as a layer that goes 00:12:23.200 --> 00:12:31.680 in between conventional RAM and SSDs and servers with the added bonus of Optane 00:12:28.320 --> 00:12:33.360 being nonvolatile like an SSD. So, if 00:12:31.680 --> 00:12:36.959 you guys liked this video, maybe check out our explainer on Optane at the link 00:12:35.360 --> 00:12:41.200 below. Make sure you're subscribed so you don't miss our upcoming video 00:12:38.560 --> 00:12:44.000 checking out some cool SSDs from Liquid. Thanks for watching, guys. I hope you 00:12:42.480 --> 00:12:48.320 enjoyed this look at a really fascinating piece of retro technology as 00:12:45.920 --> 00:12:51.720 much as I did. I will see you guys in the next video. 00:12:59.920 --> 00:13:02.920 Colton. 00:13:12.240 --> 00:13:16.040 What? What the hell?