|
BangersInMyKnickers posted:Those run ~250W PSUs right? 1.5 vs 1.35v on the dimms isn't going to be an appreciable difference on a desktop system, maybe a watt or two. It's really more for battery savings where you're trying to shave every last watt. The GPU and CPU are going to be your big draws, if you're seeing voltage dips with the config you're so far past the capacity of your PSU that ddr3l isn't going to help anything. I worked on a design once with selectable DDR voltage at runtime via a GPIO pin; it would literally toggle a pin on the DDR voltage regulator to drop to 1.35 V from 1.5 V, defaulting to booting up to 1.5 V each time. It was a deeply embedded system, so all the power savings in the world were of interest. The DDR dies themselves basically get binned if they can run down at 1.35 V as well as 1.5 V. What was not of interest were the seeming Heisenbugs that would show up due to software bugs, and forgetfulness in setting it to the proper voltage. The pin was driven by FPGA logic as well, so you'd have to boot up at 1.5 V, then drop to 1.35 V once the logic loaded, and then hope that the training parameters from that particular boot worked out OK.
|
# ? Jan 6, 2020 22:25 |
|
|
# ? May 13, 2024 21:01 |
|
movax posted:Next Unit of Computing indeed...wonder what (if) they will rebrand it too. The Unit of Computing Formerly Known as Next Unit of Computing
|
# ? Jan 6, 2020 23:45 |
|
"huh how could something be more horrific than 'touches MRC at all' i wonder"movax posted:The pin was driven by FPGA logic as well
|
# ? Jan 7, 2020 00:16 |
|
foldable screen laptops are dumb and this presentation is boring and i want a beer and a cookie!!!!
|
# ? Jan 7, 2020 02:00 |
|
Tiger Lake running is a good sign, but Intel showed 10nm Cannonlake running years before we ever saw anything out of it in retail. Disappointed we didn't get the Comet Lake-S launch when the entire board and CPU lineup is already leaked to the public. Disappointing.
|
# ? Jan 7, 2020 02:13 |
|
Cygni posted:Tiger Lake running is a good sign, but Intel showed 10nm Cannonlake running years before we ever saw anything out of it in retail. Disappointed we didn't get the Comet Lake-S launch when the entire board and CPU lineup is already leaked to the public. Yeah. New chips or why are you even here...
|
# ? Jan 7, 2020 02:27 |
|
movax posted:What was not of interest were the seeming Heisenbugs that would show up due to software bugs, and forgetfulness in setting it to the proper voltage. The pin was driven by FPGA logic as well, so you'd have to boot up at 1.5 V, then drop to 1.35 V once the logic loaded, and then hope that the training parameters from that particular boot worked out OK. I am confuse, why didn’t this system retrain at the final voltage setting, how could anyone think it’s safe to change one of the holy trinity (PVT) without a retrain??? Wait let me guess this fpga was controlling the voltage of dram connected to some other SoC with no easy/documented way of resetting the dram controller, or by then poo poo was booted and you couldn’t take dram away, or some equivalent hilarity
|
# ? Jan 7, 2020 03:25 |
|
BobHoward posted:I am confuse, why didn’t this system retrain at the final voltage setting, how could anyone think it’s safe to change one of the holy trinity (PVT) without a retrain??? It was a Zynq, so processor and FPGA in one package (but different voltage domains, technically). FPGA pin connected to the PMIC that supplied the DDR voltage rail and that pin would toggle the PMIC between 1.35 V and 1.5 V. It was mostly SW engineers or FPGA engineers forgetting to document their interfaces correctly and thinking they had set the bit in memory that toggled the GPIO, or put the correct bitstream on the device that caused issues. The first stage bootloader also configures DRAM timing / frequencies and if there was a mismatch there between the desired target voltage and desired target speed, issues would also occur. A neat idea in concept, poorly executed. The MRC Jawn refers to is an unholy mess of tens of thousands of lines of x86 that get dropped by Intel with 0 explanation. Do never touch the Memory Reference Code.
|
# ? Jan 7, 2020 04:35 |
|
I should not be surprised they haven’t done it but it would be cool if Intel put out a Zync competitor, one of the smaller Alteras with an Atom smooshed in it or something.
|
# ? Jan 7, 2020 04:53 |
|
priznat posted:I should not be surprised they haven’t done it but it would be cool if Intel put out a Zync competitor, one of the smaller Alteras with an Atom smooshed in it or something. There was Stellarton, way back before they bought Altera, but that wasn't a monolithic die, just Atom and Arria II (IIRC) dies squished together in the same package with some PCIe + GPIO routed between them. Now there is Cyclone V SoC but it got beat to market by Xilinx and I'm not sure who actually uses it / chooses it over Zynq, considering the amount of time and effort Xilinx put into supporting their Linux solution (github.com/xilinx/linux-xlnx).
|
# ? Jan 7, 2020 05:38 |
|
movax posted:There was Stellarton, way back before they bought Altera, but that wasn't a monolithic die, just Atom and Arria II (IIRC) dies squished together in the same package with some PCIe + GPIO routed between them. The Cyclone V was before the acquisition right? Gotta be, what with an ARM on there. I've put Zynqs on emulation platforms and functional validation boards before and they are awesome, love net booting them and running full linux off remote mounted file systems! Or SD Card if you are on the go. Wish we had a use case for them on the current platforms
|
# ? Jan 7, 2020 05:44 |
|
priznat posted:The Cyclone V was before the acquisition right? Gotta be, what with an ARM on there. Yeah, definitely before acquisition. Are you using a SoM for dropping Zynqs around, or do you have a core layout for a 7010 or 7020 w/ DDR and PMICs that you can drop around easily? Net booting w/ U-Boot off a tiny little QSPI and then doing everything over NFS is awesome — I love it. Back on the Intel side of things though... reading a bit of the AMD thread right before this, what's the play for Intel Xe? Most office applications using Intel CPUs w/ integrated graphics don't have a lot of pixels to push, and they already have the fixed-function video and audio hardware acceleration.
|
# ? Jan 7, 2020 06:58 |
|
movax posted:Yeah, definitely before acquisition. Are you using a SoM for dropping Zynqs around, or do you have a core layout for a 7010 or 7020 w/ DDR and PMICs that you can drop around easily? Net booting w/ U-Boot off a tiny little QSPI and then doing everything over NFS is awesome — I love it. Way back when it was a design in heavily based on the zedboard reference design even down to the display which was actually pretty useful for putting up things like host name, IP address, load avg, firmware version etc. This was at a previous company and the boards were very low run sky’s the limit cost wise so we could go hog wild, it was great. Previous to that we used Atmel AVR32s which were decent little devices but the ARMs on the Zynqs blew them away. Plus being able to directly interface with the logic that talked to the onboard DUT GPIOs was really nice, just configuring the GPIOs in Linux and they showed up as regular devices, very slick. Also for the Xe perhaps it is a way to finally kill off Matrox
|
# ? Jan 7, 2020 07:53 |
|
Hi I am excited for what tiger lake will do for Ultrabooks and probably tablets Where can I go to learn more about whatever info there is on these chips
|
# ? Jan 10, 2020 04:39 |
|
My general impression of Zen2's various boost mechanisms is that they maximize performance enough that there's often very little room for squeezing out more via a manual overclock Does a similar situation exist with Intel's chips, either on their current generation or what we can expect from Comet Lake?
|
# ? Jan 10, 2020 04:54 |
|
No, and I don't think anybody outside of Intel knows, respectively.
|
# ? Jan 10, 2020 04:59 |
|
gradenko_2000 posted:My general impression of Zen2's various boost mechanisms is that they maximize performance enough that there's often very little room for squeezing out more via a manual overclock AMDs newer chips were designed from the group up and have many (apparently thousands) of sensors in the chip that various monitor parameters related to silicon reliability. [1] That’s how Zen can boost so high, close to the stability limits but without failing or becoming unreliable. Intel’s current architectures don’t seem to have similar technology (at least not the ones directly based on Skylake, that includes Comet Lake) but even they are binning closer to the limit with something called Thermal Velocity Boost (TVB) which looks at power, time, load and temperature to extend the boost range. It’s not nearly as fine sophisticated as AMDs solution, so it has to work with safety margins that can be used for overclocking. Over time overclocking margins are shrinking or going away entirely unless the manufacturers purposefully bin down their CPUs again. The Ryzen 1600AF is one of those cases. [1] https://www.anandtech.com/show/11170/the-amd-zen-and-ryzen-7-review-a-deep-dive-on-1800x-1700x-and-1700/11
|
# ? Jan 10, 2020 08:20 |
|
Is that related to the max turbo 3.0 stuff I see on next gen stuff ?
|
# ? Jan 10, 2020 10:50 |
|
Statutory Ape posted:Hi I am excited for what tiger lake will do for Ultrabooks and probably tablets Uhhh, Computex probably? Ice Lake only launched a couple months ago so what little they teased in CES is all we're getting for a while.
|
# ? Jan 10, 2020 11:01 |
|
edit: probably not nevermind
Paul MaudDib fucked around with this message at 20:10 on Jan 10, 2020 |
# ? Jan 10, 2020 20:06 |
|
Intel is going to totally dominate dedicated GPUs in 2020.
|
# ? Jan 12, 2020 05:03 |
|
Crotch Fruit posted:Intel is going to totally dominate dedicated GPUs in 2020. Maybe they'll sell it for $29.99
|
# ? Jan 12, 2020 11:15 |
|
Isn't that card basically a stand in for where they want their integrated graphics to be for this generation? It doesn't have any PCI-E power plugs, so it can't be pulling more than 75 watts. This doesn't seem like a complete failure. If they can integrate it with reasonable power draw, it'll be good for the low end gamer.
|
# ? Jan 12, 2020 11:39 |
|
LRADIKAL posted:Isn't that card basically a stand in for where they want their integrated graphics to be for this generation? It doesn't have any PCI-E power plugs, so it can't be pulling more than 75 watts. This doesn't seem like a complete failure. If they can integrate it with reasonable power draw, it'll be good for the low end gamer. That is my assumption as well. Doing this as a separate unit makes sense to get telemetry etc probably. Develop drivers and watch how this poo poo interacts w poo poo. These will be cheap or never seen in that form by consumers I don't think they're trying to get into the traditional dgpu sphere. Probably just the low end and compute type stuff I would guess, as a rank amateur
|
# ? Jan 12, 2020 11:52 |
|
This card is also for game engine developers and the like to test and write code against. They have to work OK if and when they start integrating them into laptops or whatever.
|
# ? Jan 12, 2020 12:03 |
|
I could also see it filling the same niche for DOTA/LoL cards full because I feel like there are a ton of lower end cards that get shipped off overseas that fill literally that niche.
|
# ? Jan 12, 2020 13:37 |
|
Yeah, the segment filled by the NVidia 1650 and below is surprisingly large, and anything that can get even to 1050 levels would be able to capably play a lot of net cafe and other lower end games that, frankly, a whole fuckton of people worldwide dump a lot of play time into.
|
# ? Jan 12, 2020 14:28 |
|
I have been saying for years that not having iGPUs from either camp that can routinely do what a 750ti can is a goddamn travesty. Maybe we'll finally see that in the next two or three years.
|
# ? Jan 12, 2020 15:27 |
|
SwissArmyDruid posted:I have been saying for years that not having iGPUs from either camp that can routinely do what a 750ti can is a goddamn travesty. Maybe we'll finally see that in the next two or three years. Is the consumer desktop segment the least important segment now? I assume its the high end stuff and the mobile being the big things people would care about at this point, the stuff that drives innovation right? I assume the fact that I use a desktop the same way my dad did when he was 33 puts me squarely in the minority (not in this forum/sub forum)
|
# ? Jan 12, 2020 15:56 |
|
SwissArmyDruid posted:I have been saying for years that not having iGPUs from either camp that can routinely do what a 750ti can is a goddamn travesty. Maybe we'll finally see that in the next two or three years. 750 Ti has its own memory bandwidth and 75W of power and cooling. Maybe when DDR5 exists AMD can chiplet out something. I'd expect any huge iGPU chips to be hot as hell too, 150W+.
|
# ? Jan 12, 2020 17:37 |
|
Can't imagine 750ti GPU would use anywhere near that on a modern process
|
# ? Jan 12, 2020 17:51 |
|
DrDork posted:Yeah, the segment filled by the NVidia 1650 and below is surprisingly large, and anything that can get even to 1050 levels would be able to capably play a lot of net cafe and other lower end games that, frankly, a whole fuckton of people worldwide dump a lot of play time into. Die size on the smaller chips tends to be pad limited so a new product to fill the niche between iGPUs and a 1650 just isn't financially viable when discounted older products and the used market exists.
|
# ? Jan 12, 2020 18:03 |
|
It absolutely shouldn't. For comparison's sake, a 750ti was capable of 1.4 TFLOPs single precision. Right now, at THIS VERY MOMENT, Ice Lake G7 clocks in somewhere between 1.0 and 1.1 TFLOPs, depending on cooling and configuration. I don't know why you guys think this is some kind of unattainable goal that needs voodoo and HBM and chiplets and new memory. It's there! It's right loving there! It's so close!
|
# ? Jan 12, 2020 18:07 |
|
SwissArmyDruid posted:It absolutely shouldn't. For comparison's sake, a 750ti was capable of 1.4 TFLOPs single precision. Memory bandwidth.
|
# ? Jan 12, 2020 18:08 |
|
And I quote Paul: Paul MaudDib posted:LPDDR4 is a hell of a drug. If AMD can use it to realize what they're claiming is "59% improved performance" on their new APUs that are still using Vega-and-not-RDNA cores rewarmed, why in god's name shouldn't Intel get in on that poo poo?
|
# ? Jan 12, 2020 18:12 |
|
Raven Ridge officially supported up to ddr4 2933 which is about 23GB/s, lpddr4x 4266 does about 34GB/s or 50% more, making it real obvious where the 59% performance increase comes from. A 750ti does 86GB/s.
|
# ? Jan 12, 2020 19:22 |
|
Y'all see this thing about the 28w chip https://www.anandtech.com/show/15302/intel-28-w-ice-lake-core-i71068g7-coming-q1 I feel like a 28w chip with whatever their dgpu evolves into might be a decent type of all in one
|
# ? Jan 12, 2020 20:07 |
|
Is the IHS soldered now?
|
# ? Jan 12, 2020 20:43 |
|
Arzachel posted:Raven Ridge officially supported up to ddr4 2933 which is about 23GB/s, lpddr4x 4266 does about 34GB/s or 50% more, making it real obvious where the 59% performance increase comes from. And yet one is benched as being capable of about 10% fewer FLOPS than the other. Gee. It's almost like IPC and transistor count between a 28nm process and a 10nm process actually *means* something.
|
# ? Jan 13, 2020 08:24 |
|
|
# ? May 13, 2024 21:01 |
|
SwissArmyDruid posted:And yet one is benched as being capable of about 10% fewer FLOPS than the other. Gee. It's almost like IPC and transistor count between a 28nm process and a 10nm process actually *means* something. I messed up and the numbers should be doubled for dual channel, so it's not nearly as grim as I thought and Renoir should probably beat a 750ti as long as you're using real fancy memory Realistically, AMD's apu designs will still be heavily memory bandwidth limited until ddr5.
|
# ? Jan 13, 2020 11:00 |