PC Perspective Podcast #520 – 11/01/18
Join us this week for discussion on the new Threadripper Processors, BPX Pro SSDs, and more!
You can subscribe to us through iTunes and you can still access it directly through the RSS page HERE.
The URL for the podcast is: https://pcper.com/podcast – Share with your friends!
- iTunes – Subscribe to the podcast directly through the iTunes Store (audio only)
- Google Play – Subscribe to our audio podcast directly through Google Play!
- RSS – Subscribe through your regular RSS reader (audio only)
- MP3 – Direct download link to the MP3 file
Hosts: Jim Tanous, Jeremy Hellstrom, Josh Walrath, Allyn Malventano, and Ken Addison
Peanut Gallery: Alex Lustenberg
Program length: 1:26:22
Podcast topics of discussion:
-
Week in Review:
-
News items of interest:
-
1:08:45 Vega claims the new MacBook Pro
-
Picks of the Week:
-
Jeremy: Total Chaos – A spooky DOOM 2 mod for Halloween
-
Josh: New Center Speaker!
-
Jim: Spacedock
-
the new guy is less fun, i
the new guy is less fun, i get that he doesn’t have the power like ryan did, and he takes it very seriously, but then this becomes less of a informative friend’s gathering that i’m listening into, and more of a BBC new radio…
RELAX MAN!! HAVE SOME FUN!! it allows us(the listener) to have more fun listening to the podcast!!
the reason why i listen to the podcast inspite reading the webpage is because it’s fun, and now it’s less fun.
Hope my opinion helps!
I’m sorry.
I'm sorry.
No! no! don’t take it
No! no! don’t take it wrong!
it’s nice to see new people join! i just hope you’ll fit in much more in due time and enjoy the bants!!
If you want fun and
If you want fun and entertainment get netflix but I’d rather websites be more informative and less 5:00 News.
It’s bad enough that the majority of the online technology Press has KOed accuracy in the name of entertainment and affiliate marketing.
And the majority of readers can’t tell the difference between Ryzen the Marketing Name for a line of AMD CPUs And Zen the name that was given to AMD’s Newest x86 ISA based CPU microarchitecture. Or the readers that criticize the Vega GPU microarchitecture because they are misinformed about that single large first “Vega 10” DIE tapeout that was laid out with not enough available ROPs(64 Max on Vega 10’s Tapeout) for sufficient Pixel Fill Rates against Nvidia’s GP102 Base Die Tapeout that offers 96 ROP(Max) with Nvidia enabling 88 out of 96 ROPs on that GP102 Die Bin and creating the GTX 1080Ti with the highest pixel fill rates of any consumer gaming GPU SKU(At That Time).
So a single Die Tapeout variant is really not the way to judge a GPU’s microarchitecture and look at tha Semi-Custom “Vega M” Die Tapeout and AMD doubled the Amount of ROPs relative to Shader cores on that semi-custom product. So AMD could damn well increase the ROP counts on a new Vega Die tapeout above the only 64 ROP’s max that the first Vega 10 Big Die had available and compete just fine with the GTX 1080Ti’s 88 ROPs in pixel fill rates.
So One Vega 10(Big Die) Tapeout may be based on the Vega GPU microarchitecture but it’s not the GPU microarchitecture’s fault that AMD did not include more Available ROPs to competet better with GP102 instead of only GP104(64 ROPs max also).
If Vega 10 had sufficient numbers of ROPs included in the first Vega 10(Big Die) tapeout then the pixel fill rates would have been more competative with the GP102 based GTX 1080Ti’s 88 out of 96 enabled ROPs. And higher pixel fill rates via more ROPs directly translates into more FPS in gaming workloads. And it’s not the fault of the Vega GPU microarchitecture that Vega 10(Big Die) has insufficient ROPs. And AMD was more interested in that Vega 10(Big Die) tapeout for the Professional Compute/AI markets mosty anyways and AMD is still going to be more Interested in Vega 20 for Professional compute/AI related workloads.
AMD intentionally designed Vega 10(Big Die) to only compete with the GP104 based GTX1080 in gaming workloads and any extra Vega 10(Big Die) space was used for Shader Cores for Compute/AI workloads mostly. It’s just AMD’s good luck with Vega 10(Big Die) that it was so popular with the coin miners beacuse of that excess shader core count(4096) on the Full Vega 10(Big Die) tapeout.
In Fact Vega 56 has the exact same numbers of Shader Cores and TMUs as the GTX1080Ti but Vega 56(binned from that same Vega 10 Big Die) lacks the ROP numbers and Vega 10’s 64 ROPs max ment that the First Vega 10(Big Die) Tapeout could never Best GP102/GTX 1080TI in gaming FPS.
And AMD really is not going to be competing in Flagship Gaming with Nvidia. And even Nvidia does not want to spend any resources on consumer Flagship Gaming SKUs either because the markups are just not there in the consumer market. Any Flagship Gaming GPUs from Nvidia are sure to come via binned Dies anyways as GP102 was always ment for Quadros first and formost but Nvidia has to have a consumer market in which to sell those defective GPU DIEs that can’t be made into Quadros.
Look at the Turing generation of GPUs and even the TU104 top end Die Bin is for Quadro with the consumer getting a Binned TU104 Based RTX1080. The **102 has always been for Quadro first and formost but now under TU104 Quadro gets the top bin and consumer gaming gets a lower bin. The RTX
2070 is now based from the TU106 Base Die Tapeout so Nvidia really wants the majority of its many Turing Base Die Tapeouts targeting Quadro SKUs at many times the markup of the consumer/gaming SKUs.
AMD’s management is not going to target Flagship GPUs because the revenues to investment figures to not make sound business sense. So AMD’s Vega 20 is not for gaming and Navi is only going to be targeting mainstream gaming where the overall unit sales numbers justify AMD’s investment. AMD has so much less of the consumer gaming market that AMD would rather invest in the Epyc and Vega 20 Compute/AI market where the markups charged justify AMD’s investments in R&D and years of hard work.
So look for Flagship GPUs to become more costly and do not look for AMD to introduce any Flagship offerings until Nvidia has driven the cost of Flagship gaming so high that AMD can come in and undercut Nvidia and AMD can still charge $700+ for any new Flagship GPU. AMD is going to let Nvidia increase the gaming GPU market’s ASP for Flagsip GPUs past $800-$1000 and only then will AMD even think about entering the Flagship Gaming Market again.
You will gladly pay AMD $700-$800 for its Flagship because Nvidia is Charging $800-$1200 for its Turing high end and flagship GPUs. So in AMD’s more sound Business Judgment under Lisa Su that’s going to be sit out Flagship Gaming until Nvidia does the dirty work of driving the High End/Flagship GPU market’s GPU-ASP even higher towards the $1000+ range so AMD can come in and undercut that higher pricing and AMD can still earn revenues instead of barely breaking even or losing money.
TSMC’s 7nm node has got to be very costly for any customer and Samsung single handedly ruined the HBM2 market by trying to overcharge there and driving investment away from HBM2 for the entire market towards GDDR6. AMD will probably be still using HBM2, owing to Vega’s, and Navi’s, HBCC/HBC IP that’s a big selling proint in the professional graphics market where the markups justify the cost. But AMD really needs higher GPU-ASPs also because Memory and 7nm is very costly which is why Navi will be mostly targeting mainstream gaming mostly. Hopefully That Vega/Navi HBCC/HBC IP can make use of GDDR5x/6 as High Bandwidth Cache also but maybe by now SK Hynix’s HBM2 is more available at lower cost than what Samsung is charging for HBM2.
(same person from
(same person from before)
Wow, that was long, and tbh, I knew, from videocardz (admittedly, not the most reliable)
I mean, personally, the reason why I still listen to the podcast despite already read most of the articles from the website is for the opinions and the fun.
The articles are usually written very professionally, and therefore it’s mostly factual, whereas the podcast is more informal, and contains opinions from the writer, which is nice.
And then there’s the fun, if the podcast was super formal, then there wouldn’t be hot dog down the hallway, and I wouldn’t be buying the dumb t-shirt and no one gets it.
Long and winding but really
Long and winding but really is all about affiliate marketing and less about the whole truth. Some online sources do better jobs than others but they all fear the loss of free review samples.
Then we get the WWF wrestling match between Nvidia the 800LB Gorilla in Market Cap against the 90LB Market Cap weakling AMD. And that carefully choreographed and managed buy the online press grudge match. And any Techncal discussions concerning Nvidia’s ROP counts and AMD’s ROP counts and the pixel fill rates in the GPixels/s are rarely discussed and most likely verboden by the review manauals dictated by the GPU makers on anyone that wants to continue to receive Free Review samples.
Oh Know let’s blame it on a name, the name being a name for a GPU microarchitecture/ISA that’s really more to do with how a grouping of Shader core executes a GPU ISA. And let’s really Blame it on the Vega 10(Big Die) tapeout that was intentionally engineered to be more tuned for Compute/AI workloads than those gaming and FPS workloads that are directly affected by those Numbers of ROPs and the total GPixel/s Pixel fill rates on gaming workloads. Vega 10(Big Die) had only 64 ROPs available Max but an excess of shader cores Relative to Nvidia’s more gaing oriented GPUs.
Now the solution for any Vega microarchitecture based GPU is for gamers to raise a few hundred million dollars to give to AMD to pay for the Mask Sets and Tapeout costs for a new Gaming Oriented Vega Die Tapeout with double the ROPs available and watch that Vega/Whatever Number Die compete with any of Nvidia’s Pascal gaming SKUs.
We can reduce the numbers of Shader cores on the New Vega Die Tapeout while increasing the ROPs and maybe be tweaking some L2 GPU cache sizes. Folks Vega 56 has the same numbers of TMUs and Shader cores as the GTX 1080Ti so 88 ROPs would match that SKU but let’s match GP102’s full complement of available ROP’s at 96 and Re-Tapeout a new Vega Base Die tapeout with Loads of ROPs and a little bit more shader cores to match Nvidia’s latest Turing Cores in Shader core counts, TMUs and ROPs. Nvidia’s TU102 Base Die Tapeout still only has 96 ROPs max and that’s the same as GP102’s full ccomplement of ROPs.
This all well and good but really AMD is not going to spend any of its still limited resources on any single gaming oriented Flagship GPU tapeout as AMD’s lacks the funds still for such and undertaking. The overall market for Flagship Gaming GPUs and total limited Revenues provided by any Flagship GPU production dictates that the only cost effective way to produce any Flagship GPU Die for gaming is to produce that Die for the Professional markets first where the markups are sufficient to cover the R&D and Taprout/Mask-Sets and Fab costs and then Take the lowest Bin of that Die, that’s not qualified for professional usage, and use that Bottom Binned Die for a Flagship Gmaing GPU sold on the consumer market at almost no significant profit.
And Folks that’s what the GTX 1080Ti Die is binned from in the form of the GP102 Base Die Tapeout that generates plenty of Quadro Qualified DIEs and a very handsome markup to pay the R&D, Tapeout, Mask-Set, Hundred Million Dollar+, costs that it takes per Big Die Tapeout to produce. And AMD’s Vega 10(Big Die) Tapeout is paid for the same way by AMD, via those Radeon Pro WX 9100s and Radeon Instinct MI25’s sales. Mining Sales for Vega 10(Big Die) was just the lucky break for AMD that helped cover the costs and produce some revenues/small profit.
AMD with Navi will be aiming for the larger revenue producing, as far as consumer/gaimng market sales go, mainstream Gaming GPU market where the largest unit sales and total consumer market revenues occur. And Vega 20 is a 7nm tapeout of Vega 10 with some AI oriented instructions added to Vega GPU microarchitecture to add some more AI oriented abilities to Vega 20’s Vega ISA. I do not think that AMD will be adding any extra ROPs to Vega 20 but I could be wrong.
Navi is not going to be made up of a bunch on smaller GPU die chiplets as was orginially stated but that does not mean that any dual Navi Die on a single PCIe card SKUs could not be made wired to each other across the Card’s PCB Via that xGMI(Infinity Fabric Based) multi-GPU interface that’s a fair bit better and more cache coherent in a similar manner to Nvidia’s NVLink in scope. So even if Navi is still a single monolithic larger GPU DIe based design there is nothing stopping AMD from creating any Dual Navi GPU DIEs/Single PCIe card SKUs and using xGMI(IF Based) to wire up the Navi DIEs to look like a larger logical GPU to the OS/Software. Ditto for Nvidia and using NVLink to do the same.
Both NVLink and Infinity Fabric have inherent in their respective designs built in scalaibility in mind with extra GPU to GPU cache coherency Protocol Feratures that can actually wire up 2 GPU dies in a more logical manner. And that includes wireing up 2 GPUs to act and to appear to software as one logical larger GPU. So with the arrival of the NVLink and Infinity Fabric IP look for GPU makers to begin offering more professional oriented Dual GPU DIEs/One PCIe card Variants and maybe AMD will begin offering Dual Die Navi GPU DIEs/One PICe card offerings in the future.
That Many Smaller GPU Die/Chiplets IP still has some packaging kinks to be worked out along with some Physical data fabric verification to be certified yet. But as far as Dual GPU DIEs/One PCIe card both NVLink and Infinity Fabric are mostly ready for that and Dual GPU Gaming. Nvidia is not likely to fully enable it total NVLink Featuer Set on its Consumer Branded GPUs as Nvidia will segement that to its Quadro/Tesla lines. But AMD will be more likely to be the one to make use of Infinity Fabric based IP with Mainstream Consumer GPU Dual Navi DIEs/one PCIe card offerings, if only because AMD lacks the funds to afford too many monolithic DIE Tapeout Variants even with Navi as a midsize GPU DIE tapeout.
So scaling of 2 larger Navi GPU DIEs on a PCIe Card/Module via the Infinity Fabric will be the easiest obtainable first step Towards Many Smaller GPU Die/Chiplets. Ditto for Nvidia and NVLink and 2 larger GPU DIEs/One PCIe card also as a first step. Just Don’t look for Nvidia to enable all if its NVLink feature set on Coneumer Gaming Cards.
No worries, I meant the
No worries, I meant the apology sincerely. It was not my plan to end up as host, and this is not my natural strength. But I love these guys and want to do what I can to help things keep moving in a way that the audience finds enjoyable.
I appreciate your feedback and hope you'll continue to suggest areas in which I can improve (or attempt to, anyway!).
Feedback is always nice, but
Feedback is always nice, but I think it’s fair to give Jim a chance to settle in. He’s not Ryan, and he is probably a bit nervous… I’m sure it will get back to a similar feel after a few more shoes, but the dynamic has changed and sometimes change can be refreshing.
Has anyone done similar
Has anyone done similar ramdisk IOPS test under linux?
Haven’t seen any in a cursory
Haven't seen any in a cursory check, but that seems like it's right up the alley of the folks over at Phoronix.