Problems? See this thread at archive.org.
So this part on anti-consumer is a fundamental misunderstanding of the tech implementing SAM (Smart Access Memory). This is why I am excited for PCIe 5.0 and Intel's CXL!
Let me back up. AMD created a pre-cursor to CXL on PCIe 4.0 PHY. This is what allows for heterogeneous memory usage where the CPU can instead of using and having to duplicate in system memory just use the GPU memory on the card. AMD also is doing direct SSD access for loading into GPU memory. Intel's CXL goes beyond this to cache coherency and access across devices, which is why you need both CXL on PCIe 5.0, which AMD signed onto, and Gen Z PCIe for internode connections that are going to be revolutionary. But that is getting into server talk. Back to consumers.
AMD took their non-proprietary, open, standard for PCIe 4.0 and implemented it in their products. I was not expecting to see this consumer side this quickly (which means Nvidia and Intel will BOTH have to look at adoption of this moving forward or guarantees consumer implementation of CXL on PCIe 5.0 in the future OR could cause Intel and Nvidia to adopt, to a degree, AMD's way of doing it on PCIe 4.0, thereby allowing that to impact how CXL will be implemented in the future). This standard was known to Nvidia, and although they did not have a Zen 3 chip to optimize and implement (because not out yet), they could have asked AMD how to implement it for server side, then put that in their desktop cards without knowing if any hardware would ever support it, or ask Intel to implement AMD's solution for PCIe 4.0.
Either way, since Intel never really put out a PCIe 4.0 product, all product development has been done on AMD CPUs. In fact, AMD and Nvidia have worked together for some server offerings. So this is NOT anti-consumer. In fact, this is forcing movement of cutting edge server tech that has barely been even implemented in servers to the consumer. Look up the whitepapers on it from 2018.
AMD never restricted the implementation of this from Nvidia NOR Intel.
Edit: Found the name again: CCIX (Cache Coherent Interconnect for Accelerators; https://en.wikipedia.org/wiki/Cache_coherent_interconnect_for_accelerators). That is a known project and likely is the basis for the entire SAM implementation, as well as could work with Intel's CXL implementation.
Yes, that will be fun to see the back and forth. And then next gen, with this shot across Nvidia's bow, things can get REAL GOOD! It is like how the 290x from AMD competing caused Maxwell to drop. So definitely exciting!!!
Turns out, RT in WoW is DXR, which AMD directly supports Direct X Raytracing (DXR). From rumors, the DXR is similar to Turing's, or slightly better, but still lags Ampere.
-
AMD Ray Accelerator appear to be slower than NVIDIA RT Core in this DXR (Ray Tracing) benchmark videocardz.com
Convel, Rage Set, Robbo99999 and 3 others like this. -
-
Holy Cow... Hello guys.
Yes I'm alive and well haha. Very interesting times, I hope everyone has been well and staying safe. I've been very busy with family, great things despite this crazy pandemic. Life is good... many blessings to be grateful for..
I have a lot of reading to catch up on so it'll take a while, however....
I do have a 3080 right in front me and am eager to start having some fun testing it.
-
-
-
-
-
There are some games that showed much lower performance, sometimes being only a couple percent over the 2080 Ti, while with the 3070, you see it lose significantly to the 2080 Ti. You pick ONE GAME. ONE. Then act like that is representative when we know some games fill up the 11GB of the 2080 Ti already.
Now, I'll agree, it generally is edge cases at this moment in time. We both agree there. Because the number of titles are currently extremely limited that hit that barrier on VRAM. But that may change in the coming years, especially with AMD competing at the high end and 16GB being given all the way down to a 3070 competitor, with the 3070 only having 8GB, something low end 580s had years ago.(edit: Also, both consoles have 16GB of VRAM available, meaning that Nvidia is the limiting factor on ports, hilariously enough, whereas AMD's Big Navi cards with the 6900 XT and 6800 XT and 6800 will all support that much. Who knows the VRAM on the upcoming 6700 XT which uses Navi 22 and has 40 CUs instead of the 80 CU big brother, along with the 32CU Navi 23).
Nvidia screwed up there. It is plain and simple. They were greedy.
So, two things: 1) You are wrong that the 1080 Ti had more cuda cores than the Titan X Pascal, and it also had fewer cores than the Titan Xp, and 2) this is why looking to Quadros informs of the full die size, generally.
So, let's start with the core counts of the Titan X pascal and the 1080 Ti, then move to the second. Titan X pascal had 3584 shaders or cuda cores. That was NOT the full die. The 1080 Ti had 3584 shaders or cuda cores. That looks like the same number of cores to me.
https://www.techpowerup.com/gpu-specs/geforce-gtx-1080-ti.c2877
https://www.techpowerup.com/gpu-specs/titan-x-pascal.c2863
In fact, the 1080 Ti also cut the number of ROPs down from 96 to 88 for the 1080 Ti, along with going from 12GB to 11GB of VRAM and cutting the memory bus from 384-bit to 352-bit bus.
Now, what is the difference between those two cards then? Well, part of it is frequency of both the Cuda Cores and the memory. The Titan X Pascal had a base clock of 1417 MHz, a boost clock of 1531 MHz, and a memory clock of 1251 MHz which gave a memory bandwidth of 480.4 GB/s. The 1080 Ti had a base clock of 1481 MHz, a boost clock of 1582 MHz, and a memory clock of 1376 MHz which gave a memory bandwidth of 484.4 GB/s. So, even though it had less VRAM, it was clocked higher and had more memory bandwidth and faster memory, going from an effective 10Gbps to 11Gbps, thereby giving better performance. It's almost like how people overclock their cards to get more performance. Who would have thought higher clock speeds and faster memory would increase performance?
To be clear, TechPowerUp has the difference in performance between the two cards at 3%. That's it. Just 3%. Which makes sense with only a 50MHz boost and 4GBps memory bandwidth, while nothing outside of professional workloads could have taken advantage of that extra GB of memory at the time, although I remember AI training COULD and DID! But that was pre-tensor core Volta cards. And any extra difference between the cards can be accounted for in the driver optimizations for each line, which geforce cards are optimized for gaming whereas Titan cards have extra optimizations for professional workloads.
Let's now move onto the Titan Xp and the Quadro P6000 cards. The Titan Xp had 3840 shaders, while the P6000 had 3840 shaders. The P6000 had double the memory, at 24GB, whereas the Titan only had 12GB. The Titan Xp had a base clock of 1405 MHz, a boost clock of 1582 MHz (like the Titan X Pascal), and a memory clock of 1426 MHz which gave a memory bandwidth of 547.6 GB/s. The Quadro P6000 had a base clock of 1506 MHz, a boost clock of 1645 MHz, and a memory clock of 1127 MHz which gave a memory bandwidth of 432.8 GB/s. They likely gave the Quadro extra core clock to make up for the slow memory speed from getting ram with 2GB instead of 1GB per chip, which ran slower.
This brings me to my second point, that you look to the Quadro 6000 series to look for the non-cut-down variant of the 102 die. They charge a premium for these cards and so pick the choicest silicon off the wafers with no defects. So if you ever want to know what the full die size is, you instantly go to the Quadro 6000 entry to understand how many shaders is on the card.
If you look at the RTX 6000, there were 4608 shaders. Same on the RTX Titan. The 2080 Ti had 4352, or about 94% of the total cuda cores on a TU102 die. The RTX Titan had a higher boost clock, higher memory bandwidth, and more than double the memory of the 2080 Ti. In fact, there was no cut down between the Quadro and the RTX Titan this year, similar to the Titan Xp, whereas they DID cut the die down to make the Titan X Pascal line. Almost like Nvidia planned on releasing another die a year later to bilk Titan purchasers.
Now, if you jump forward to today, you have to start at the non-cut-down die of the Quadro lineup.
https://www.techpowerup.com/gpu-specs/rtx-a6000.c3686
As you can see, the RTX A6000 Quadro card has a total of 10752 shaders on the GA102 die. That is the full die and full core count. You cannot go above that without jumping to the 100 die manufactured at TSMC, which they haven't released a 100 die to consumers in nearly a decade. In fact, they've been making the 100 variant a high priced commodity, with the V100 costing $10,000 at launch, although the Titan V did get a cut down variant of it with fewer ROPs and less memory and mem bandwidth.
From here, we look at the 3090 core count and the 3080 core count. The 3090 has 10496 shaders and the 3080 has 8704, which is 97.6% and 81% of the full possible shader count, respectively. The 3090 has a base clock of 1395 MHz, a boost clock of 1695 MHz, and a memory clock of 1219 MHz (or 19.5 Gbps effective) which gives a memory bandwidth of 936.2 GB/s. The 3080 has a base clock of 1440 MHz, a boost clock of 1710 MHz, and a memory clock of 1188 MHz (19 Gbps effective) which gave a memory bandwidth of 760.3 GB/s.
And since the boost is higher by one 15MHz stepping, but slower memory and 17% fewer cores than the 3090, but the difference in performance at stock is only 11%, anything you do for a 3080 Ti would have to play in the ground between the two, since the 3090 is practically not cut down (it is by 2.4%, but that is to increase usable yields and would matter very little in the result). So, let's go to the rumored 3070 Ti and 3080 Ti specs.
https://www.techpowerup.com/gpu-specs/nvidia-ga102.g930
So in theory, you are getting the same memory bandwidth as the 3090, but only 12GB of memory on the card, while getting 92.9% of the full GA102 die. That is much more die than the 3080, but is 4.7% less die than the full 3090.
If memory is the same effective speed, without the dual channel, you might be able to clock it higher, while also not having memory on the backside of the PCB to cool, which could be better for OCers, but at stock, you get no benefit. That leaves the question of if the clockspeed will be kept the same as the 3090 or will be increased. If it can clock higher, then the 3080 Ti might add some performance due to extra frequency, but there are likely going to be limits on that.
So you are WRONG that the Ti series has gotten the full die.
Not even the 980 Ti, which got the 200 die (so like a 100 series, but the 800 series maxwell was a poo stain that they made go away because it ran hot and crappy). For the 1080 Ti, I already showed that was a cut down die.
So, automatically committing yourself to buy something is a little absurd, as is over-predicting the specs of the card and its performance. It will be closer in performance to the 3090 than the 3080, but as to beating the 3090, nothing I've seen nor shown here today really suggest that.
As to performance, that leaves the question are you buying it for raytracing or DLSS? On traditional rasterization, it looks like the RX 6900 XT beats it, both on memory size, not so on memory bandwidth, and in raw TFlops. But, once you add in if you are buying for raytracing, then I get your point.
But think of what you said. You just committed to paying a 20% premium over a competing product with more raw horsepower and more VRAM, but with lower performance in still niche use features. You are basically telling them to charge you more and undercutting your leverage as a consumer.
Companies scrape data everywhere they can and crunch their own estimates on performance and pricing to come up with the end price. Them doing absurd pricing on the 2080 Ti and RTX Titan and Titan V are what led to them raising the 3090 to $1500, which for all intents and purposes, that IS where the former Ti slots, NOT the Titan. They then told people the 3080 is the flagship and priced it where the 2080 released at, but was the price of the 1080 Ti. Consumers saying they would pay more year after year is what drove the prices to DOUBLE in a span of 2 generations.
Bit I digress.
In any case, I hope this better informs you. I leave you with the NAVI 21 information, so you can compare shaders, TMUs, ROPs, and clocks. Might even inform you why AMD does better at rasterization. When Nvidia kicks into int/fp mode instead of fp/fp, you basically divide the shaders by 2 to figure out what you are actually getting on that performance.
https://www.techpowerup.com/gpu-specs/amd-navi-21.g923
Edit:
This let's you see where the performance of the RT is before AMD releases their version of super resolution. This is also on their 6800, not the 6800 XT. There is one raytracing core per CU. So the 6800 XT has 20% more RT cores. That likely places is where the 2080 Ti and 3070 are with DLSS on. The super resolution isn't introduced for a month or two after the card is released, which may further scale the performance on the AMD cards. AMD is using Direct X Machine Learning (DXML). Since DXR and DXML are poised for use over the proprietary RTX and DLSS on many games, just because of consoles in part, you may see limits to adoption of Nvidia specific solutions. Their raytracing cores and tensor cores will still be used, however. And we won't know AMD's full RT performance until reviews, then the super resolution effect for a month or two after that.
I just hope AMD has the review embargo lift THE DAY BEFORE THE LAUNCH. And I really think a 6800 XT with a good waterblock may be the way to go this time. It took both Rage and SAM to meet the 3090 on the 6900 XT, so a good mod and OC on a 6800 XT should have me going toe to toe with a stock 6900 XT easily (or their mild OC, as Rage is about 2% and the SAM is like 4 or 6%, so they had a 6% or 8% extra performance going on the 6900 XT, meaning OC the lower card, save $350, put $150 toward a good water block, maybe $50 toward a good backplate, then however much for additional pads for whatever you want to sink beyond that. Or, if Optimus creates a waterblock, that will likely cost $389 or so on that, so you could pick up an optimus with Fujipoly (what they ship with)).
Edit: Aside from 2 times where I put "is" instead of "are," I was wrong on the timeline of 100 series as a Ti (which was shown above) as the 980 Ti was released in 2015. So, just correcting it to about 5 years instead of 10 years, for fairness to Nvidia.Last edited: Nov 1, 2020iunlock, tps3443, Robbo99999 and 2 others like this. -
https://www.techpowerup.com/gpu-specs/nvidia-gp102.g798
https://www.techpowerup.com/gpu-specs/nvidia-tu102.g813
Just to give the 102 die lineup for the Pascal and Turing cards, since Maxwell 200 and Ampere GA102 were shown above in the last post. But since you can only attach so many screenshots/pics, adding this, just for future reference and rebuttals. All of those come from techpowerup GPU spec library.Robbo99999 and Papusan like this. -
You’ll be ok. Plenty of cpu power. You may or may not agree with this on a nice shiny $1,000 dollar processor, but because the 10980XE is soldered and cannot be delidded, I would recommend spinning it on a flat surface to check how flat the IHS is. You may want to consider lapping it on some glass to make it super flat.
You’ll probably manage 4.8Ghz at 1.2V or even less. The silicon seems pretty good compared to my 7980XE.
![[IMG]](images/storyImages/6-DD585-A1-DB06-4-B84-BD66-A4-B3-FF6-EC281.jpg)
Last edited: Nov 2, 2020ajc9988 likes this. -
Can't wait to see the MSRP on this bad boy:
https://www.evga.com/articles/01454/kingpin-3090/
-
So my assorted set of 5 groups of shunts per milliohm type came in today. I bought two packages, so I have ten 10 mOhms, ten 15 mOhms, ten 5 mOhms, etc etc. I actually have 3 more identical packages coming from Aliexpress, so that's going to be shunts for life.
- SMD Package 2512, max. 2 Watt
- 10 values, 5 pieces each
- 1, 2, 5, 8, 10, 15, 20, 25, 50 and 100 mOhm
https://www.amazon.com/gp/product/B07B2PM5TV/
https://www.aliexpress.com/item/32368758351.htmlRobbo99999, tps3443, electrosoft and 2 others like this. -
-
Here are some very early runs with the RTX 3080. I had to do a tune up on one of the test benches last night and it took me several hours from cleaning out the water blocks, radiators, flushing everything and everything... phew... it got late fast, but before I called it a night I took a quick spin down the 1/4 mile.
All in all, not bad. Very different characteristics with the 30 series as expected. I've analyzed some data on its behavior and it's quite interesting. The raw performance is pretty nice and it'll benefit those demanding titles a bit, but nothing astronomical.
Next this is going on the main test bench after I gather and record some more data.
It's that time of the year again...
GPU on air with the stock cooler of course.
https://www.3dmark.com/fs/23903049Attached Files:
Rage Set and Robbo99999 like this. -
-
-
-
The 1080Ti is such a great card still. It's one of the few cards that I've kept as selling it was actually taking a loss for all that the card offers. I have one in a gaming rig in the guest bedroom and the other on a test bench being used as just a GPU to turn on the monitor lol. Sad I know... both water cooled.
You're a monster. Really informative and great posts. It was nice reading familiar content written in detail. Good job.
Hey Rage, how's it going buddy? Yea I'm with you... about the VRAM talk I would love for anyone to test DCS (realistic flight sim) as that thing eats up VRAM for lunch and not only that it taxes your system Mem like no other. A large part of the system RAM usage is due to it being built on an older engine with mem leaks etc, but still the raw usage will max out your system memory with Maxed out settings.
*Sorry for the camera shot, it was taken while I was in game real quick with one hand.
If there are any DCS pilots in here, DM me... I'll fly with you.
Here are some screenshots to hopefully motivate someone to try it out LOL... because I really would like to see your RAM usage + VRAM.
This missile shot with the AIM120 from my F15 was pretty epic in taking down a cocky F14 guy that wouldn't stop yapping after him spamming all his AIM-54 Phoenix missiles at me. For those that don't know the AIM-54s from the F14 are very long range, uber fast and hard to dodge.
Formation flying during a competition SATAL 4v4 match. It's serious business lol... a ton of fun.
Oh my way back home returning to base after taking out the enemies AWAC.
Attached Files:
ssj92, Papusan, Charles P. Jefferies and 2 others like this. -
-
-
-
-
*Official* NBR Desktop Overclocker's Lounge [laptop owners welcome, too]
Discussion in 'Desktop Hardware' started by Mr. Fox, Nov 5, 2017.
