Nvidia's 3000 Series RTX GPU [3090s with different memory capacity]

That's not the only issue with graphics APIs, another one is strict ordering requirements, objects must be processed in order they were submitted, which would limit parallelism significantly without synchronizing the work between chiplets.
There is little doubt, front end must be reworked to account for chiplets. I guess the new front end must be very smart to distribute the work across multiple chiplets in an efficient manner -- it should batch more draw calls and track more state than ever to keep utilization high.

I’m thinking it’s the exact same challenge in monolithic dies. The only difference is lower latency communication because it’s on die. Question is whether the increased latency and lower bandwidth between chiplets is enough to require a new approach to distributing work.
 
I’m thinking it’s the exact same challenge in monolithic dies.
Yes, but this should be less of an issue for narrower monolithic dies with higher bandwidth and lower latencies.

Question is whether the increased latency and lower bandwidth between chiplets is enough to require a new approach to distributing work.
Not sure about completely new approach, but without any changes the wide multi chiplet design would result into lower utilization for sure.
 
Ran across a somewhat lower level description of Turing and Ampere’s Load-store/Tex memory pipeline.
https://docs.nvidia.com/nsight-graphics/AdvancedLearning/index.html#gt_l1_throughput
https://docs.nvidia.com/nsight-graphics/AdvancedLearning/index.html#gt_l1_throughput

What does it mean to write back to the SM and not the RF?

In this simplified model, memory & texture requests follow these paths:

HLSL Wave Broadcast are counted here
  • Texture/Surface Read → TEXIN → T-Stage → TEX Data → Filter → Writeback → SM
    • Includes texture fetches, texture loads, surface loads, and surface atomics
  • Surface Write → TEXIN... → T-Stage → LSU Data → Writeback → RF
    • Surface writes cross over to the LSU Data path
  • Memory Barriers → LSUIN & TEXIN → … flows through both sides of the pipe
  • Primitive Engine attribute writes to (ISBE, TRAM) → LSU Data
  • Primitive Engine attribute reads from ISBE → LSU Data
  • Local/Global/Texture/Surface→ T-Stage [miss!] → M-Stage → XBAR → L2 → XBAR → M-Stage → Cache Data SRAM
 
What does it mean to write back to the SM and not the RF?
To me it implies that the result of a texturing operation, which would normally end up in a vector register, is simply sent to the operand collector where it is used as the source for an instruction at the appropriate time.

In general you can avoid allocating registers for texture results if you know that you can always collect texture results and issue them "soon" to the SM.

In the worst case, the compiler might produce an instruction which stores the texture result to a vector register, if that result is needed two or more times. If the result is only required once there's no reason to use a vector register.
 
Blender 3.1 – How Fast Are GPUs in Cycles X, Eevee & Viewport? – Techgage
May 12, 2021
It’s almost hard to believe, but Blender’s 3.0 version released a full six months ago. With it came a vastly overhauled Cycles engine, the introduction of HIP (Heterogeneous Interface for Portability) for AMD’s Radeon graphics cards, the arrival of the much-anticipated asset browser, a doling out of additional geometry nodes, and more.
...
At this point, Blender 3.1 has been out for a couple of months, so this article should have come a lot sooner. Truth be told, though, the performance we evaluate was largely unchanged from 3.0 to 3.1, so there’s not much this article could tell you that the previous 3.0 one couldn’t. That said, we do have two additional GPUs in the rankings here that weren’t available at 3.0 launch, including AMD’s Radeon RX 6500 XT, and NVIDIA’s GeForce RTX 3050.
 
I do think it was a gag, it seems like it was a voice over edited in after the fact by the way they cut away from him just before he said it and returned to him right after.

Just an opinion, but the alternative is nVidia's PR is morons and while they might be evil and ruthless I've never thought of them as totally incompetent. No way they'd put that out after all the flack they took for the original tie line.
 
Interesting watching price trends on Amazon. Back at the end of April, 3070's were still around 1000 USD or more. And they'd been at or above 1000 USD for about a year now. And they've basically been over 900 USD since launch.

Many of those same cards are now under 700 USD and occasionally dipping below 600 USD on Amazon. Still well above the "MSRP" of 499 USD that NV claimed at launch, but at least it's approaching relatively sane pricing levels. :p

Still far more than I'm willing to pay for a new graphics card, but it's getting closer.

Regards,
SB
 
Interesting watching price trends on Amazon. Back at the end of April, 3070's were still around 1000 USD or more. And they'd been at or above 1000 USD for about a year now. And they've basically been over 900 USD since launch.

Many of those same cards are now under 700 USD and occasionally dipping below 600 USD on Amazon. Still well above the "MSRP" of 499 USD that NV claimed at launch, but at least it's approaching relatively sane pricing levels. :p

Still far more than I'm willing to pay for a new graphics card, but it's getting closer.

Regards,
SB

Also im seeing quite low prices on the used market for GPU's, could be gpus that have been mining though.
 
The price drops are good if you for sure need a GPU now but for more opportunistic buyers (like myself) it's a bit of a mixed bag due to how volatile the market situation is going to be in a few months. People can try to rationalize it post fact but it's going to horrible if you buy a GPU now and say find out in only 3 months it's potentially halved in value which can be hundreds of dollars. At the same time you risk potentially still 6 months+ of nothing better coming along and wasting that time as well.

Like for myself just in the past month or so just for Nvidia -

I had the opportunity to buy a 3080 FE for MSRP (well CAD MSRP at $1030) a few weeks ago. Passed on it and already seen some dominos fall with a 3080 12GB being on sale for $1000. I did the math and a post mining and next gen collapse could mean what $500 loss for just a few months of use?

Ordered a 3070 FE for MSRP ($680 CAD). Just cancelled it, didn't even pick it up. Even at MSRP it's not a great value especially given the VRAM (I won't go into this debate too much here).

Ordered a 3060 for a bit over MSRP ($480 CAD, I guess MSRP in CAD would be $450?). Didn't want to risk the heatsink (noise sensitive) among other things and so returned it on the last day unopened.

Coincidentally I lucked into a 3060ti FE at MSRP ($550) on that last day. But now I'm not even entirely sure if I want to go with it even if it's the best Nvidia deal at least.

My other problem is I feel both Ampere and RNDA2 will have aging issues (which is a separate debate) and utility issues for myself for what you pay (even if they were MSRP).
 
Still not seeing anything close to MSRP in the US, prices haven't dropped in over 2 months. Still want to buy an Ampere for my wife but not paying $200+ over what they should be.
 
But everyone knew that Nvidia can't sell/make profit from such a big die and that $699 MSRP was fake. That is why day one sales blitz, were people stood in line to get MSRP was a sales gimmick, bcz they were the only people offered MSRP.
 
Back
Top