A glimpse at future GPUs...but will they play Crysis ?

[J7SC_Orion]

...All too often, gamers think that they get to buy the 'latest and greatest' in terms of graphic processors, but the real money for and innovation by vendors is in HPC and enterprise solutions (AI and other). What is happening in that market (where a single GPU accelerator can cost close to $20k and doesn't even come with HDMI or DP I/O) can sometimes foreshadow what gamers get as cut-down hand-me-downs in future desktop and even workstation GPUs.

 

With that in mind, below is a brief screen-shot summary of the latest by AMD, Intel and NVidia - the information and pics are straight from their respective corporate sites, so some of the performance numbers should be taken with a grain of salt (in addition to some numbers being doubled up via two physical 'cards' being used). Also, this is by no means an in-depth look, just a brief fly-over. Still, it is clear that multiple processing cores or 'tiles' per single card are here to stay. In addition, these enterprise HPC/AI offerings are highly scalable, both horizontally (including but not limited to adding more of the same accelerators, often w/ VRAM pooling) as well as vertically (integrating it into the system RAM pools). That in turn also underscores another 'trend':

 

All three vendors want to max their revenues and minimize their clients' developer time by offering turn-key solutions. AMD has been making huge inroads into the server market over the past few years on the 'CPU front', and AMD Epyc Genoa with up to 96 cores/192 threads, 12 channel DDR5 and 128 PCIE 5.0 lanes looks set to extend AMD's inroads. Intel though is pushing a new combo ('Max' series) of both CPU and GPU that are highly integrated via Intel's oneAPI for heterogenous computing, with AMD releasing its own new software packages to go after the current leader, NVidia. NVidia's problem is that its attempt at purchasing ARM (RISC) failed, though they certainly have the licenses to design their own RISC 'CPUs' to round out their impressive 'accelerator' offerings.

 

Eagle-eyed readers will note that Intel's latest offerings include the 12VHPWR connectors (at least on the PCIe versions) and looking at the power consumption figures, some of the 'cards' are rated at well over 500W. So enjoy some 'nerdy' numbers such as 128 GB of HBM+ VRAM, massive memory bus widths, 100 billion transistor XTOR GPUs and such...

 

Will they play Crysis ? No idea... But when the hand-me-down consumer / desktop versions of these monsters make it to market, we get to try 

 

alphabetical order, images and tables sourced from AMD, Intel and NVidia respectively

 

  

 

 

Edited December 2, 2022 by J7SC_Orion

[bonami2]

Missed this thread. Am all in for Multi Die. Just need game to support it.  Could even have close TDP to a single chip with the power scaling issue.

[J7SC_Orion]

11 minutes ago, bonami2 said:

Missed this thread. Am all in for Multi Die. Just need game to support it.  Could even have close TDP to a single chip with the power scaling issue.

 

mGPUs via tiles are supposed to appear as a single GPU to the OS and drivers...but there might be latency issues on the gaming desktop front, at least initially...

[bonami2]

7 minutes ago, J7SC_Orion said:

 

mGPUs via tiles are supposed to appear as a single GPU to the OS and drivers...but there might be latency issues on the gaming desktop front, at least initially...

Well what are they waiting for!  I want a 6 Die 4090 at 185w tdp lock to each so like 700w and 80m PPD per days

Edited December 3, 2022 by bonami2

[J7SC_Orion]

35 minutes ago, bonami2 said:

Well what are they waiting for!  I want a 6 Die 4090 at 185w tdp lock to each so like 700w and 80m PPD per days

 

...they are probably waiting to make a few $ billion first from the HPC / Big Data /  Ai  corporate folks. However, if you're in a hurry and own your own power station (for that ~4 kW), you can try out this today...

 

[Sir Beregond]

18 hours ago, J7SC_Orion said:

 

mGPUs via tiles are supposed to appear as a single GPU to the OS and drivers...but there might be latency issues on the gaming desktop front, at least initially...

Yeah I'd imagine similar to when Ryzen first came out. I'll be interested to see how that plays out starting with RDNA3 and its multi-chip design.

[J7SC_Orion]

...some more OP-related head-spinning future tech for GPUs...somewhat speculative of course, but interesting nonetheless...

 

 

 

[J7SC_Orion]

...and another 'tech' video involving some of the accelerators in the OP I watched last night. It is time-stamped re. losses just from wiring re. memory and for large data centers and AI operators this adds up real fast. The rest of the video is also worth watching (as is the channel in general) re. just how much energy gets wasted, how they're starting to hit the 'memory wall' and the drive to get to processor-in-memory chips to cut efficiency losses and get past the memory wall. 

 

 

[J7SC_Orion]

...a 'nerdy' update slide presented by Intel a few days ago (via YT/TechTechPotato). Tiling - also per original and other posts above - will become the main-stream in the future, and if Intel (along w/ AMD and NVidia) perfect the on-card 'inter-links' we will be in for some real treats, never mind connectivity between separate discreet cards that carry tiles. Some of this will make it down to the consumer / gamer level, eventually...

 

[J7SC_Orion]

...while this thread is about the latest and even upcoming workstation (& other) GPUs, including via mGPU 'tiles', and chiplets, DerBauer released a highly interesting video about the last true dual AMD GPU, from 2021 - what a monster ! 

 

I'm not a friend of proprietary connections, though...

 

 

[J7SC_Orion]

...It has been a while for this thread, but the vid below is extremely interesting, if a bit tedious at times. @Sir Beregond will appreciate the AMD vs NVidia wafer cost differentials, and @bonami2 might find his next graphics card in there some where...

 

The long and the short of it is that AMD is about to fire a major shot and NVidia re. AI hardware, though NVidia's ecosystem on related software still rules. In any event, enthusiasts and gamers will just get the bread crumbs that fall off the AI table

 

 

[J7SC_Orion]

...an interesting if lengthy video referencing not only some of the AI APU/GPUs but also points to the real challenge for anybody but NVidia - the software ecosystem. But never say never...