Latest Intel driver makes reference to DG2 GPUs

[HEXUS]

Remember, DG2 will spearhead Intel's discrete graphics in the gamer / enthusiast segment.

Read more.

[Tabbykatze]

With Tiger Lake at 96EUs, it gets a 3D Mark 11 score of 6710 and Fire Strike of 4820 whereas (being nice) compared to a 6800XT it gets 52046 and 49735 respectively which is and 675% and 931% difference in performance.

Now, obviously, we have to account for the fact that the 96EU tiger lake is 400-1350mhz in a mobile low power configuration and is their version 1.5 of their DG1 system. So if DG2 is perfect scaling with no improvement (so they focused on upscaling first) that would be a 530% increase in performance putting it substantially behind a 6800XT. So the DG2 generation performance increase on IPC and frequency will have to pull out all the stops to beat that deficit.

What also has to be remember is increase of cores traditionally does not provide a perfect scaling of performance, Intel may be different with the DG2, maybe it does but it's unlikely.
I don't think DG2 will compete on the high end, I think they'll play in the volume market first like AMD did for all those years trying to catch OEM wins etc. They will catch wins and the DG2 will saturate the market because Intel will use it's financial horsepower to make it so.

Frankly, it looks like they're targeting the volume, mid-market initially because it looks like it'll barely be scraping an RX580 level performance which altogether isn't bad. A second frankly, I bet Intel will push themselves heavily into the OEM space and do the dirty like before just so their GPUs can get mindshare because not many will want a mid-range, first generation (DG1 can barely be called "first") untested GPU in their system for gaming.

[KultiVator]

These cards seem to have been in development for a seriously long time... it'll be interesting to see whether they offer Nvidia and AMD any serious competition and whether a third player in the gaming GPU space helps drive better price/performance for consumers.

[edmundhonda]

Now, obviously, we have to account for the fact that the 96EU tiger lake is 400-1350mhz in a mobile low power configuration and is their version 1.5 of their DG1 system.

The discrete DG1 with dedicated GDDR6 and presumably a less restricted TDP hit 5528 in Fire Strike, although there's the chance newer drivers have made further gains...?

If this is the 'enthusiast' card it's a little underwhelming.

[DanceswithUnix]

These cards seem to have been in development for a seriously long time... it'll be interesting to see whether they offer Nvidia and AMD any serious competition and whether a third player in the gaming GPU space helps drive better price/performance for consumers.

As long as GPUs are all competing for the same limited TSMC & Samsung silicon wafers, I don't see it helping.

[Tabbykatze]

The discrete DG1 with dedicated GDDR6 and presumably a less restricted TDP hit 5528 in Fire Strike, although there's the chance newer drivers have made further gains...?

If this is the 'enthusiast' card it's a little underwhelming.

That's a whole 14% faster with the memory focused RAM that puts the 6800XT at 899% faster than the DG1, so my scale up figures aren't changed by much.

As I alluded, the DG1 was just hot garbage but it was simply because it was a first generation alpha product that really shouldn't have seen the light of day as a "GPU" but more of a test bed for development which really is what it became. But that was during the same that Raja was doing his whole overmarketing shtick before he got his choke chain pulled so that's why it was made out to be far more than it actually was.

[Ericthetroll]

"Intel's first gamer / enthusiast targeted discrete GPUs"

The i740 AGP graphics cards took that title long ago and failed miserably, much like this probably will.

What Intel really need to do to capture another market is to release Clear Linux in an LTS version.

[JayN]

Intel has commented that the ray-tracing hardware in DG2/Xe-HPG has also been designed to accelerate their oneapi image rendering applications.

at IMS20, Intel described their "Rambo Cache", which is being used with the Ponte Vecchio GPUs in Aurora, as a discrete SRAM chip that provides higher performance for matrix operations than HBM memory.

Intel announced their hybrid bonding stacked SRAM had taped out in Feb 2020. That could be an interesting development.

The IMS20 video describing the Xe-HPC Rambo Cache is Frank Hady's, about 23 minutes in. There's a link to all those IMS20 presentations on Raja's twitter page, Rajaontheedge

[Tabbykatze]

The IMS20 video describing the Xe-HPC Rambo Cache is Frank Hady's, about 23 minutes in. There's a link to all those IMS20 presentations on Raja's twitter page, Rajaontheedge

Firstly, you're double (triple) posting and secondly, not sure anyone cares what you have to say because your sole purpose on this forum is to shill for Intel and when contested with realistic contestation, you're quieter than a mouse which is quite telling.

Edit: additionally, the rambo cache isn't that innovative, it's just a natural progression.

[JayN]

What also has to be remember is increase of cores traditionally does not provide a perfect scaling of performance, Intel may be different with the DG2, maybe it does but it's unlikely.

Intel demoed 1,2 and 4 tile execution of a Xe-HP, which max'd out at 42 TFlop FP32 for 4 tiles. The demo showed that it did scale almost linearly.

The externally manufactured Xe-HPG is built on the same tiled architecture as Xe-HP, but I believe it is currently monolithic, while the Xe-HP GPUs are chiplets, stitched together with Intel's emib technology.

[DanceswithUnix]

Firstly, you're double (triple) posting and secondly, not sure anyone cares what you have to say because your sole purpose on this forum is to shill for Intel and when contested with realistic contestation, you're quieter than a mouse which is quite telling.

Perhaps they are really a deep learning neural net running on a prototype Xe card. Just a few bugs that need ironing out

[Tabbykatze]

Intel demoed 1,2 and 4 tile execution of a Xe-HP, which max'd out at 42 TFlop FP32 for 4 tiles. The demo showed that it did scale almost linearly.

The externally manufactured Xe-HPG is built on the same tiled architecture as Xe-HP, but I believe it is currently monolithic, while the Xe-HP GPUs are chiplets, stitched together with Intel's emib technology.

You sure it's not glued together?

And a multi-tile system can have linear increase in performance the same way a physical multi gpu system can if it is managed correctly but your error was moving from a single GPU to a multiple GPU core, my statement was you can't just throw more power into a single die and for it to scale linearly.

If you put 4 Nvidia 3090s in quad SLI, in the right circumstances it's a linear increase in performance.

All Intel is doing is taking a leaf out of AMDs 6990, using old methods in really not very new ways.

Hmm, is your name Omar perchance, you have a very similar single focus as him on the OC3D pages.

[nobodyspecial]

These will at worst, steal WAFERS from AMD (cpu or gpu, hurts either way). Nice move Intel. Like I said before Intel has many ways to skin this cat, even legally. If Intel decides TSMC makes cpus for a year or two until their fabs get back to business in Q1, then I think things get worst FAST for AMD. It would be better for AMD if Intel's 7nm fab fix is REALLY working (said in Q call it was in and working now), so they won't be fighting with yet another player over already short wafers across the board for everyone it seems.

I just think if Intel is forced to do cpu at TSMC for a year, AMD gets screwed a year earlier and maybe before they get any real NET INCOME. There is already only about 1yr left for that IMHO anyway as Intel will be back in one way or another Q4. IE, chip that kills Q1 looks good (Q4), and many packaging options then helping also either way (TSMC or Intel fabs, Intel has made chips from 4 different fabs in a single slab, impressive). That said, I see a larger problem coming from Apple than AMD for Intel (gpus too, so trouble ahead for NV likely also maybe a year or two after Intel). Apple wants to make everything internally if possible, which is about all they can do to grow now besides buy other companies. So knocking out a cpu and gpu are top jobs for them for the next few years and they've been advertising it for at least that long already (Intel kick info came years ago, gpu rumored for ages now).

[nobodyspecial]

These will at worst, steal WAFERS from AMD (cpu or gpu, hurts either way). Nice move Intel. Like I said before Intel has many ways to skin this cat, even legally. If Intel decides TSMC makes cpus for a year or two until their fabs get back to business in Q1, then I think things get worst FAST for AMD. It would be better for AMD if Intel's 7nm fab fix is REALLY working (said in Q call it was in and working now), so they won't be fighting with yet another player over already short wafers across the board for everyone it seems.

I just think if Intel is forced to do cpu at TSMC for a year (whatever amount of time), AMD gets screwed a year earlier and maybe before they get any real NET INCOME. There is already only about 1yr left for that IMHO anyway as Intel will be back in one way or another Q4. IE, chip that kills Q1 looks good (Q4), and many packaging options then helping also either way (TSMC or Intel fabs, Intel has made chips from 4 different fabs in a single slab, impressive). That said, I see a larger problem coming from Apple than AMD for Intel (gpus too, so trouble ahead for NV likely also maybe a year or two after Intel). Apple wants to make everything internally if possible, which is about all they can do to grow now besides buy other companies. So knocking out a cpu and gpu are top jobs for them for the next few years and they've been advertising it for at least that long already (Intel kick info came years ago, gpu rumored for ages now).