Intel Xe DG-2 topic :)


  • Here the latest news & info of the all new Intel Xe DG-2 GPU:
    it seems the performance will be between a Radeon RX 6800 and Radeon RX 6800 XT, with 512 Execution Units aka EU's it will have a total of 4096 cores.
    https://www.pcgamer.com/intel-xe-dg2-g…d-specs-rumour/


    The point of all this is that we can talk about the top-end 512 EU Intel Xe DG2 card coming with 4,096 stream processors, which would potentially put it above the AMD RX 6800....If this latest GPU list is correct, the 384 EU card would also sit above the AMD RX 6700 XT, clock speeds notwithstanding...
    if Intel can get some out while Nvidia and AMD are still struggling with stock, then it could be a hell of a time for Intel to offer a third way for gamers to get graphics cards.

    This is good news imo, a 3rd player will give the buyer more choice just as long these cards will be consumer friendly.

    Maybe this is a bit early to make, but on the otherside if any of you have news of this new exciting graphics chip, feel free to add your finds in here as well.

    Rock on and stay safe.

  • Auch wenn ich sofort für einen dritten, starken Spieler im GPU Feld bin, einfach Shadereinheiten nach Anzahl miteinander zu vergleichen ist ein Blödsinn. ;) Das kannst ja z.B. bei der RTX 3080 und der RX 6800 XT auch nicht machen; RT beiseite lassend sind die so ungefähr im selben Leistungsbereich, aber die RTX 3080 macht das mit 8704 Shadern und die RX 6800 XT mit 4608. Andere Architektur halt, das kann man so nicht wirklich vergleichen.

    Zudem kommt Intel's massiver Rückstand im Softwarebereich, den aufzuholen sicher kein Pappenstiel wird. Treiber, Tools, SDKs usw. Da sind nVidia und AMD meilenweit voran. Man bedenke wieviele kleine Probleme NV und AMD Treiber andauernd haben, da wird dauernd was gepatched. Und Intel muß hier nahezu von null starten. Ich meine, ich lasse mich nur zu gerne positiv überraschen, aber ich bin beim initialen Xe Highend Produkt vorerst erst Mal skeptisch. Einfach Mal warten und schauen was dann wirklich dabei rauskommt. :)


    Even if I'm all for a third, strong player in the GPU field, to simply compare shader counts is a tad bit foolish. ;) You can't do that for e.g. a RTX 3080 and a RX 6800 XT either; RT set aside those two are roughly in the same ballpark in terms of performance, but the RTX 3080 does it with 8704 shaders, whereas the RX 6800 XT does it with 4608. Simply a different architecture, you can't compare it just like that.

    In addition to that Intel is way back in terms of software, to make up for all that won't be no easy feat. Drivers, tools, SDKs, etc. nVidia and AMD are miles ahead in that field. Think about how even NV and AMD drivers have constant problems and receive patches all the time. And Intel needs to start from scratch almost. I mean, I'm always ready for a positive surprise, but I'm still sceptical when it comes to the initial Xe high end product. Just waiting and seeing what we'll really get from them. :)

    1-6000-banner-88x31-jpg

    Stolzer Besitzer eines 3dfx Voodoo5 6000 AGP Prototypen:

    • 3dfx Voodoo5 6000 AGP HiNT Rev.A-3700

    [//wp.xin.at] - No RISC, no fun!

    QotY: Girls Love, BEST Love; 2018 - Lo and behold, for it is the third Coming; The third great Year of Yuri, citric as it may be! Edit: 2019 wasn't too bad either... Edit: 2020... holy crap, we're on a roll here~♡!

    Quote Bier.jpg@IRC 2020: "Je schlimmer der Fetisch, desto besser!"

  • yeah I took notes from that link rather globally, then again I am curious how Intel will do seeing that NV's Tom Peterson & AMD's Raja Koduri work for them for their GPU Department that is at least some helpful knowledge to get them probably something quite decent, I just like standing open to new inventions in their field of computing, but you are right Intel is light years behind, but with the right minds in their team they might do alot better than most would expect, and hey I could be wrong, anything can happen in the end :)

    Ah well when there is news for these new Intel bred GPU's I'll post it here for sure :)

  • Wird echt interessiert was Raja Koduri alles für Intel tun kann oder wird getan haben werden. Ich bin ja schon beeindruckt daß AMD ohne ihn noch so gut fährt wie sie's tun. Was Peterson angeht wird es interessant wie er seinen Rollenwechsel managen wird. Immerhin geht er scheints aus dem technischen Marketing zurück in's Engineering. Üblicherweise passiert sowas nur in die Gegenrichtung. Aber es steckt sicher mehr hinter Xe als wir bisher von Intel in Sachen iGPU gesehen haben. Und da waren die Treiber auch eher müllig. Ich frage mich wer an der Xe Software arbeitet?


    I'm actually interested in how much Raja Koduri will do or will have done for Intel. I'm quite amazed AMD can still go as strong as they do after he left. As for Peterson it'll be interesting how he'll manage his change of role, given he's seemingly gone back to engineering after having tended to technical marketing at nVidia. Things usually go the other way around. But there's surely more backing Xe than was previous iGPU products. And those had pretty crappy drivers too. I'm wondering who's working at the Xe software?

    1-6000-banner-88x31-jpg

    Stolzer Besitzer eines 3dfx Voodoo5 6000 AGP Prototypen:

    • 3dfx Voodoo5 6000 AGP HiNT Rev.A-3700

    [//wp.xin.at] - No RISC, no fun!

    QotY: Girls Love, BEST Love; 2018 - Lo and behold, for it is the third Coming; The third great Year of Yuri, citric as it may be! Edit: 2019 wasn't too bad either... Edit: 2020... holy crap, we're on a roll here~♡!

    Quote Bier.jpg@IRC 2020: "Je schlimmer der Fetisch, desto besser!"

  • Wird echt interessiert was Raja Koduri alles für Intel tun kann oder wird getan haben werden. Ich bin ja schon beeindruckt daß AMD ohne ihn noch so gut fährt wie sie's tun.

    I'm actually interested in how much Raja Koduri will do or will have done for Intel. I'm quite amazed AMD can still go as strong as they do after he left.


    Maybe AMD's graphics solutions are as good as they are because Koduri left.

    If you believe in rumours, he wasn't substantially involved in development of Navi.

    Ich werde mich von keinem einzzzigen Prozzzessor trennen.
    Jedoch lockt es mich beinahe, ihn Dir zu überlassen, nur um zu sehen, wie er Dich in den Wahnsinn treibt :evil:

    Meine Begehren

  • "wird getan haben werden?

    Meinten Sie: getan haben wird?

    Jop, das stimmt wohl eher.

    1-6000-banner-88x31-jpg

    Stolzer Besitzer eines 3dfx Voodoo5 6000 AGP Prototypen:

    • 3dfx Voodoo5 6000 AGP HiNT Rev.A-3700

    [//wp.xin.at] - No RISC, no fun!

    QotY: Girls Love, BEST Love; 2018 - Lo and behold, for it is the third Coming; The third great Year of Yuri, citric as it may be! Edit: 2019 wasn't too bad either... Edit: 2020... holy crap, we're on a roll here~♡!

    Quote Bier.jpg@IRC 2020: "Je schlimmer der Fetisch, desto besser!"

  • For the 46 CU, the die size is 190 mm^2 which gives between 9.4 and 11.6 BB Transistors depending on efficiency estimate (AMD's first gen 7 nm efficiency was 43% and 2nd gen 7 nm efficiency was 53% and this is the range I used), which would put the 46 CU at slightly less to slightly better than a 5700XT

    For the 64 CU, I estimate 13-16 BB transistors, which would put it less than an Xbox Series X to slightly better than a 6700 (but of course this depends on my die size estimate, in addition to the efficiency estimate)

  • Wird echt interessiert was Raja Koduri alles für Intel tun kann oder wird getan haben werden. Ich bin ja schon beeindruckt daß AMD ohne ihn noch so gut fährt wie sie's tun. Was Peterson angeht wird es interessant wie er seinen Rollenwechsel managen wird. Immerhin geht er scheints aus dem technischen Marketing zurück in's Engineering. Üblicherweise passiert sowas nur in die Gegenrichtung. Aber es steckt sicher mehr hinter Xe als wir bisher von Intel in Sachen iGPU gesehen haben. Und da waren die Treiber auch eher müllig. Ich frage mich wer an der Xe Software arbeitet?


    I'm actually interested in how much Raja Koduri will do or will have done for Intel. I'm quite amazed AMD can still go as strong as they do after he left. As for Peterson it'll be interesting how he'll manage his change of role, given he's seemingly gone back to engineering after having tended to technical marketing at nVidia. Things usually go the other way around. But there's surely more backing Xe than was previous iGPU products. And those had pretty crappy drivers too. I'm wondering who's working at the Xe software?

    Hmm very good question yea, not sure how well Intel's drivers are with current GPU's they make from what I know laptops always ran pretty decent with them as for the ones I worked with or tried out.

    All in all a very interesting thing going on at the Blue team though, that is one thing for sure, bottom line we finally have a thrid player for the GPU market and it doesn't matter if they start small , who knows how it will all evolve over the years, technology does shift fast with things like these, so I am curious how they will do to compete against AMD & NVIDIA but also how long it will take them to catch up, money wise Intel is still in the strongest position, so as that goes I wonder how well their strategies will be in the upcoming years.

    For the 46 CU, the die size is 190 mm^2 which gives between 9.4 and 11.6 BB Transistors depending on efficiency estimate (AMD's first gen 7 nm efficiency was 43% and 2nd gen 7 nm efficiency was 53% and this is the range I used), which would put the 46 CU at slightly less to slightly better than a 5700XT

    For the 64 CU, I estimate 13-16 BB transistors, which would put it less than an Xbox Series X to slightly better than a 6700 (but of course this depends on my die size estimate, in addition to the efficiency estimate)


    Welcome to Voodooalert.de Skyye :) Glad to see you made it here as well, even I linked this thread in our discord at the NJA :)
    It still makes me wonder how Intel's 128, 384 & 512 EU Cards will do, Execution Units is their name for Compute Units I believe and no math isn't my kind of thing hence why I ask :)

    5 Mal editiert, zuletzt von Gold Leader (28. Februar 2021 um 12:22) aus folgendem Grund: typo's mostly

  • Thank you for the invite to this board!

    Both Nvidia (with its 3000 series) and Intel (with DG-2) just confuse the issue with compute units and shaders. It's all bullshit. If it sounds too good to be true, it probably is. Like Nvidia's published shader count for the 30 series is twice the real #. Intel's EU counts are 8X what the actual number of compute units is. They both do it to confuse people and obfuscate the truth.

  • Welcome to the forums, Skyye! :)

    Not to confuse the different compute units or shader cores is actually a very good point.

    I think there is much more to it than the amount of cores or CUs when comparing the different ASICs from NV, AMD and Intel in the future.

    AMD's GCN was very strong in pure compute with for instance Fiji being faster than NVidia's Maxwell back in the day.

    The problem with GCN's gaming performance seemed to be more at the front end as well as the geometry throughput.

    The following was put together based on scattered memories of the chip schematics... if something seems wrong, by all means feel free to correct me ;)

    RDNA 2 now has just like RDNA 1 so called WGPs (Work Group Processors / 40 in total) each combining two CUs. One WGP group contains 4 blocks.

    Each block as 32 Lanes for FP32, FP16 and INT32 as well as 2 Lanes for FP64. The results are way more memory bandwidth as well as compute power.

    One Ampere CU (82 in total) contains 4 blocks. Each block has 16 ALUs and 16 FPs. Only block 1 has 2 DPs, the others don't. The 16 ALUs are accessible through port 0 and seem to handle FP16, FP32 and INT32. The remaining 16 FPs are accessible via port 1. Looks like the interesting part is that only FP32 instructions can be scheduled onto both issue ports 0 and 1 which gives them access to all 32 ALUs that can double their performance. That's probably why they talk about "10496 Shaders".

    I'm very curious about what Intel might be able to pull off. I don't expect a killer chip that will compete in the high ranks with both AMD and NVidia.

    However a solid start in the right direction is something I'm hoping for.


    DEU:

    Die verschiedenen shader cores und CUs nicht miteinenander zu verwechseln, ist ein guter Hinweis.

    Ich denke auch, wenn man AMDs, NVidias und später Intels ASICs miteinander vergleicht, gilt es mehr zu beachten als die reine Anzahl der CUs.

    AMDs GCN Architektur wer sehr stark im reinen Compute. Zum Beispiel Fiji vs. Maxwell damals.

    Das Problem bei GCN schien damals am Front End und am Durchsatz der Geometry zu liegen.

    Folgendes habe ich aus meiner nebulösen Erinnerung an die Blockdiagramme zusammengetippt... korrigiert's mich also bitte, falls ich wo falsch liege ;)

    Bei RDNA heißen die Dinger Work Group Processors (40 insgesamt), von denen jeder zwei CUs zusammenfasst. Ein WGP besteht aus 4 Blöcken, von denen jeder 32 Lanes hat, die FP32, FP16 und INT32 können. Für FP64 sind zwei DP Lanes pro Block vorhanden. Das Ergebnis sind mehr Speicherbandbreite und mehr Leistung.

    Eine CU bei Ampere hat ebenfalls 4 Blöcke. Jeder Block besteht aus 16 ALUs und 16 FPs. Nur der erste der vier Blöcke hat zwei DPs. Auf die 16 ALUs kann über Port 0 zugegriffen werden, was Berechnungen in FP16, FP32 und INT32 erlaubt. Die übrigen 16 FPs sind erreichbar über Port 1, was angeblich nur den FP32 Instructionen vorbehalten ist.

    So hat man für FP32 über beide Ports alle 32 ALUs zur Verfügung, was die Performance verdoppeln kann. Wahrscheinlich redet man deswegen von "10496 Shadern".

    Ich bin sehr neugierig was Intel rausbringen wird. Ich erwarte keinen Killerchips, der es mit den großen von AMD und Nvidia aufnehmen kann.

    Aber ich hoffe auf einen guten Start in die richtige Richtung.

    Voodoo-Kiste (seit 3dfx BayernLAN 2007):
    TUSL2-C, P3-S 1400, 512MB SDR, Voodoo 5 5500 AGP, 9,1GB IBM SCSI (OS), 40GB HDD (Games), Adaptec 2940-U2W SCSI
    Dualatin: GA-6VTXD, 2x P3-S 1400, 1GB SDR, 2x 60GB, Hercules Radeon 8500 128MB

    Work (CG): MSI TRX40 Creator, AMD TR 3970X, 128GB RAM, Titan RTX 24GB, Win10Pro

    Zock & Office: MSI X58 Pro-E, Xeon X5680 4,2GHz, 24GB RAM, RX590 8GB, KDE Neon

  • Hi eLuSiVe, thank you! Glad to be here!

    I am not an engineer. I have an MBA in Finance and own AMD stock, so keep track of performance to price on AMD, Nvidia and Intel CPUs and GPUs and try to forecast performance to price for upcoming products, as well.

    So just thinking big picture, Techpowerup shows the 82 CU 3090 with 10,496 shaders vs. 5120 for the 80 CU 6900XT, yet both have similar PASSMARK and timespy scores. 1/2 the 10,496 Nvidia shader number (5248) seems to make more sense.

  • Lassen wir uns überraschen was passiert.

    Eine Matrox Parhelia war damals auf dem Papier auch einer Geforce Ti 4600 oder einer Radeon 8500 überlegen...

    Let us wait what happens I think?

    A Matrox Parhelia was expected to beat a Geforce 4 Ti 4600, but only managed to beat a Radeon 8500, well actually it was thought that the Parhelia was going to beat the 9700 pro this card crushed the Ti 4600 by a mile so as that goes no P512 didn't do so well mainly a flaw within it's Ringbus memory controller and terribly unoptimized drivers & lacking of game optimization it all played along other than that the P512 is still one of the most interesting non 3dfx gaming cards made being that it is the world's first DX 8.1/ DX (.0a hybrid, it does do things no other could dream of doing.

    Could be that this new Intel GPU may be a Hybrid as well but then hopefully done right or at least better than what the P512 was back in 2002 to 2003.
    Here's my P512 AGP 256MB 256Bit DDR Rev.A 5103, love this card to bits:

    Einmal editiert, zuletzt von Gold Leader (1. März 2021 um 17:22)

  • Grade gelesen, daß die Xe DG2 GPUs nicht von Intel selbst gefertigt werden sollen. Nachdem ich die Technologien von IBM, GlobalFoundries und UMC für zu rückständig halte, bleibt fast wieder nur TSMC. Samsung ist - glaube ich - nicht ganz so beliebt im Foundry Business?

    Das würde bedeuten, daß die Lage mit den Chipverfügbarkeiten durch einen Xe DG2 Launch wieder verschärft werden würde. Ist ja bekanntlich voll am Arsch mit Navi 21, Ampere, den neuen Konsolen und Zen 3. Und dann auch noch die neue Intel GPU - da kommen Zweifel daran auf, ob Intel wirklich von den mangelnden Verfügbarkeiten bei AMD und nVidia würde profitieren können, wenn's auf die selbe Foundry losgehen. Gerüchte scheinen zumindest auch auf TSMC hinzudeuten.


    I read just now, that the Xe DG2 GPUs are not supposed to be manufactured by Intel themselves. Since I consider the technologies of IBM, GlobalFoundries and UMC as too old, this leaves mostly TSMC, again. Samsung is - or so I believe - not quite that popular in the foundry business?

    That would mean that the chip availability situation would be once again tightened by an Intel Xe DG2 launch. As we know it sucks quite hard right now with Navi 21, Ampere, the new consoles and Zen 3. And on top of that, a new Intel GPU. That raises doubts about Intel being able to profit from AMD and nVidia availabilites, should they have to approach the same foundry. Rumors appear to point towards TSMC as well.

    1-6000-banner-88x31-jpg

    Stolzer Besitzer eines 3dfx Voodoo5 6000 AGP Prototypen:

    • 3dfx Voodoo5 6000 AGP HiNT Rev.A-3700

    [//wp.xin.at] - No RISC, no fun!

    QotY: Girls Love, BEST Love; 2018 - Lo and behold, for it is the third Coming; The third great Year of Yuri, citric as it may be! Edit: 2019 wasn't too bad either... Edit: 2020... holy crap, we're on a roll here~♡!

    Quote Bier.jpg@IRC 2020: "Je schlimmer der Fetisch, desto besser!"