Nvidia slowed RTX 4090 GPU by 11 percent, to make it 100 percent legal for export to China
- Reference: 1703805759
- News link: https://www.theregister.co.uk/2023/12/28/nvidia_4090_returns_to_china/
- Source link:
Dubbed the RTX 4090D, the device appeared on Nvidia's Chinese-market [1]website Thursday and boasts performance roughly 10.94 percent lower than the model Nvidia announced in late 2022. This shows up in the form of lower core count, 14,592 CUDA cores versus 16,384 on versions sold outside of China.
Nvidia also told The Register today the card's tensor core count has also been been cut down by a similar margin from 512 to 456 on the 4090D variant. Beyond this the card is largely unchanged, with peak clock speeds rated at 2.52 GHz, 24 GB of GDDR6x memory, and a fat 384-bit memory bus.
[2]
As we [3]reported at the time, the RTX 4090 was the only consumer graphics card barred from sale in the Middle Kingdom following the October publication of the Biden Administration's most restrictive set of export controls.
[4]
[5]
The problem was the card narrowly exceeded the performance limits on consumer cards with a total processing performance (TPP) of more than 4,800. That number is calculated by doubling the max number of dense tera-operations per second — floating point or integer — and multiplying by the bit length of the operation.
The original 4090 clocked a TPP of 5,285 performance, which meant Nvidia needed a US government-issued license to sell the popular gaming card in China. Note, consumer cards aren't subject to the performance density metric that restricts the sale of much less powerful datacenter cards like the Nvidia L4.
[6]
As it happens, cutting performance by 10.94 percent is enough to bring the card under the metrics that trigger the requirement for the USA’s Bureau of Industry and Security (BIS) to consider an export license.
While slower Nvidia insisted the performance his is much less than you might think.
"In 4K gaming with ray tracing and deep-learning super sampling (DLSS), the GeForce RTX 4090D is about five percent slower than the GeForce RTX 4090 and it operates like every other GeForce GPU, which can be overclocked by end users," an Nvidia spokesperson said in an email.
[7]
The fact that the card can be overclocked means that conceivably some performance lost by the lower core count could be recovered by customers willing to do a bit of tinkering.
[8]Why Nvidia and AMD are roasting each other over AI performance claims
[9]China's GPU contender Moore Threads reveals card that can cope with Nvidia's CUDA
[10]US willing to compromise with Nvidia over AI chip sales to China
[11]Nvidia's China-market H20 chips hit another speed bump
This isn't the first time Nvidia has nerfed the performance of its graphics cards to comply with US export controls. In late 2022, following a wave of restrictions on the sale of AI accelerators in China, the American giant cut down the interconnect speed of its popular A100 units and [12]relaunched it as the A800. The card was followed by a new version of the H100, predictably called the H800. (Eight is the luckiest number in Chinese numerology: the naming of the A800 and H800 was no accident.)
The fresh round of export restrictions announced in October saw Nvidia again deliver chips designed for the Chinese market. The L2, L20, and H20 [13]limbo danced right under Washington’s performance caps.
Nvidia's antics quickly caught the attention of US Commerce Secretary Gina Raimondo who warned chipmakers against testing those bans. "I'm telling you, if you redesign a chip around a particular cut-line that enables them to do AI, I am going to control it the very next day," she said during a [14]forum in early December.
Raimondo later told Reuters the Commerce Department was working closely with Nvidia to [15]ensure that GPUs and AI accelerators that pose a threat to US national security weren't sold to China, but that the chip house could and should be able to do business there. ®
Get our [16]Tech Resources
[1] https://www.nvidia.cn/geforce/graphics-cards/40-series/
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZY5SdRFYancHB1hCMqohBAAAAI0&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://www.theregister.com/2023/10/19/china_biden_ai/
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZY5SdRFYancHB1hCMqohBAAAAI0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZY5SdRFYancHB1hCMqohBAAAAI0&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZY5SdRFYancHB1hCMqohBAAAAI0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZY5SdRFYancHB1hCMqohBAAAAI0&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://www.theregister.com/2023/12/21/nvidia_amd_benchmarks/
[9] https://www.theregister.com/2023/12/20/moore_threads_mtt_s4000_gpu/
[10] https://www.theregister.com/2023/12/13/us_commerce_dept_nvidia/
[11] https://www.theregister.com/2023/11/27/nvidia_china_h20_accelerator/
[12] https://www.theregister.com/2022/11/08/chipmakers_alter_products_china_ban/
[13] https://www.theregister.com/2023/11/09/nvidia_china_gpu/
[14] https://www.youtube.com/watch?v=5hqSnQzy7Yg
[15] https://www.theregister.com/2023/12/13/us_commerce_dept_nvidia/
[16] https://whitepapers.theregister.com/
It will be in the firmware
NVIDIA has form for crippling cards using firmware changes, that’s how they historically separated Quadros from their GeForce lineup. Enforcement this time round will totally be done using digital signature checks plus natural binning (e.g. if a card has some defective cores, ensuring they are not used but shipping the card anyway, as the spec demands fewer working cores anyway).
In the event someone does defeat the restrictions, they’re left with a card which is statistically less likely to work properly anyway, allowing for legitimate plausible deniability, and if any shenanigans happen, the US government would not want to lose face, nor risk destabilising stocks/shares for which and wealthy investors. So nobody will admit anything and the CCP would probably play along even when their people do find a way round it by not shouting from the rooftops, just continuing to buy more.
Re: Slower Version
I'm sure I heard somewhere that the earliest IBM AT PC's, which shipped with a 6MHz 80286, could be given a very cheap speed bump by replacing the 6MHz clock crystal with an 8MHz equivalent.
Once IBM heard about this, either the design or components were changed to prevent this from working. IBM weren't about to let anyone have a quicker computer without paying for it!
Re: Slower Version
The first IBM AT had a 6MHz bus, designed with pre-release Intel 80286 parts. When it became clear that their existing design could be fitted with an 8MHz processor and overclocked to 8MHz, their response was to release the 8MHz version. The 6MHz version was never stable just boosting the clock speed: Intel shipped 6MHz processors defined by "can't reliably do 8Mhz".
Re: Slower Version
I've also heard tale of how companies will "repurpose" chips that may not quite pass muster. So, everything starts out life as the top of the line 4090, but during QA some of them might have some minor defects. Like not enough of the tensor cores are fully functional to meet the requirements of the 4090, but otherwise everything works perfectly, so they rebadge it as a 4080. Then some may not meet the specs for 4080 and become a 4070 and so on until either all the chips have been rebadged or you run out of models and scrap the rest. I think this is how AMD came up with their old tri-core CPUs. They were quad-core designs, but one of the cores didn't pass testing, so they disabled it and sold it as a tri-core CPU.
Re: Slower Version
Yep, this is pretty well known. You'll often see it referred to as "binning". Defects are so common that it doesn't make sense to discard hardware that's 99.9% good. Though, I think manufacturers still tend to have a couple product lines that they try to manufacture, not just one.
Re: Slower Version
> This makes me think of the -- urban tech legend? ..
After asking people to leave the room, a tech would take the panel off the printer and move a belt up a notch, that doubled the print speed.
Re: Slower Version
Dell's hobbled iDRAC which magically springs to full life when you enter an enterprise licence key in the BIOS? BMW's heated seats? There're many examples.
Easiest limbo ever
Glad to hear it was an easy fix! (as a bonus, this may help prevent the power plug from melting!)
Slower Version
This makes me think of the -- urban tech legend? -- of some of IBM's various, compatible mainframe computers being "the same" inside, and that when a customer paid for an upgrade to a higher model, a tech would come out, install a jumper, and laze around pretending to work for an hour or two. The advantage to IBM (or any other manufacturer) of such a scheme is having to have only one production line for all of its compatible models across a CPU speed range. The disadvantage to IBM (or other manufacturers) is the customer or some hacker might learn of this, and install a "free" speed upgrade for themselves. These days, if such a scheme exists, it probably involves a BIOS change.
I can't help but wonder if nVIDIA has some magic sauce involving test pins and/or JTAG to enable all the cores, and raise the clock speed, of its chips, thus changing the crippled "export" model into the "regular" model.